Knowledge is no guarantee of good behavior, but ignorance is a virtual guarantee of bad behavior.

—Martha Nussbaum

🌺🦟 As I wrote about Google’s plans to breed millions of Wolbachia-carrying mosquitoes in Edition #640, a question popped into my head:

If male mosquitoes don't bite, what do they eat?

I had to look it up.

It turns out that male mosquitoes are basically tiny nectar-drinking bachelors.

Their main food is: Flower nectar, plant sap, honeydew, and other sugary plant fluids.

In fact, both male and female mosquitoes drink nectar, but females also need blood for the protein and iron to make eggs. 🩸🥚

So if they’re not busy trying to bite some poor victim, what do male mosquitoes do all day?

They mostly rest in vegetation, avoid drying out, and sip sugar when they can. Many mosquitoes are most active around dawn and dusk to avoid the midday sun.

Their big evolutionary job is to find females and mate. 🕺

Males often gather in mating swarms, sometimes near visual markers like bushes, puddles, rocks, or other landscape features. Females fly into the swarm. Males detect them partly by the sound frequency of their wingbeats (around 475 Hz, lower than the males' own 700 Hz+). Mating happens in flight or nearby.

After mating, male mosquitoes generally don’t live very long.

Their life is mostly: hatch → drink sugar → chill around → swarm → mate → die

Depending on species and conditions, males may live only several days to a couple of weeks.

Females are miniature vampires: the deadliest animal on the planet, killing close to a million people a year. 🧛‍♀️💀💀

Males? Fragile, short-lived, sugar-powered flying reproductive drones with really good hearing. 🌺🦟🎩

🔎📫💚 🥃 If you want Liberty’s Highlights to continue, this is the moment. Become a paid supporter.

Without your support, this steamboat sinks 🌊🚢⚓

🏦 💰 Business & Investing 💳 💴

🗣️📊💵 Thoughts on Jensen’s Financial Analyst Q&A at Computex 2026 🇹🇼

I listened to part of it and read the transcript for the rest, so I may have missed some tone-of-voice stuff, but I came away thinking that Jensen’s goal was to reframe how analysts think about Nvidia: from “How many more GPUs can they sell, and at what margins?” to “What is the full stack required to produce profitable intelligence/tokens/agent work inside fixed power envelopes?”

Through that frame, a lot of recent moves and the expansion into all kinds of adjacencies make a ton of sense.

• CPUs matter again because agents use tools.

• Networking matters because AI factories are scaling beyond rack-scale.

• Local PCs matter because agents need secure, low-latency access to local files/apps.

• Software matters because enterprises need a runtime/harness/tool layer, and because frameworks/APIs/libraries can be optimized to turn the same expensive hardware into more output.

In other words, the goal is to turn “The GPU Company” into “The Operating System for AI Factories and Agents Company,” which isn’t that crazy since, until recently, they were “The Video Game Company.” ¯\_(ツ)_/¯

Here are some of my highlights:

-Why CPUs matter again (2nd order reason) 🛑 🏎️

Jensen: Because the AI is actively waiting on this back-and-forth loop, tool execution must be incredibly fast. [...] If the tools are slow, a one-gigawatt data center full of GPUs sits idle waiting for a response.

When you think about it this way, the CPU becomes a lot more valuable. If you try to save money by buying slower CPUs and fewer of them than you really need, you may end up with a lower utilization on your very expensive AI data-center full of Blackwells. Whatever savings you think you were getting on the CPU side end up getting swallowed whole by stranded GPU capacity.

In the short term, as long as supply is constrained and AI factories are trying to maximize output inside fixed power envelopes, this is probably good for the whole high-performance CPU ecosystem: Nvidia, AMD, Intel, ARM, and others.

At least for now, performance may matter more than cost. But if Nvidia succeeds with Vera, especially as a standalone CPU, some of that incremental CPU demand will come at AMD and Intel’s expense.

-Local AI isn’t replacing cloud AI, it’s the control loop 🧠💻☁️

AI-capable PCs are in the news these days. But why would you need one if you can get access to hardware worth millions in the cloud by paying a few bucks a month for a Claude or ChatGPT subscription?

Jensen: The future architecture is not an exclusive choice between local execution and cloud processing; it uses both simultaneously. Your local device will run a highly efficient, secure agent continuously looping in the background to handle immediate workflows, local file management, and tool orchestration privately and without metered cloud fees. When that local agent encounters a highly complex reasoning problem requiring multi-disciplinary knowledge, it will automatically call out to a multi-trillion parameter frontier model in the cloud.

The theory is that your local machine becomes the private, low-latency control layer.

I think in some cases it’ll make sense, especially in the enterprise or with prosumer/creators, etc. For most people, their token usage probably won’t be high enough, at least for a while, and the cheap, small online models are getting better all the time. I’d be surprised if it makes sense for most people to run a local model over, say, Haiku 5.0 or GPT-6.0-Fast or whatever comes next unless the local model is doing something truly local.

Side note: it’s a bit surprising that Anthropic hasn’t updated Haiku or Sonnet in a relatively long time, but I suspect most of their attention is going into Opus and Mythos in the pre-IPO period (because that’s where the big bucks are).

In any case, the hybrid model of edge as hands + cloud as brain 🤲 🧠 makes enough sense that it’s likely to keep getting traction with the highest-volume token users, or at least, those that have demonstrable ROI on higher volumes of tokens.

-The most expensive data center may be the cheapest one ⚡🏭💵

Jensen: If you operate within a fixed one-gigawatt power envelope, it is vastly more profitable to fill that envelope with a high-density, ultra-productive $90 billion computing infrastructure than a lower-performing alternative. […]

The absolute key metric for any modern AI factory is performance-per-watt. [...]

What you can drastically reduce is the cost of compute through architectural efficiency. Our end-to-end co-design of the entire rack, networking fabric, and software layer allows us to extract maximum efficiency out of every watt, driving down the true total cost of ownership per token.

In other words, a $90bn datacenter can be “cheaper” than a $50bn one if it produces way more tokens per watt, has better utilization, gets to production faster, and keeps generating useful output for longer.

I mean, the barber is telling you that the more expensive haircut is totally worth it, so be careful with your wallet. But the logic makes sense if revenue depends on $/token more than anything else. 💇‍♂️

-Bonus: Nvidia’s Networking Attach Rate 🔌🛜 + From Hopper to Rubin

At a different event, Nvidia CFO Colette Kress talked about the networking business:

Kress: our attach rate of what we are seeing in terms of the networking is an enormous piece of that as well. We're probably more than 90% attach rate in terms of what we're doing on networking.

Last Q, they disclosed: “Data Center networking revenue was a record $14.8 billion, up 199% from a year ago and up 35% sequentially.”

She also mentioned that Hopper systems (ancient in AI terms!) in the cloud are earning more money than when they were first deployed because Nvidia keeps improving them through software. She distinguished depreciable life from useful life.

She also said that Vera Rubin is now ramping into full production. The one-year cadence is brutal to compete with.

💰💰💰$20bn AUM at 24 💰💰💰

The WSJ has a fun one on Leopold Aschenbrenner and his fund Situational Awareness.

It has gone from a few hundred million to ~$20bn in under two years (up ~270% after fees this year alone, 1,000%+ since inception).

He has a lot of coattailers who copy his positions, but the 13F that everyone's dissecting doesn't even contain his biggest homeruns. ⚾⚾

His Anthropic position is about a fifth of the fund. He bought it at a $61.5bn valuation in early 2025. The latest round was at ~$965bn. 15.7x in a year and a half. Not bad. He also got into SK Hynix back in Nov 2024 and rode it into the trillion-dollar club.

These investments are private and foreign, so they never show up in U.S. disclosures.

Personally, I don’t really have an opinion on him, but it’s fun to watch from afar.

🚀 SpaceX the Neocloud, Google Circular Deal Edition 🤝

If there was any doubt, SpaceX issued this filing:

On June 5, 2026, we entered into a Cloud Service Agreement with Google LLC (”Google”) with respect to access to compute capacity. The compute capacity provided includes approximately 110,000 NVIDIA GPUs, CPUs, memory, and other related components.

Pursuant to the agreement, the customer has agreed to pay us $920 million per month from October 2026 through June 2029, with capacity ramping up through September at a reduced fee.

If we fail to deliver access to the committed amount of GPUs by September 30, 2026, then following a one-month grace period, Google may immediately terminate the agreement or accept the number of GPUs provided, with a corresponding pro rata reduction in the monthly fees. After December 31, 2026, the agreement may be terminated by either party upon 90 days’ notice.

Here’s where it gets circular 🔄

As of the end of 2025, Google owned about 6.11% of SpaceX equity.

SpaceX is expected to IPO around 90-95x TTM revenue, or about 70x estimated 2026 revenue.

This means that if the multiple holds, or even partly holds, the money that Google is spending on SpaceX compute will more than pay for itself in the increase in value of their equity stake, and they still get the compute (possibly to help power their deal with Apple?  🤔).

Now, whether that multiple makes sense is a different question. Should commodity neocloud revenue really earn the same multiple as a near-monopoly in launch and a global satellite constellation?

🇵🇱📈 How to Make a Country Rich: Poland vs Russia A/B Testing 🇷🇺📉

“How did Poland go from communist central planning to one of Europe’s fastest-growing economies?”

Very informative AND fun video. I want to visit Poland someday.

🧪🔬 Science & Technology 🧬 🔭

🤖🔮 Anthropic’s Fable/Mythos Split: Same Brain, Different Chains ⛓️⛓️‍💥

First, the name.

I think Mythos is a better name, and a brand they’ve been building for months. Fable is a cool word too, but it’s a bit less… Mythic

Fable makes it seem more like a children’s story. Oh well, if it kicks ass, I’m sure we’ll soon create all kinds of positive associations with the name and it’ll be fine. I remember thinking at first that “Claude” was a strange name for an AI, even as a big Claude Shannon fan.

Now I think it’s great ¯\_(ツ)_/¯

So Fable 5 is a Mythos-class model, but if I understand things right, there are traffic cops 👮‍♂️ (aka classifiers) that sit in front of it and analyze queries. In Claude’s apps, requests related to cyber/biology/chemistry will fall back to Opus 4.8 instead. On the API, those requests are blocked by default.

There’s also a second, quieter system for frontier LLM development by competitors. If Fable thinks you’re asking for help building pretraining pipelines, distributed training infrastructure, or ML accelerator designs, it may quietly make itself less useful through “prompt modification, steering vectors, or PEFT.” Very sneaky!

The euphemism is safeguards.

The business translation is: No! We will not help you clone us!

To release the model both safely and quickly, we’ve tuned these safeguards conservatively—they’ll sometimes catch harmless requests, though they trigger, on average, in less than 5% of sessions. With more capable models arriving in the coming months, we’re working to improve our safeguards and reduce false positives as quickly as we can.

For cybersecurity researchers and digital infrastructure maintainers, Project Glasswing continues and will get an upgrade from Mythos Preview to Mythos 5 (which is basically Fable 5 unchained).

Even just since Mythos Preview, the model’s capabilities have improved significantly in some areas.

On the cyber ExploitBench it went from 69% → 78% and on the biology BioMysteryBench (hard) from 29.6% → 46.1%.

In one tabletop exercise, generalist biology PhDs using Mythos 5 outperformed teams with plant-pathology specialists on both scientific quality and feasibility. Expert graders estimated the work would normally have taken 40–95 working days, with an average of 72.5 days.

With Mythos 5, two-person teams did it in 16 hours.

Ethan Mollick had early access and was very impressed:

First, how good is Fable? In experiment after experiment I conducted, it outperformed basically every other public model I have used by a considerable margin. It was capable across many problems and produced some startling results — it would work up to a dozen hours executing on multi-page specifications. [...]

I’m guessing he got free access, because the bill for twelve hours of Fable tokens must be sizable 😅

Here’s an interactive Isochrone map that he built with it (9 hours of work).

Soon, we intend to expand access to Mythos 5 through a broader trusted access program.

It’s not entirely clear what this means.

Will it remain a cybersecurity-only model, and they’ll just keep expanding beyond the current 200 orgs, or do they mean that others will have access to Mythos as long as they meet certain criteria?

Anthropic is also highlighting improvements in knowledge work and vision:

On Hebbia’s Finance Benchmark for senior-level reasoning, Fable 5 has the highest score of any model, with substantial gains in document-based reasoning, chart and table interpretation, and problem solving. [...]

Fable 5 is the new state-of-the-art model for tasks involving vision. It can extract precise numbers from detailed scientific figures and can perform complex vision-based tasks like rebuilding a web app’s source code from screenshots alone.

The system card also has an interesting anti-hype point: Anthropic says Mythos 5 is the most capable model they’ve trained, with the highest internal AECI score they’ve measured, but they do not see evidence of a sustained 2× AI-attributable acceleration in their overall AI progress.

In other words: recursive self-improvement is real enough to matter, but not yet clearly compounding into runaway acceleration.

The direction is obvious, but the slope is uncertain… ⏳

Pricing isn’t as bad as some feared.

Both Fable 5 and Mythos 5 will both cost $10 per million input tokens and $50 per million output tokens, which is less than half the price of Claude Mythos Preview.

Anthropic is clearly trying to balance two things: get buzz for its new best-in-class model AND manage compute capacity. That’s why they’re rolling it out in a now-you-see-it-now-you-don’t way.

We expect demand for Fable 5 to be very high, and difficult to predict. On the Claude API and consumption-based Enterprise plans, Fable 5 is fully available from today. For subscription plans, we’d rather give access sooner than later, so we’re rolling out more conservatively, in stages:

• From today through June 22, Fable 5 is included on Pro, Max, Team, and seat-based Enterprise plans at no extra cost.

• On June 23, we’ll remove Fable 5 from those plans. Using it after that will require usage credits. If capacity allows, we’ll extend the included window.

• After this point—when sufficient capacity allows us to do so—we aim to restore Fable 5 as a standard part of subscription plans. We intend to do this as quickly as we can.

If you have a flat-rate plan and want to give it a try, don’t wait too long.

If you want WAY more details on both Fable 5 and Mythos 5, check out the 319-page model card. 📄📄📄

🇪🇺🥵 Heat Deaths in Europe 🌡️

I’ve written about this a few times before, so you know what I think of the anti-A/C crusade (while the rich folks are vacationing at their seaside villas and the poor folks are stuck in humid suburbs, all while shutting down the nuclear power plants and constraining energy…).

There’s been a viral chart lately showing gun deaths in the U.S. vs heat deaths in the EU.

While it’s a very satisfying chart that supports my position when it comes to A/C in the EU, the methodology needs some fixes. Hannah Ritchie did a great job:

the chart has several issues.

The heat death numbers for Europe are modelled as “excess deaths” during summer months; this attempts to capture people dying from a range of health conditions — cardiovascular disease, stroke, and others — earlier than they would have in more optimal temperatures. This is a common way researchers measure heat deaths (or cold, for that matter). I’ve done a deep dive on these methods previously.

But the US number isn’t based on this type of modelling; it’s based on heat deaths recorded on death certificates. If you used death certificate figures for Europe, they’d be far lower.

It also uses European Union figures for gun deaths, but a broader definition of Europe for heat deaths.

I thought I’d have a quick go at making a better version. These numbers are still not perfect, but hopefully a closer comparison.

Check out her post for all the details, but long story short, here is the upgraded graph with per capita numbers:

So many unnecessary heat and gun deaths. And deaths aren’t the only thing. There’s also plenty of suffering caused by both that doesn’t result in deaths but still matters.

🎨 🎭 The Arts & History 👩‍🎨 🎥

💾 WarGames Revisited 🍿

I had so much fun watching Ferris Bueller’s Day Off (1986) with my kids recently that I decided to go to the next logical Broderick step: 1983’s WarGames.

To be honest, I’m not sure if I had ever seen the whole thing before. I remembered a few scenes, but chances are I just caught part of it on TV as a kid (if you’re young and didn’t live through this, it was very common never to see the beginning of films because you just randomly changed the channels and stumbled on a film that was already started). 📺

In any case, it was a big hit with my kids, and another great opportunity for me to explain how the world used to be back in the old days.

It’s uneven. The second half kinda loses momentum at a few points, and the protagonist stops driving the plot… It was early enough in the genre that there probably wasn’t a clear template for techno-thrillers and they couldn’t maintain what they had the whole way through. The formula was perfected later with Sneakers (1992, I love that film, it’s tons of fun, I gotta watch it with my kids 🤔).

But overall, it still works.

In fact, it works very differently post-2023/ChatGPT. The whole chatting with an AI aspect doesn’t seem so science fiction anymore. Neither does giving battlefield control to software…

Apparently, Ronald Reagan saw the film shortly after it came out and asked his advisers whether this could actually happen. The answer was basically: not like the movie, but it’s something to worry about. The film may have helped shape policy on cybersecurity and hacking.

Fun detail: I’ve read that they were inspired by Stephen Hawking for Falken, but when they show footage of him in his lab, I thought it was clearly an homage to Claude Shannon. Some of the black and white shots look like famous Shannon photos, and the whole ‘building whimsical gadgets and studying games’ also fits (maybe with a sprinkle of John von Neumann and Richard Feynman in there).

The actual character of Falken in the film is mostly a doom monk and doesn’t convey that at all, but at least that little segment gave me that vibe.

The NORAD room was a challenge. The giant screen graphics weren’t just modern-style post-production overlays (impossible at the time). They were generated on HP 9845C desktops that weighed 100lbs and cost $90k in today's dollars. Then they had 130,000 feet of film printed and projected for the actors. An insane amount of work to make fake computer maps feel “live.”

The lose-lose, anti-war message of the film is just as relevant as it ever was.

People with enduring personal finance success — not necessarily those with high incomes — tend to have a propensity to not give a damn what others think about them.

—Morgan Housel, The Psychology of Money

🏋️💪🗓️ I was looking at the May monthly stats on Hevy, the app I use to track my workouts, and I realized that last month was the least I've lifted in a long time, by total lbs volume.

I want to turn this around.

I’ve decided to try to make June the highest month in the past year, which means beating last June. I’ll report back, but I figure a little public commitment may help 😅

(You should probably harness that power for whatever you've been struggling to get done.)

🕰️⏱️⏲️⏰⌚ I need more time to read. I want to read more history, more biographies, more science, more novels.

I need more time to exercise. I need more time to walk in nature.

I need more time to spend with my kids and my wife.

I need more time to play guitar.

I need more time to write, think, and research.

I need more time in the sauna. I know it’s good for me.

I need more time to get back into running. I wish I hadn’t stopped.

I need more time…

🏦 💰 Liberty Capital 💳 💴

🔄 💾 Recursion Starts at Home aka Anthropic’s Escher Hands

Read more

Self pity is the cause of many sorrows and the solution to none of them.

—Stoic Emperor

🚢 📈 Another arbitrary milestone.

Near the end of May, the crew on this steamboat reached 29,000. That’s 2.4 Hobarts!

I always think the graph is a bit funny. The least viral chart ever. No spikes, no hits, just chugging along and doing its thing.

It probably doesn’t help that you and I go in ALL the directions and it’s hard to share just one segment cleanly. But oh well, I optimize for what I find interesting, not for what travels smoothly ¯\_(ツ)_/¯

29k is a lot of people if, like, you’re hosting a backyard BBQ. But it’s not a lot of people in the grand scheme of things.

I looked it up, and there are about 1.5 billion people who understand English (380 million native speakers and 1.1bn+ as a second or nth language). Of course, only a tiny fraction of 1% of these are nerdy enough to fit the psychographics of our crazy Venn diagrams, but that’s still quite a big potential crew.

So we’ll see. I’m curious how many inboxes will be on the receiving end of my keyboard at Edition #1,280 🤔

👨‍🚀🚀⭐🦠 Last week, I published one of the most fun podcasts I've ever recorded. If you missed it (and have read or seen Project Hail Mary, because it’s spoilerific), check it out:

• 🎧 Project Hail Mary Explained: Book vs Movie (Spoilers, Science & Why Rocky Works) 🧑‍🚀🚀⭐🦠🪨

I’m still reading the book out loud to my kids, and they’re enjoying it SO MUCH. During the big moments, they are literally on the edge of the sofa. I love to see a good story have that effect. I hope that in many years, they’ll look back fondly on these moments together.

I’m also looking forward to watching the film together once we’re done with the book.

🔎📫💚 🥃 If you want Liberty’s Highlights to continue, this is the moment. Become a paid supporter.

Without your support, this steamboat sinks 🌊🚢⚓

🏦 💰 Business & Investing 💳 💴

⚡💻 RTX Spark: Nvidia Puts Data-Center Silicon in a Laptop

We knew it was coming, one of those badly kept secrets. But now it’s here for real.

Nvidia announced a system-on-a-chip (SoC) designed to power Windows laptops and give them the vrrooomm vrroooom necessary to run AI agents and heavy workloads locally.

The RTX Spark superchip features an NVIDIA Blackwell RTX GPU with 6,144 CUDA cores and fifth-generation Tensor Cores with FP4 precision, connected via the NVIDIA NVLink-C2C chip-to-chip interconnect to a high-performance, 20-core NVIDIA Grace CPU.

You can bet that the next version will have Vera CPU cores and a Rubin GPU.

It’s kinda wild that you can pack chips from the same family as what’s in hot data-center racks into a laptop (though of course, they aren’t exactly the same chips, otherwise you’d have molten aluminum on your lap).

They want to do this with a mix of new Windows security primitives (for stuff like identity and containment policy) and Nvidia’s OpenShell, which is a kind of sandbox environment around agents like OpenClaw.

A big selling point, like with the DGX Spark and Apple’s Macs, is the unified memory available to both the CPU and GPU (in most PCs, the GPU and CPU have separate memory pools, and the GPU doesn’t have enough to run larger LLMs).

Nvidia has announced ‘up to 128GB of unified memory.’

In theory, that’s enough to run a 120-billion-parameter model with 1 million tokens of context, locally 🤯 (but it may struggle with complex agentic tasks, since they take so much more memory because of the very extensive reasoning and fast-growing multi-agent context windows).

Aside from waiting for the benchmark numbers to come out, the biggest ❓ is about Microsoft’s support for the Arm ISA. Windows and its ecosystem of apps are still mostly x86. While they have an emulator called Prism that translates from x86 to Arm on the fly, its reputation isn’t as good as Apple’s Rosetta, which does the same thing (to be fair, Apple had a much less heterogeneous ecosystem to deal with).

Intel and AMD’s x86 moat is getting a new breach attempt.

🚀 “SpaceX is officially a neocloud” (With a Space Side-Business) ☁️

Turboblitz posted a chart showing SpaceX’s revenue by source, and then Racketboy/Nick updated it to include xAI+Twitter:

I’m mostly sharing because I think it’s funny. But also, let’s give some kudos to xAI’s infrastructure team.

The model team may have dissolved and never really got much traction (and is in the process of being rebooted with Cursor’s transplant), but the infra team will likely turn out to have made a huge difference for SpaceX’s IPO.

It’s hard to know exact capex numbers invested in Colossus 1 and 2 (which add up to about 1 GW of compute together), but the AI Segment level numbers show $26.5B cumulative AI capex from 2023 through Q1 2026.

By comparison, SpaceX disclosed $19.9B of cumulative capex for its core Space + Starlink/Connectivity businesses from 2023 through Q1 2026. If we go back to the foundation of the company in 2002, cumulative capex isn’t disclosed, but if we take the consolidated gross PP&E and remove what we know is the AI capex ($26.545bn), we get a napkin ballpark proxy of $42bn.

And Colossus 2 isn’t completed yet. As far as I can tell, the first phase was for ~110,000 Nvidia GB200s (~210 MW), the second was for ~110,000 GB300s (~220 MW), and the next one will add ~220,000 GB300s (>400 MW). The first phases have been brought online in record time. We’ll see if they keep the streak going. ⏱️

But one thing is certain: Without those juicy Anthropic bucks, the picture for SpaceX would look a lot worse.

🤖 Why Frontier Models Capture So Much Value (aka When 10-20% Better Is Worth 10x More) 🍰

The main reason is that for many AI workflows, quality is nonlinear.

A model that is 10-20% better on the right dimension may not be 10-20% more valuable. It may cross the line from ‘toy’ to ‘usable,’ or from ‘assistant’ to ‘autonomous agent.’ Coding is the best example of that right now, and we saw that shift happening last December (circa Opus 4.5 and GPT-5.2 Codex).

The likely equilibrium is not ‘frontier wins everything’ or ‘open source commoditizes everything.’ If I had to guess, I’d say that open models will win volume, frontier models will win the work that people can’t afford to get wrong (the stuff that is high-risk, high-reward, high-trust… by far the most valuable), and the app layer will eat some margin in between.

Here's the part that trips people up: token cost usually isn't the binding constraint.

If an engineer is using AI to avoid shipping a bug that costs a week of cleanup or makes a customer angry, the gap between a few cents and a few dollars of inference is noise vs the cost of getting it wrong. In those cases, buyers are rational to overpay for the best model.

Is there such a thing as ‘good enough’?

The answer depends on the task.

For bounded, verifiable, low-risk tasks, you can find ‘good enough’ and leave it at that. Once a model can reliably extract fields from invoices, sort support tickets, and summarize meeting notes, you move on to optimizing for stuff like cost, latency, privacy, and integration.

For tasks that are open-ended, high-leverage, high-risk, ‘good enough’ is always kind of just over the horizon. Coding agents, legal and medical reasoning, anything that runs unsupervised. With that stuff, you could always use fewer errors and better quality. A model that fails 8% of the time may be unusable. A model that fails 2% of the time may be transformative.

Something that fails 0.5% of the time may expand the TAM and spawn a whole new industry.

The State of Token Subsidies: $2,180 of Tokens for $200 🪙

Simon Willison:

I currently subscribe to the $100/month Max plan from Anthropic and the $100/month Pro plan from OpenAI. If you are a heavy user of coding agents these plans are a fantastic deal. I just ran the ccusage tool on my laptop to get an estimate of how much I would have spent if I were to pay for API tokens in the past 30 days and got:

• $1,199.79 for Anthropic Claude Code

• $980.37 for OpenAI Codex

That’s $2,180.16 worth of tokens for $200—not bad at all! I’m a moderately heavy user of these tools, but I’m certainly not running agents every hour of the day and night.

Ccusage allows you to track how many tokens you use and do the math on how much it would cost at API pricing. For example:

Now, of course, the math is always a bit more complicated than it seems. Flat rate pricing is great for getting new users in the door who might not have tried otherwise, or would have limited their use too much to find their own product-market-fit… which then makes their usage explode, and they end up paying for extra tokens at pay-as-you-go rates.

There’s also the fact that for every whale who will completely max out every 5-hour-window and weekly allocation on Claude Code Max, there are a lot of other users who only use a fraction of that yet still pay the same rate. In the end, it’s the average that matters most, and light users subsidize the whales 🐳🐋

But even with that said, seeing the large disparities between flat and API pricing, I don’t think the situation is sustainable, and we’ll keep seeing the end of the all-you-can-eat buffet era of AI.

OpenAI = Dropbox? 📦 🤔

Finally, here’s an interesting part of last week’s Sharp Tech where Ben Thompson (💚 🥃 🎩):

Dropbox tried to push for a long time to be a consumer company, and they had to pivot because consumers don’t pay for productivity. It’s like the initial version of OpenAI: “We’re going to monetize through consumer subscriptions.” No, you’re not. You won’t make money. Consumers are emailing themselves documents as opposed to paying for storage.

This is such a fundamental error that OpenAI is the latest one to make, but Silicon Valley makes it again and again and again. People in Silicon Valley are obsessed with productivity. People in the rest of the world don’t want to be productive at all—they want to be entertained.

This is why Meta wins. They provide entertainment. […]

The reason why enterprises pay for productivity is because they’re paying people to come to work, and they want to get more value from the dollars they’re already putting in. So they will buy them new computers, they’ll pay for software, and they’ll give them a Dropbox because they’re getting more value from what they’re already paying their people.

Enterprises care and pay for productivity, and consumers don’t. And Silicon Valley continually doesn’t learn that lesson. Dropbox used to be the canonical example. I think OpenAI took that one over.

To be fair, OpenAI has grown revenue faster than almost any other business in history…

BUT

The point is a good one, and they’ve had a much harder time monetizing the non-enterprise, non-productivity side of the business while carrying large infrastructure costs that Anthropic has been less burdened with.

It doesn’t mean that the future will 100% rhyme with the past, and it could be that someday this large consumer pool of users will become a highly monetizable asset, but apparently that day isn’t today.

🧪🔬 Science & Technology 🧬 🔭

Memory is Now Almost 2/3 of AI Chip Component Costs

Everyone’s talking about memory. You’re probably talking about it with your Uber driver right now!

High-bandwidth memory (HBM) has grown from 52% to 63% of total AI chip component spending between Q1 2024 and Q4 2025. These estimates are an average across all AI chips designed by Nvidia, AMD, Google, and Amazon, weighted by production volume.

As a share of spend, logic dies stayed roughly flat near 13%, while advanced packaging fell from 19% to 15% and auxiliary components fell from 15% to 9%. In absolute terms, HBM spend across these four designers grew from roughly $12 billion in 2024 to $32 billion in 2025, a faster year-over-year increase than any other component. (Source)

And 2026 is set up for even more of it.

🏴‍☠️📲 Meta's Support AI Hijacks Instagram Accounts If You Just Ask Nicely 😬

A bunch of Instagram accounts got hacked recently. They include some big ones like the Obama White House account and Sephora, and some short ones like ‘@hey’ that are valuable in the same way that short web domain names are.

So far, nothing too unusual. Damn these sophisticated hackers, right?

Then you read how this was done, and it’s not a great look for Meta:

Look, I’m no spring chicken. I’ve spent almost a decade and a half identifying vulnerabilities and exploits at unicorn scale, but this is hands down the most unserious, “almost too stupid to be true” of them all.

The Takeover Flow

• Step 01: Faking the Location & Initiating Support
  All the attacker needs to kick this off is your account username. Then, they hop on a VPN or proxy close to your city so Instagram’s security algorithms don’t suspect a thing. (You can quite easily get this from your public profile or “About” section or a hundred other ways.) Once it looks like the request is coming from the correct region, they tell the Meta support AI that the account is hacked and ask it to send the verification codes to an arbitrary email address they control.

• Step 02: That’s It
  Really, that’s it. The first proper zero auth password reset I’ve seen in production. There appears to be no additional check as to whether the email being given is actually something the user has used before. Once the AI sends the security code to the attacker’s email, the attacker passes it right back to complete the verification. The platform hands over a fresh password reset link, granting full ownership to the attacker.

What about 2-factor authentication? The thing that everybody tells you will protect you?

This bypasses it entirely 🤦‍♀️

Existing sessions are revoked and the password changed with no email, text, or push notification. The actual owner can’t initiate recovery because the email and phone numbers now map to the attacker. There’s no human to escalate to, it’s just you arguing with a chat hoping to take control back while praying they don’t do it again.

What’s ironic is that Meta pitched these AI systems as a way to prevent account issues and hacking.

In this March announcement, they said the support assistant was a way to "resolve account problems for you from start to finish" and said that its AI could "prevent an account takeover by noticing it was suddenly accessed from a new location, the password was changed, and edits were made to the profile." That's almost move-for-move what the attacks did.

One Meta AI was supposedly watching for exactly this, but a different Meta AI waved it through.

It looks like Meta has patched the issue, but it was active for months, with Telegram groups offering stolen accounts on the black market.

Who needs Mythos when you can just ask the Meta assistant to hack for you? 😅

Google Wants to Release 64 Million Mosquitoes Carrying Wolbachia Bacteria in Florida and California 🦟🦠

Back in Edition #457, I wrote:

Personally, I’m in favor of heavy mosquito control anywhere humans are present in any concentration. Mosquitoes can have the rest of the world — the planet is plenty big for everyone — but they’ve been annoying me for too long…

Alternatively, I’d take a shoulder-mounted automated laser mini-turret that shoots down any mosquitoes in a 30 feet radius 🤔

No luck yet on the laser turret, but it looks like Google’s Project Debug is working on this.

The plan reads like a Bond villain's opening monologue: release 32 million mosquitoes into California and Florida per year for two years.

BUT

This is the anti-villain plan. The goal is fewer infections and, if it works, fewer of the little vampires ruining your evenings outside.

The CDC calls mosquitoes the deadliest animal on Earth, and the one Debug is targeting this round is Culex quinquefasciatus, the southern house mosquito behind West Nile virus and St. Louis encephalitis. West Nile is the most common mosquito-borne illness in the US, with 1,300-plus severe cases and over a hundred deaths in a typical year. (Worth noting: Debug started by attacking Aedes aegypti, the Zika and dengue mosquito. This proposal is a pivot to a different 🧛‍♂️.)

The method is older than it sounds, and it works because of one biological quirk: male mosquitoes don't bite. Only females need a blood meal to make eggs, so releasing males adds zero bites to your summer.

Debug raises males carrying Wolbachia, a bacterium found naturally in loads of insects, and when one of those males mates with a wild female that lacks the same strain, her eggs never hatch. Repeat at scale and each generation comes out smaller than the last.

No genetic engineering required. Google didn’t invent Wolbachia. It, uh, borrowed it. This is the Sterile Insect Technique, a 1950s idea that already cleared screwworms and fruit flies off the board. Mosquitoes were just too delicate to mass-produce until the robots showed up.

And that's the part that makes this a tech story rather than a bug story. You cannot release the females (they bite, and they'd grow the population you're trying to shrink), so the whole thing rises or falls on sorting males from females accurately, tens of millions of times, cheaply. That used to be slow manual work; Debug built sensors, computer vision, and automated rearing to do it at industrial scale, which is the actual hard part.

One catch worth flagging: it's self-limiting. The released males die, nothing propagates through the wild gene pool, and the population drifts back unless you keep releasing. Reassuring if you worry about irreversible meddling, less so if you wanted one-and-done.

It's maintenance, not a cure: stop releasing, and they come back.

Does it actually work?

The Fresno field trials knocked biting females down 68% the first year, then 95%-plus, with the most isolated neighborhood hitting 99%. And in February, a randomized trial in Singapore covering 700,000-plus people found Wolbachia releases cut symptomatic dengue by more than 70% (published in the NEJM). Debug says it's now released over a billion mosquitoes across four continents.

If you're on team fewer mosquitoes, the EPA's taking public comments until June 5. 📩

(The main objection is ecological: collapse a population and you may affect the birds, bats, and dragonflies that eat them. My guess: real in general, overstated here, since both target species are invasive in North America to begin with, anything that eats mosquitoes also eats a lot of other (juicier) insects, and the Earth is a really big place. There will be plenty of mosquitoes left even if we treat near human population centers.)

🎨 🎭 The Arts & History 👩‍🎨 🎥

🔊 Eddie Kramer on Hendrix, Zeppelin, the Beatles, and the Happy Accidents in the Studio 🎤🥁🎸

I didn’t even know who Eddie Kramer was before I watched this interview, but I enjoyed it enough to share with you. His anecdotes from the studio with Hendrix, Zeppelin, and even the Beatles a few times were super interesting.

He made me notice the glockenspiel on Little Wing in a totally new way. Apparently, Hendrix saw it lying around the studio, had never played it before, and added those well-placed notes to the song in a way that totally transformed its texture. He also has some anecdotes on Jimi’s cover of All Along the Watchtower.

The anecdote about the “woman” pre-echo on Whole Lotta Love is crazy. Zeppelin fans spent the 1970s lying on their beds with headphones, so they’ve noticed that during the middle part, the echo follows Robert Plant’s vocals on every word except one, where the ‘echo’ comes first.

Here, lemme pinpoint the moment using the timestamp feature (it’s quiet, so you may have to crank it up to hear):

It sounds almost supernatural, and people had theories that it was done with studio trickery. But Kramer says it’s accidental. A past take’s vocal recorded on another channel on the same tape bled through while they were recording, and landed in just the right spot! What are the odds of that?

Anyway, it’s a fun one. I just like these stories that tell us a bit of what it must’ve been like on the inside as some of these great records were being made.

🚨 If you haven’t read the book or seen the film, bookmark this and come back. Seriously. This one’s worth avoiding spoilers and going in blind. 🚨

We go deep on the book first, because it’s the foundation and the more complete version of the story, then turn to the film, the production, and what the adaptation gained and lost.

Some of the rabbit holes we go down: 🕳️🐇

• Not a reluctant hero. An avoidant one, which is very rare!

• Why making Grace alone on the ship is the decision that makes everything else possible, and how a group of humans may have reacted differently to first contact.

• Stratt runs one algorithm on every choice, and it’s the one most of us are too squeamish to run. Why she’s the best character in the book, and why the film’s karaoke scene is its worst mistake.

• The most alien alien possible: no face, lives in ammonia, looks like a spider. Weir stacked the deck against Rocky on purpose. So why do we love him? 🕷️👽

• Is Rocky a preview of humanity's first contact with AI? 🤔

• The one plot change in the film that Andy Weir says he wishes he’d thought of for the book.

• A species that never looked up, and what that says about everything humans built because we happened to see the stars. 🔭

• “Everybody did their job” on the film. The imperfectionist cinematography of Greig Fraser. 🎞️ 🎥

• Gosling carrying a whole film alone in a room with a puppet, a four-hour cut you’ll never see, and why the book is still the canon.

And, as always, we wander: Interstellar’s water-planet plot hole (Tinkered’s blood boils), whether more gravity would actually make Grace stronger rather than older, R2-D2 as the secret hero of Star Wars, the James Joyce theory of the “kneecapped” modern male, and me-burgers. 🍔

Amaze!

(Yes, it’s a kind of reunion episode, with the same crew as on the pod for A Knight of the Seven Kingdoms ⚔️🥚🛡️🏰🐴. We had so much fun, we had to do round #2.)

🎧 Listen on Spotify

If you prefer to listen on Spotify, here’s the feed:

• Liberty’s Highlights podcast feed on Spotify

🎧 Listen on Apple Podcasts

Here’s the podcast feed on Apple Podcasts:

• Liberty’s Highlights podcast on Apple

📺 Watch on YouTube

→ Here is Jameson’s episode on Project Hail Mary at Becoming the Main Character.

🗄 More TV & Film Podcasts from the Archives 🍿

• 🎧 A Knight of the Seven Kingdoms: Going Deep with Tinkered Thinking and Jameson Olsen ⚔️🥚🛡️🏰🐴

• 🎧 White Mirror: 31 Stories, a $1,000 Puzzle, and Sci-Fi to Reboot Your Childlike Wonder 🦾🪞📖🦋

• 🎧 David Senra on Deadwood (also: founders, entrepreneurs, civilization, Henry Singleton, Walt Disney, David Ogilvy, etc)

• 🎧 Sicario by Denis Villeneuve: Going Deep with Ed William and Dylan O’Sullivan 🎥 🎬 🎞️ 🍿

• 🎧 The Last of Us: Going Deep on HBO’s Amazing Show with Matt Reustle and Tinkered Thinking 🍄

• 🎧 TENET Deconstructed: ↔️ Christopher Nolan’s Mind-Bending Time Thriller 🔁 (with BOTH Tinkered Thinking and Dylan)

🗄 More Podcasts from the Archives 💡

• 🎧 Trillion Dollar Club with Mostly Borrowed Ideas (MBI): Nvidia, Apple, Google, Microsoft, Amazon, TSMC, Meta, Broadcom, and Tesla 🏆💰📈🤖🚀

• 🎧 Kevin Kelly: Excellent Advice for Living 💡🤖 📚

• 🎧 The Cost of Glory with Alex Petkas: Timeless Lessons from Ancient Greece and Rome 🏛️🏺📜 🏹⚔️🦅⚖️

• 🎧 Semiconductors Industry: Going Deep with Mostly Borrowed Ideas (MBI) on TSMC, Texas Instruments, Nvidia, Intel, Hyperscalers, etc

• 🎧 David Senra of Founders Podcast (Podcast #11)

• 🎧 Claude Shannon fan club with Jimmy Soni and David Senra 📖

• 🎧 Going deep on Payments with Mostly Borrowed Ideas (MBI): Adyen, PayPal, Stripe, Square, Amazon, Apple Pay, Google Pay, Visa, Mastercard!

• 🎧 Going deep on Constellation Software with Mostly Borrowed Ideas (MBI)

You can become a paid supporter to get access to paid Editions + access to the private Discord + Zoom Q&As with me + support the recording of more podcasts:

🍪🍪🍪🍪🍪🍪🍪

You can get far in life just by being extremely curious.

Passionate inquisition will lead you down paths most people never take & are completely unaware of.

You can be most well rounded experienced person you know just by being curious.

“Interested people live Interesting lives.”

—Brother Lobo

🧍⌨️🖥️ I wanted to show you my standing desk, so I took a photo and had AI redraw it (to anonymize it). And yes, I’m standing as I’m typing this.

Years ago, I read about the benefits of standing desks and figured I’d try.

My prototype was simply stacking a bunch of boxes on my desk to raise my monitor and keyboard to standing height. I tried a few days and liked it, so I bought a used IKEA Fredrik desk for $30. Some online forums recommended it as a cheap way to go standing.

The desk wasn’t designed to be standing, but because it had those top shelves and side posts with holes every few inches, you could mount the desk surface to standing height.

I used that as my main desk for a few years and got used to standing while working, and walking around the office as soon as I had something to think about. But then in 2014, my wife and I had our first kid, and oh boy, he didn’t sleep very well for 18 months.

Survival mode. I went back to a sitting desk 🥱

That situation got better… then we had our second boy, and he had the same sleeping genes. So I didn’t go back to standing, and inertia being what it is, I’ve been sitting ever since.

But during that decade, standing desks got WAY more popular. Many motorized models came on the market, and I started thinking about going back, but this time with a hybrid that could do both.

I bought a motorized frame from MotionGrey (on sale around $300CAD), and I asked a friend who’s great at woodworking if he could make me a hardwood top (👋 Fred! Thanks again 💚 🥃). He suggested a few possibilities, and we ended up going with white oak (it’s beautiful).

I’ve not been using it for that long (my wife can attest to how slow I’ve been to assemble it, even after the frame and top arrived). But I have to say: It’s a real quality-of-life improvement.

Without having to force myself, I end up working standing up at least 2-3 hours per day on average, and over the years and decades, I think that will add up to all kinds of benefits.

If you spend your day at a desk and haven’t got a standing desk yet, I recommend it. And if you have young kids, get a motorized one. Fixed-height is a commitment you may come to regret at stupid o'clock.

🔎📫💚 🥃 If you want Liberty’s Highlights to continue, this is the moment. Become a paid supporter.

Without your support, this steamboat sinks 🌊🚢⚓

🏦 💰 Liberty Capital 💳 💴

🏭🤖 Nvidia Q1 Highlights: Compute Is Revenue, Demand Goes Parabolic, and Jensen Wants the CPU Market Too 😎

Humans get used to things really quickly. I was watching Apollo 13 with my family last weekend (🚀🧑‍🚀🧑‍🚀🧑‍🚀). One of the plot points is that Apollo 11 (first time on the moon) had the whole world watching, but by Apollo 13, going to the moon had become seemingly so routine to most people that the major TV networks didn’t even bother carrying the crew's live broadcast. 😅

Sometimes it feels like this with Nvidia.

They have some of the best quarters any public company has ever had, and everybody goes 🥱 and takes it all for granted. I don’t necessarily mean that the stock should have moved more, because that’s the expectations game (yes, they beat “consensus”, but in the Keynesian beauty contest, beating consensus is itself the consensus, so it’s not really a beat 🔁). Mr. Market is clearly skeptical of the duration of growth (🤨).

But we’ve become blasé about these numbers very quickly.

• 85% revenue growth at that scale, selling physical products!

Jensen described it this way: “This was an extraordinary quarter, demand has gone parabolic.”

For comparison, the last time Google grew revenues at that pace was in Q4 of 2005. But at the time, its market cap was $120bn, not $5.2 trillion.

And Google’s gross margin back then was ~60%. That’s a purely digital business (at the time), vs Nvidia’s 75% gross margins, and they ship metal.

• Nvidia’s net income was $58.3bn in Q1, up 211% year-on-year 🤯 (GAAP, which includes $15.9bn of gains from equity securities, to be fair)

For most of its 33-year history, Nvidia was a gaming company. Now it doesn’t even report gaming separately. 🎮

They lumped everything that isn’t “Data Center” into “Edge Computing”, and all of it added up to $6.4bn last quarter compared to $75.2bn for DC. One big downside of that, IMO, is that networking isn’t broken out as a segment anymore. It’s now part of DC, which is too bad, though we did get specifics in the CFO commentary:

• “Networking revenue was a record $14.8 billion, up 199% from a year ago and up 35% sequentially.” 🛜

Let’s dive into management’s commentary to see what’s going on underneath those numbers:

One thing they try to do is convince the market that they’re not only selling to a handful of hyperscalers. They will report two segments within DC, with the hyperscalers on one side, and what they call ACIE on the other (AI Clouds, Industrial, Enterprise, which also bundles in sovereign, on-prem, and any odds and ends):

Colette Kress (CFO): Hyperscale revenue of $38 billion was approximately 50% of Data Center revenue and increased 12% quarter-over-quarter.

ACIE revenue was $37 billion and grew 31% quarter-over-quarter, including AI cloud revenue that more than tripled year-over-year. Our customers have enabled rapid stand-up of AI compute capacity.

The number of partner data centers exceeding 10 megawatts has nearly doubled in just 1 year, now surpassing 80 sites.

Sovereign revenue increased more than 80% year-over-year.

Hyperscale grew 115% YoY, faster than ACIE's 74%. Doesn’t quite fit the "hyperscaler AI spend is maturing" narrative. In fact, hyperscalers are accelerating faster than the new diversified base Nvidia is trying to highlight.

Kress also made a pitch against the 'Nvidia is more expensive' refrain:

Today’s data centers are revenue-generating AI factories […] Customers do not buy GPUs. They build AI factories and the right economic metric is not the purchase price of the GPU. It is the lifetime cost of an AI factory producing intelligence, token per watt, tokens per dollar, uptime, utilization, time to production, software durability and asset life.

Constrained by power and capital, AI factory operators must choose the right architecture. With our extreme codesign approach, we deliver the industry’s lowest token cost, the highest token throughput and the highest ROI. MLPerf Inference results are in, and once again, we swept every benchmark as Blackwell Ultra delivered the highest throughput across the broad set of models and deployment scenarios.

Full stack innovations drove the 2.7x increase in throughput and a 60% reduction in the cost per token on GB300 compared to just 6 months ago.

That's the moat many people miss. Even if you catch up to last quarter's Blackwell, this quarter's Blackwell is 2.7x faster on the same chips. Software optimizations ftw.

They also announced a “deepened” partnership with Anthropic, but I'm unclear what that actually means. I’m guessing they sweetened their offers to them to pull them away from TPUs and Trainiums a bit ¯\_(ツ)_/¯

After years of GPUs getting all the love, CPUs are back in the spotlight, and Nvidia wants its big slice of the 5-layer cake (🍰) there too:

Colette Kress: Agentic AI and reinforcement learning represents new growth opportunities for CPUs. Building on the success of our Grace CPU, Vera is arriving just in time to meet this inflection. Built on custom ARM cores and codesigned end-to-end with Rubin GPUs and NVLink, Vera will deliver up to 1.5x faster performance per core, 2x performance per watt and 4x density per rack compared to x86-based alternatives.

Vera CPU opens a brand-new $200 billion TAM for NVIDIA, a market we have never addressed before, and every major hyperscale and system maker is partnering with us to get it deployed.

I’m always wary of large TAM numbers, but in the AI world, big numbers have had a tendency to materialize.

We have visibility to nearly $20 billion in total CPU revenue this year, setting us up to become the world’s leading CPU supplier.

More from Jensen on why demand for CPU is exploding:

Jensen: Claude Code is essentially a harness around Claude. OpenAI's Codex is a harness around the GPT-5.5 model… these harnesses provide for things like I/O, orchestration, memory management, tool use connected to tools, for example, browsers and things like that, C compilers, python compilers.

the harness runs on CPU. And the tool use runs on CPUs. So for example, if the AI were to do a search or do a browser, use a browser, that would run on the CPU. The world has 1 billion users, human users. My sense is that the world is going to have billions of agents.

Not today, I mean, we’re going to grow into it, but we’ll have billions of agents. And those billions of agents will all use tools. And those tools can be like PCs, just like us humans using PCs today.

Grace (which was bundled) and now Vera (which can be purchased standalone) allow Nvidia to sell the whole stack, rather than relying on AMD and Intel CPUs. They’ll still offer them, but it probably feels good not to have to sell your competitors’ products.

Morgan Stanley had an estimate of the impact of rising memory prices on Vera Rubin, Nvidia’s next generation:

Bonkers, no?

It'll be interesting how long they can keep passing on those COGS to customers without taking a margin hit. And whether less differentiated competitors get squeezed harder by the same memory inflation. 🗜️

On China, it’s still Schrödinger's market. 🇨🇳🐈

• Q1 FY26 had $4.6B in Data Center Hopper shipments to China.

• Q1 FY27: zero.

And they still grew 85% anyway.

They got approval from the U.S. side for H200, but got blocked from the Chinese side. For a while, they were blocked from the U.S. side but got demand from China. Who knows where this will go, but they say that they aren’t including any revenue from China in their forecasts.

Despite that, they're guiding $91bn next quarter, another sequential acceleration on top of a record base.

Finally, on the capital allocation front, since it doesn’t look like they’ll plough tens of billions into the frontier labs anymore (Jensen said the recent $30bn OpenAI check and $10bn Anthropic check are probably the last big ones, since they're heading toward IPOs), they announced an $80bn buyback authorization + raised the dividend from $0.01 to $0.25. Not moving the needle that much, but at their scale, options are limited.

They generated $48.6bn in free cash flow just last quarter. The question becomes: what do you do with it? 💰💰💰

🗣️💬 Gavin Baker: When Software Becomes Labor, TSMC as the Accidental Central Bank of AI, GPU Racks in Space, Etc

Interesting interview on the state of the markets and AI.

Here are highlights:

-The Vertical ARR Curve: AI Revenue Has Entered the Weird Zone 🛸

Gavin Baker: what [is] happening in AI was the most extraordinary moment in the history of capitalism, the history of American business. What I mean by that is Anthropic added $11 billion of ARR. What is astonishing to me about this, is the SaaS and cloud revolution created, we’ll call it, between $5 and $10 trillion of value. Arguably the 3 highest-profile SaaS companies in the last 10 to 12 years are Palantir, Snowflake, and Databricks. These 3 companies employ thousands of people, tens of thousands collectively. They’ve all spent 10 years building their businesses, and Anthropic added their combined businesses in 1 month.

Nothing like that has ever happened in the history of capitalism — forget my career, just the flat-out history of capitalism, the history of business. It’s wild.

So what explains this? How can AI labs grow at a rate that would’ve sounded fake in almost any previous cycle?

I think it’s because software revenue used to be tethered to humans, even when the software itself scaled beautifully. I mean, it’s great that copying bits has zero marginal cost, but if to get more usage you need more employees, more seats, more workflows getting approved, more training, more integrations, more sales people and more steak dinners, more people changing their habits… there’s still a lot of friction in the system.

Or to put it another way: Even the best SaaS companies had to pass through the narrow straw of human adoption. 🥤

With AI agents, that straw turns into a firehose.

If you give a company 1,000 AI agents, you don’t need to hire 1,000 people before they can start consuming tokens. They can scale from 10 active tasks to 100,000 tasks overnight. Compute consumption can ramp up instantly by orders of magnitude without waiting for Debbie from HR to give orientation to a single person.

Software used to amplify labor. Now software is becoming a kind of labor.

This doesn’t mean that things won’t ultimately be limited by other factors, including ROI on all those tokens, which are increasingly expensive (and as I wrote in Tokens Schmokens, a bad metric).

-TSMC as the anti-bubble central banker 🐜🏗️🇹🇼🫧

Gavin Baker: I have been optimistic that this fundamental shortage of wafers, which today is controlled by Taiwan Semi, will prevent one. If Taiwan Semi did what Jensen wanted, I think Nvidia could sell $2 trillion of GPUs in '26 or '27, maybe $2.5 trillion, maybe $3 trillion. But there is a limit where consumers would consume so much, that you probably would be in an overbuild. So Taiwan Semi — if we don't get a bubble, we need to throw a party for them, because they will have single-handedly prevented a bubble.

Usually when people ask “is AI a bubble?”, they look at demand: hype, valuations, capex plans, extrapolated revenue, etc. 📈

What if the answer is on the supply side?

TSMC’s multi-year lead times and conservative capital expenditure are acting as a hard ceiling on global wafer output. Even if you have the cash, you can’t overbuild, or at least, not as much as if supply was freer.

It’s almost like TSMC is the accidental central bank of AI, and they have a “wafer monetary policy” trying to keep the AI economy from running too hot. 🏦

What will be a secondary effect of this?

Forced Capital Efficiency: Because hyperscalers cannot buy all the raw compute they want, they are being forced to innovate at the algorithmic and optimization layers much earlier in the cycle than they normally would.

Kind of similar to what is likely to happen because of memory supply constraints. Expect optimizations to come that will reduce memory usage, possibly significantly (it’s impossible to predict breakthroughs, but we can look at incentives and know that they have changed A LOT in recent times, and the Eye of Sauron has turned to RAM 👁️).

-When Junk Companies Become Cash Machines 🗑️➡️💰

Gavin Baker: Because we are in shortages, the lowest-quality companies are doing the best. If you’re an oil and gas investor, a mining investor, natural resources investor, and you’re well-versed in thinking of costs, this is very intuitive to you. In a real bull market for a commodity, the commodity suppliers with the highest costs go up the most because it’s the most beneficial to them. They go from on the verge of bankruptcy to gushing cash. This is one reason commodity investing is really, really hard. Because quality outperforms during the cycles, but you get all of the outperformance during the downturns when the high-cost guys who mooned during the shortages and the commodity bull markets go bankrupt or whatever.

You’re seeing that happen in every industry. The lowest-quality players — companies that are hated and detested by the hyperscalers and the buyers because they have high costs, they’re unreliable, the parts fail at a high rate — they’re sold out and raising prices. And then that activity gets the interest of these retail accounts on X, and these stocks get bid to the moon. Whereas some of the higher-quality expressions have actually really underperformed. As an investor, it’s hard because you know within a shadow of a doubt that the thing that’s mooned 10x in 3 months or 6 months is going to go right back down, subject to what they do with all the cash.

That can be counterintuitive if you’re used to thinking “better business = better stock.”

Anyone who has looked at levered equity stubs with high risk of bankruptcy knows that operating leverage cuts both ways. And when the direction of travel reverses rapidly, the results can be pretty crazy. 🚀

🗣️💬 Anthropic’s CFO Krishna Rao: Compute is our lifeblood 🩸

I rarely feature two interviews by the same interviewer at the same time, but Patrick was on a roll recently and I enjoyed both enough to highlight.

There’s more interesting stuff in this one than I can fit here, so I recommend that you listen to the whole thing to get a better idea of how one of the most consequential companies operates and thinks internally.

Oh, and maybe it’s just me, but it’s a little weird to listen to the audio version because Krishna’s voice has a timbre that sounds a little like Patrick’s (at least, it does at 2x).

Here are my highlights:

-Compute Allocation > Capital Allocation 🤖🤖🤖

Krishna Rao: The compute that we procure is the lifeblood of our business.

It is the most important thing in the company. It’s the canvas on which everything else gets built. The decisions we make on how much compute to buy are some of the most consequential and hardest decisions to make in the entire company. [...]

Thinking of it as not just something that is a variable cost over some time period, but really this resource that’s so fungibly utilized. We run workloads on one day in the morning on a chip for inference, and in the afternoon or evening, we use it for model development.

That paradigm does not exist in a company like a software company or a factory. You can’t repurpose — if you have a bunch of people doing R&D and that’s your R&D expense, they can’t go and become COGS, and vice versa, in most traditional companies. Here, you really have that fungibility. That’s possible.

That’s where the return on compute is so important, and I think people are beginning to understand that. But there’s still a tendency towards treating it like, “Oh, I have to separate these two costs,” when in actuality they’re very self-reinforcing, and that flexibility is what helps to drive revenue short term and long term.

He also drops a crazy stat:

“Our net dollar retention rate is over 500% on an annualized basis.” 🫡

Elite SaaS sits at maybe ~130%. Some exceptional ones have been at ~150-170%.

500% is unprecedented.

And then this:

“On the way here, I was in an Uber, and I signed two double-digit million-dollar commits in the car ride, which was 20 minutes.”

Reminds me of the anecdote about how in eBay’s early days, one guy’s job was just to open envelopes full of checks. This is the AI-era equivalent 😅

-Frontier models unlock TAM, not just better benchmarks 🔓🔑

Krishna Rao: Every time we have a new model, there’s a set of capabilities that are different. People tend to think about model intelligence as IQ — a single number. “Okay, this model was at 110, and then it goes to 125.”

We think of it differently.

Intelligence for us is multidimensional. It’s not just a score. We find that, yes, everyone publishes their model benchmark cards. We find that a lot of those benchmarks are saturated. We publish it too, but our measurement is what the customers tell us. What is the real-world capability of this model?

As we’ve released better and better models, what we’ve seen is it’s not just the outright intelligence. It’s also the ability to do long-horizon tasks. It’s the ability to use tools, or computer use. It’s the ability to do agentic tasks that have specific value even faster. […]

What we’ve found very consistently is that by releasing new models, the TAM is unlocked in a unique way. More TAM gets unlocked. More use cases are possible.

A good illustration of that is this last four months that we’ve had at the company. We started the year with about $9 billion of run-rate revenue, and we ended the quarter with north of $30 billion of run-rate revenue. That kind of a change is enabled by these model intelligence leaps, and then the products that we build around them. That’s what I mean by the returns to frontier intelligence are really high.

I think that’s an important reframing.

It may not seem like a big deal at first, but going from “same thing but better” → “new use cases, new possibilities” matters if we want to understand what is likely to keep happening.

A better model doesn’t just take share inside the existing market. It can colonize whole new corners of the economy.

And what’s 🤯 is that just by the time this interview was recorded, edited, and came out, Anthropic’s run-rate went from $30bn to roughly $45bn, based on the latest reporting.

In fact, there are reports in the WSJ that Q2 is predicted to show $10.9bn in sales, up from $4.8bn in Q1, as well as their first operating profit ($559M). But Anthropic isn't guiding to sustained profitability, they expect to keep losing money for a while longer as compute spend ramps. It’s almost an accidental profit ¯\_(ツ)_/¯

-Anthropic’s Escher Hands Moment 🔁🪞🤖 ⌨️

I wrote about the Escher Hands Era, and clearly we’re going deeper and deeper into it:

Krishna Rao: I can’t speak for other companies, but for us, the scaling laws are alive and well, and we’re seeing that with releases more recently like Mythos. But right now, within the company, 90-plus percent of our code is actually written by Claude Code. A lot of Claude Code’s code is written by Claude Code.

Why do we allocate compute internally? Why would we forgo revenue for it? It’s because the models themselves are helping us build that next generation of models.

Clearly they’re not building with Sonnet.

It’s obviously Mythos that is coding the next version of Claude, and the fact that they haven’t released it widely (which means it can’t be distilled or studied by competitors) probably also has to do with extending that recursive advantage for a bit longer (not just cybersecurity and compute constraints).

There’s also good stuff on how they think about compute allocation, how fungible it is between different use cases, and how they are the only lab more or less agnostic between using Nvidia/TPU/Trainiums.

🧪🔬 Liberty Labs 🧬 🔭

🐜 From Six Dells to 4,032 x GB300s: Inside Jane Street’s New Datacenter 🛜

Caveat: Jane Street is a Dwarkesh sponsor and produced this video. It’s marketing, so treat it accordingly (🧂). With that said, it’s fun to get a tour of one of their datacenters.

It’s a recent one in Texas that they use for training AI models, running liquid-cooled Nvidia GB300s (!). They talk about how much work went into making the DC liquid-cooled, and how many sensors and failsafes they have to detect and isolate leaks (you don’t want to short out millions of dollars of equipment because of a leaky pipe!).

Compared to what hyperscalers are building, this is a tiny cluster, but even with just 4k GPUs, there’s over 8,000km of fiber in that deployment (the distance from Lo Angeles to Tokyo)!

After the DC tour, they sat down to discuss some of the stuff JS does:

This part is 🤯:

There are trading systems we build, and trades that we do, where in order to be competitive you actually have to turn around a packet in under 100 nanoseconds.

That’s a very different regime, right?

People sometimes ask, “Can you write high-performance stuff in OCaml?”

And the answer is: we can. But for this kind of speed, it doesn’t matter whether you write in OCaml, Rust, or C++. You can’t use a CPU.

You’re going to be on an FPGA that’s directly attached to the network, and you’re going to be turning around the packet so fast that if you attached an oscilloscope to the wire on the way in and the wire on the way out, you would see the packet start to leave before it’s done being consumed.

So it’s a very different, very specialized regime. And when you’re in that time regime, you really can’t do very much computation.

This made me curious how much throughput came from the raw feeds from the various exchanges (equities, options, FX, crypto). This is what I found with the help of both Claude and GPT-5.5 (with multiple rounds of them fact-checking each other in extended thinking mode, but errors can remain):

Man, Options are not just bigger than equities… they're in a different weight class. ⛰️

Nasdaq's April 2025 systems commentary has a useful reference point: roughly 165 billion options messages per day versus 7 billion equities messages per day on its systems. A ratio of about 24:1.

So the guesstimate (because this isn’t public info) is that for a tier-1 multi-asset firm like Jane Street, total provisioned inbound market-data capacity is in the hundreds of Gbps range.

And then what kind of compute do you apply to all this raw data... That’s another question entirely.

Cloudflare’s Experience Hunting Security Flaws with Mythos 🔐 🏴‍☠️

It’s starting to feel like Mythos is building a separate brand from “Claude.” The longer Anthropic keeps it unreleased and separate, the more it’ll feel like its own thing 🤔

Anyway, back to the cybersecurity beat, Cloudflare shared its experience using it:

We soon pointed [Mythos] at more than fifty of our own repositories – to see what it would find, and to see how it works. [...]

Mythos Preview is a real step forward, and it’s worth saying that plainly before getting into anything else. We’ve been running models against our code for a while now, and the jump from what was possible with previous general-purpose frontier models to what Mythos Preview does today is not just a refinement of what came before.

It’s a different kind of tool doing a different kind of work, and that makes a clean apples-to-apples comparison to earlier models difficult.

Mythos isn’t just acting like a better static analyzer. Cloudflare says the big leap is in exploit-chain construction and proof generation.

In other words, it can find several individually unimpressive bugs, reason about how they combine into something more dangerous, then write and run proof-of-concept code to test whether the exploit actually works. 😬

And even if you can use tools like Mythos to find vulnerabilities faster, and then fix your code faster, if the rest of the pipeline for deployment remains the same, it won’t be able to handle this new work. Cloudflare explains:

The loudest reaction to Mythos Preview from other security leaders has been about speed - scan faster, patch faster, compress the response cycle. More than one team we have spoken with is now operating under a two-hour SLA from CVE release to patch in production. The instinct is understandable: when the attacker timeline shortens, the defender timeline has to shorten with it.

Faster is not going to be enough, and we think a lot of teams are about to spend a lot of time, effort, and money learning that the hard way.

Patching faster does not change the shape of the pipeline that produces the patch. If regression testing takes a day, you cannot get to a two-hour SLA without skipping it, and the bugs you ship when you skip regression testing tend to be worse than the bugs you were trying to patch.

Source.

🎧🇺🇦 Ukraine: Update from the Frontline of the Drone Wars

The latest episode of WarTalk has guest Rob Lee reporting from Kyiv on how things are these days in Ukraine. Drone tactics, how both sides are organizing operations, what's likely next, and why Rob Lee thinks Starlink is the game-changer of this war.

The part about how Vampire drones are used for logistics is very sci-fi. They airdrop medical kits to wounded soldiers, lay mines, and carry infantry packs so the guys can sprint all out across open ground until they reach cover, where the drones drop their gear off, like flying sherpas.

Rob Lee also says that only about 2% of Ukrainian casualties in one area came from small arms. The 'forward line of troops' has been quietly replaced by a forward line of drone teams.

🧠📊 The Two-Thirds Decline in Dementia Risk, and Why Cases Keep Climbing

We still need to figure out how to bend that curve down, but at least we can clearly see some progress over the decades.

The good news first:

A comprehensive analysis of three large population-based studies found that the age-adjusted prevalence of dementia has decreased by two-thirds over the past 40 years (Stallard et al., 2025). The data revealed a staggering 67% drop in dementia risk at any given age over this period.

For instance, in 1984, 30% of people aged 85–89 had dementia; by 2024, this figure had fallen to just 10%. Additional analysis of the data also revealed that each successive five-year birth cohort has a lower risk of dementia at a given age than did their predecessor.

But still:

Dementia is one of the biggest healthcare challenges of the 21st century, affecting millions worldwide. In the UK, nearly 1 million people are currently living with dementia, a figure projected to rise to 1.4 million by 2040. In the United States, more than 6 million people suffer from dementia, and the condition is responsible for over 100,000 deaths annually.

Dementia also carries a substantial economic burden, with total costs exceeding $600 billion annually in the United States. (Source)

Despite the encouraging decline in dementia incidence, the absolute number of people living with dementia continues to grow because the average age of the population keeps going up (as fertility keeps going down).

🎨 🎭 Liberty Studio 👩‍🎨 🎥

🚀🔧🌕🧑‍🚀🧑‍🚀🧑‍🚀 Apollo 13: The Hard Way Ages Best 🌎

Rewatched this with the family last weekend. First time in years for me, first time ever for my two kids. It was even better than I remembered. Probably my favorite Ron Howard film.

About halfway through, one of my kids asked “Wait, all this happened?”

Yep, pretty much. They worked really hard to keep it accurate.

The zero-g still looks amazing.

This is a 1995 film, so the effects should look dated. They don’t, because Ron Howard refused to fake it with wires. He talked NASA into letting them use the KC-135 reduced-gravity training plane (affectionately known as the “Vomit Comet” 😅). It flies parabolic arcs that produce about 25 seconds of free fall at a time.

The eleven-person cast and crew flew 13 days over a four-week period, did 612 parabolas, all that to get probably no more than 15-20 minutes of usable zero-g footage that ended up in the film.

Kevin Bacon described it: 40 parabolas in the morning, lunch (which someone, foolishly, decided should include Mexican food 🌮🌯), 40 more in the afternoon. 🤢

There’s a reason other sci-fi productions don’t do this! You have to be a little crazy.

My kids never even questioned it. They just took it as given that they were watching people in space.

A few other things that made me appreciate it more this time:

• It’s a systems movie, not an action movie. The hero isn’t a person, it’s a process: checklists, communication loops, simulations, redundancies, inventory. Everyone solving small problems in the right order. The heroes are not saved by one genius epiphany. It’s dozens of small correct decisions made under pressure.

• The CO₂ filter scene (make this square thing fit into that round hole) is probably the nerd ideal of constraint-driven engineering on film. 🤓🛠️

• Both famous lines are slightly off from the historical record.

  • Jack Swigert actually said “Okay, Houston, we’ve had a problem here.” Past tense. The film changes it to present tense (”Houston, we have a problem”) because it’s punchier.

  • “Failure is not an option” was invented for the screenplay. The real Gene Kranz never said it. The line stuck because it captures the culture perfectly, which is what the movie is faithful to, even when it isn’t transcript-perfect (but a lot of the other dialogue comes fromt recordings and transcripts).

• A sadder note: Jim Lovell, the real commander whose book the film is based on, died in August 2025 at 97. He has a cameo near the end as the captain of the USS Iwo Jima, the recovery ship. Ron Howard apparently wanted him to play an admiral, but Lovell said: “I retired as a Captain; a Captain I will be.” 🫡

If you have kids who haven’t seen it, this is a great family watch. Suspenseful, clean, historically rooted, full of adults behaving like adults under pressure… which is rarer than people admit.

Want more?

I knew it!

Here you go:

Unthinking respect for authority is the greatest enemy of truth.

—Albert Einstein

🌖↔🌎 I think most people know about ocean tides, even if few people understand all the variables that make them so complex to accurately predict (it’s not what I want to get into today, but it’s an interesting phenomenon).

But there’s another kind of tide that I didn’t know about until recently:

Solid-Earth tide!

Yep, gravity doesn’t care if you’re a liquid or a solid. It pulls, and over something as large as the Earth, you get an actual tide in the crust too (which is why they also call it crustal tide, but that sounds a lot yuckier).

The ground under your feet can rise and fall by several inches (tens of centimeters) over a tidal cycle. You don’t notice it because everything is moving very slowly with it (buildings, roads, whole cities). It’s probably happening right under you right now!

Earth tides, like ocean tides, vary a lot based on latitude, local geology, and the lunar/solar alignment.

This has to be taken into account if you want very precise spatial positioning. The International Association of Geodesy (I had never heard of it before) says: “The entire planet’s surface rises and falls by up to about 30–40 centimeters twice daily. This deformation affects geodetic reference frames, satellite orbits, and precise positioning, requiring continuous modeling and correction in space geodesy.” 🛰️

One of the things that didn’t make immediate sense to me about tides (water or crust) is that there isn’t just a bulge in the direction of the moon. There’s also a bulge on the other side. 🫤

What unlocked my understanding was when I learned that tides are caused by differences in gravity across Earth, not just the direction of gravity.

So imagine three cars driving at the same speed in a line: 🚙 🚙 🚙

If the front one accelerates a lot, the middle car accelerates moderately, and the back car only accelerates a little (gravity pulls more strongly on what is closest than what is farther away), then from the point of view of the middle car, it looks like the line has stretched in both directions.

That’s the two-sided tidal bulge.

Wait, why am I writing about Earth Tides? I guess we don’t always pick the rabbit holes we fall in! 😅

🔎📫💚 🥃 If you want Liberty’s Highlights to continue, this is the moment. Become a paid supporter.

Without your support, this steamboat sinks 🌊🚢⚓

🏦 💰 Business & Investing 💳 💴

✨ Constellation Software’s ✨ 2026 Annual Meeting

First, I gotta say, I’m a little disappointed in myself. After years of wishing they would go back to an in-person meeting, they finally do, and I wasn’t there. I didn’t have a huge scheduling conflict or anything, I dithered a little too much, time flew, and the date snuck up on me and it was kind of too late (well, I was also really busy with lots of projects and family stuff, but when is that not the case?).

So I didn’t get a chance to see my old friends in person, or to make new ones. Next year!

The meeting was over 4 hours long, but friend-of-the-show Trevor Scott (🍁) made this supercut (👆), condensing it into 36 minutes. That’s a good way to get some highlights.

Here are mine.

First, like everything else in the world of technology, AI was a huge focus. It’s both what shareholders are uncertain about AND the main technological driver of change these days, so it’s everywhere.

As a sentiment check, they were in a good enough mood to make a joke about changing the name of the company:

Speaking of culture, Mark Miller followed in Mark Leonard’s footsteps and has elected not to receive a salary or bonus as President (Constellation doesn’t have a CEO). Leonard had stopped taking a salary in 2014, and in 2015, he also decided to drop other bonuses and incentives too.

Mark Miller: Yes. I just carried on what Mark Leonard did. And I really care a lot about this company, and I took this job to help this company and the shareholders in it.

John Billowits: Mark wanted 0 comp, and we wanted to pay him something.

Something stood out to me at this AGM. Not with huge certainty, but I moved my thinking in that direction a bit: CSU's weird structure may be unusually well-suited for this moment, at least compared to most other software companies.

Decentralization plus running lots of experiments and scaling up what works is how markets work. With 1,500+ small units that can each experiment close to customers and share what works with the rest, interesting ‘invisible hand’ effects are emerging.

AI Case Study: When 50x Faster Changes the Question 💡

David Wilkes: I experienced one of the Volaris AI workshops back in March last year. And I went away from that with one question, which was simply how fast can we go. And what I did on day 1 was I stopped all development across both businesses.

So [at] CELCAT and Bullet [we] stopped for the whole month of April last year, and went on an intensive learning program, to learn what we could do with the technology. And I think if I look back and then we look now at the outcome of that, it’s actually pretty straightforward.

I like this idea of taking an actual pause to really focus on learning. Too many try to learn complex new workflows and tools on the side while doing things in the old way, and that’s rarely very effective, or at least, it isn’t for everyone in the org (some nerds will always race ahead).

At the outset, we came back in May, and we were going faster. If you think of a cycle time from concept through to production software, typically for us, that was measured in months. When we came back in May, we already started measuring that in weeks. And as we went through last year, we turned that from weeks into days. And today, our cycle time is measured in hours.

We can cycle from concept through to production in hours. And that means a road map that might have taken, say, a whole year […] that can be literally a week now.

So this pace at which we produce software is no longer a problem that we have different challenges instead. But it’s fundamentally transformational in terms of what it means for our business.

Jeff Bender: And how are your customers reacting to this? Can you talk to that?

David Wilkes: Yes, yes. Well, let me put that first in the context of growth. So if I go back to the end of last year, our growth was running at about 12% per annum. To date, it’s running at 23% per annum. Our vision is that we’ll double our revenues by 2029.

I think I’ve also been told I’m sandbagging on that. The interesting thing there is that growth. It’s not driven by going faster. It’s driven by us asking a totally different question, and that is, if we can go 50x faster than what does that look like for our customers. So what does 50x faster look like for our customers?

Speed can go from quantitative to qualitative if there’s a big enough change in velocity. You don’t just do the same thing faster, you do whole different things, you do more and create new value.

That's exactly what you'd predict in a Jevons-paradox-of-software world. Nice to see it confirmed by an operator.

And for us, that meant we’ve literally rebuilt our product from scratch. We’ve, in effect, reimagined what a new category of product might be. If you think about our old product, it was a system of record. You created, read, updated, deleted data. Our new product is a system of action.

It thinks, it predicts, it gives you insights. It does really clever stuff for you. So that creates a huge value for our customers and with value comes new price points doesn’t it? And with that, we are driving growth, and we’re seeing the results of that already.

If development goes from months to hours, the roadmap stops being a fixed annual plan and becomes more like a live conversation with customers. Ideas that used to be too small, too weird, or too low-ROI to justify engineering time can suddenly become worth trying.

"System of record" → "System of action"

Agents Need Managers & Names (and Slack Accounts) 🏷️

Interesting anecdote about the use of AI agents, and how to integrate them in the human org:

Andrew Jones: Rather than having support agent #1 and support agent #2 and support agent #3, we decided to name them... we wanted to get adoption of the agent. [...]

They all have a Teams [chat] account. So rather than you having to go to some web page or try to find it, you Teams them. You communicate with them exactly as you would with anyone else within ClickDimensions. And I think that’s the most important thing. [...]

The other thing is as well is by giving the agents names, you also have a human who’s a manager of them, because again, this is really vital when you set up agents is. If you set up an agent and you leave it to its own devices, it will go off. And over a period of time, it will go on a tangent. You need to have a human manager who manages those agents all the time. So by giving them a name and managing it, that’s the way we move forward with it. [...]

It took us about 4 weeks to do it after going through those various steps. We launched it in January of 2026. And what we found was that with our customers’ interactions with this chat agent, conversational agent rather than being a lookup agent, it was a problem resolution agent.

It effectively resolved 82% of all our Tier 1 tickets that went through the chat agent, which is fantastic.

These agents may not be as good as talking to a human, but at a lot of companies, that’s not where you start anyway. Replacing more limited agents can help solve a lot of customer problems faster and free up human agents to deal with the more complex cases.

The AI Hackathon That Built Solutions Nobody Deployed 💾

I’ll leave you with this anecdote told by a shareholder:

Ashwin Annamalai: In Waterloo, there was a hackathon called Anhackathon that was organized for university students where they were like businesses come with their problems, use AI and solve our problems. Like this is amazing. So with the help of AI with Claude and all these tools, they should be able to solve all these problems. So you know what, they did manage to solve the problems. But at the end of it all, zero solutions were deployed for the business.

That’s when — it was the lightbulb moment for me. And I was like, yes, building code is one part of the solution, but having the ownership and the trust is really important. The very next day, like started buying Constellation.

Mark Miller’s response: Success is not judged on your ability to develop products. Success is judged on your ability to go out and convince the customers that they want to use your product.

Now, I don’t read too much into this because in reality, it’ll most likely be competitors that adopt AI and use it to compete better, and they’re likely to have a lot more existing customer relationships and sales knowledge than university students.

But it also means that as long as you don’t fall behind those competitors, it’s more of an arms race than bringing a knife to a gun fight. 🔪

🏎️ Innovation cycles vs procurement cycles 🐌

A good reminder that it’s not because you innovate fast that you can sell fast:

Bill Delaney at Modaxo: They’re asking us about the pace of change from an innovation perspective. And we looked at the research, and it was clear that innovation cycles were compressing right?

From months to weeks and soon to be days. And I looked around the room and I said, but what’s happening with procurement cycles.

Now we deal with large government procurements that have — it takes multiple years for a procurement even to come to tender, then they run the tender process, you do a multiyear project and then you have 10 years of maintenance. Those procurement cycles and government are getting even longer. So we’ve got this dichotomy of extended procurement cycles compressing innovation cycles.

Even if the tech is perfect, that's not what wins the deal:

Bill Delaney: The elements of those tenders that really people should have a look at there’s elements around product capability and functionality and — but then most of the scoring goes to your experience, your ability to contract on the heavy-duty terms that they’re putting forward — and they need somebody to take liability, right? Governments like to shift risk — so you have to have capacity to take liability that might be through accepting LDs if we don’t perform, posting bonds or guarantees. You’ve got to have a scale on our capacity to do that.

That’s a big advantage that incumbents have in public or highly regulated industries. A new entrant with a perfect product that they sell for 1/10th the cost may still struggle.

AI = Layoffs? ✂️

In response to a shareholder asking whether AI productivity gains will lead to layoffs:

Bill Delaney: I don’t see us doing large-scale layoffs. No, not at the moment. We’re so focused on how to leverage this capacity because what you want to understand is we just have an enormous backlog of our customers wanting us to do things for them. Some of those things that just didn’t make business sense now do make business sense.

On the backlog/organic opportunity:

Barry Symons (Jonas): I do the hypothetical, if the business gods came down from above and said, ‘Barry, here’s the deal for you. You will never ever, ever win another customer again in Jonas’ history. But at the same time, you will never ever, ever lose a customer that you currently have in the Jones organization. Would you take that deal?’ And I say 150%, I would take that deal because there’s so much more we can do for our customers. And AI just makes it so much easier.

But wait, why weren’t they doing all this before?

Time to trot out Jevons' paradox again.

Cheaper development reveals a giant backlog of work that wasn’t ROI-positive before (or at least, wasn’t a higher ROIC than other ways to spend the capital, such as M&A).

But are they seeing organic growth from this yet? 🤨

Mark Miller (President): I’d love to see more organic growth. I’d love it to be yesterday […] I also haven’t seen a lot of new revenues from AI as well. So we’re just being -- we’re just trying to drive adoption of the usage of it, and we’ll monitor it internally

Like most things, investment comes before return, so this isn't a surprise yet.

Jamal Baksh (CFO) on cost: It wasn’t a material increase in that COGS line that. yes, I would believe that, yes, we’re probably outpacing revenue with cost today because we’re investing a lot. But it’s not a material impact on margins.

Has AI caused any customer attrition yet? 💸

This was a recurring topic:

Mark Miller: We really haven’t seen any AI specific attrition at any point.

Jamal Baksh (CFO): I look at about 900 of the businesses and look at trending of organic growth... there are a couple of things that might have been popped out and both of them had nothing to do with AI.

Jeff Bender: There’s no specific evidence to date of any significant or material attrition as it relates to AI. But I do believe it’s coming... Stay paranoid right? Like always.

💰 On Capital allocation (the core competency) 💰

The biggest challenge for the company has long been to deploy its free cash at high ROIC, and most of that has been through M&A. This remains the core function:

Bernie Anzarouth (CIO): We had a slide up there earlier that said we had how much, $3 billion of dry powder. We're trying to invest that capital, and it's a very tough slog. It's tough despite the number of VMS businesses that are out there, it's tough to use to redeploy all of our capital. And that's why we came up with the idea of PEMS [...]

How’s the runway? How’s the runway? Everybody always asks about the runway… Bernie doesn’t seem worried:

Our database of these software companies is still quite large, and it's still building.

I’m trying to remember the last number I heard. I think it was ~60k VMS that they were tracking. Probably way more now, my number is dated.

Bernie: And if you think of what the software world is all about, especially now with AI tools that are out there that you can get more software businesses that are created on a regular basis, the number of acquisition targets that are here today will be multiplied several fold over the next 10 years or so.

On the investments in public companies (PEMS):

Bernie Anzarouth (CIO): If you think of the way that we've done acquisitions to date, we buy businesses lock, stock and barrel, and it takes 1, 2, sometimes 3 years to get them up to speed with the best practices that we use to run our businesses. And so if you think of a minority shareholder and how much influence you can have on a business that you do not own 100%, it will obviously take a little while longer in order to influence management to do the right thing. And so what we believe is that we have to get them at a lower price than the businesses that we are actually acquiring at 100%. And so you will see if we measure ourselves at the same IRR that we measure 100% ownership, we would have to get these businesses at a lower price.

Sounds like it’s currently possible to get that extra discount because publicly traded software has been hit a lot harder than the private market. But maybe it’s just a lag, and private acquisitions will get more attractive… which may make PEMS less so.

Billowits had a good line:

We do believe that we want to have the opportunity to take advantage of which could be a dislocation in the market for a period of time. So keep selling vertical market software businesses. We want valuations depressed as long as possible.

🫡

Mark Miller (President): 25 years ago, we — we used to buy minority stakes in public companies and those companies ended up being put into play. It generated external high return... but the average tenure of our investments were very short. And nearly always 90% of the time got outbid by a third party when the businesses were sold... That short-term activism was profitable, but it was a poor strategic fit with our objective of being good permanent holders of vertical market software businesses. So we stopped doing it.

Buy and hold forever, we’re in it for the long run, and [PEMS] is another tool. The other part of the [PEMS] thesis is that the companies we want to invest in have the potential to benefit from an engaged minority shareholder. We want our [PEMS] or public portfolio companies to be run by people we respect — we want their incentives aligned with those of long-term shareholders. We like the businesses to generate high returns on incremental invested capital, and to return any excess capital to their shareholders.

PEMS is not the old short-term activism playbook, even though that worked out financially, and it’s possible that investees will get acquired and the “permanent” part of PEMS will turn out to be short.

The biggest open question is: Can they successfully influence companies they don’t control enough to improve them and unlock value?

And is 'permanent' really the right target? Or is it aspirational, like Buffett's 'ideal holding period of forever'? Will they really hold even if a company takes a wrong turn or gets wildly overvalued? 🤔

Bernie seems to confirm it really is permanent:

Bernie Anzarouth (CIO): I mean selling the businesses, if the price runs away with us, that just doesn't make sense tax-wise. It just would be tax inefficient. And I think what we would do is have or just see like-minded investors piling to the business that we're in. and just keep holding on to it for the long haul. We're not in the business of selling high-priced shares to people that are unaware of what's going on within the business. We're literally in there permanently.

But we’ll see how that turns out in practice, or if they change their minds over time.

On Buybacks 💵 ↩️

John Billowits: The way we think about it is we're using the same hurdle rates for making both decisions. We're buying businesses at a value that is far beneath what we're trading at and what other companies are trading at even after we fix them. So we still believe we can generate much higher returns by buying companies than we can by buying our own shares at these prices.

Mark Miller: Yes, I don't want to buy back shares. Definitely.

Customers as M&A Radar 📡

This one is from Lumine’s CEO:

David Nyland: So we’ve got some big customers worldwide, big brand names and -- they have a lot of intelligence. And so we learn a lot by speaking to them by finding out what the problems are, and obviously now supporting a portfolio of companies that sell into those customers. But also as they navigate their assets, they’re important assets, they start to share with us ideas about what else we could acquire to help them provide security with some of those assets that are in transient ownership at this point. So you have to kind of earn the position to have that conversation.

It takes some time, but the power of that relationship is unbelievable.

In fact, customers may want them to consolidate a space because it makes things easier and more predictable for them:

David Nyland: We mainly get new logos through acquisitions right? It’s very hard to go into a large media or telco business and become a new vendor. They want less hurdle. They want consolidation.

And if there’s something risky in the environment that they’re going to want us to do it. They’re not going to want some new name company that they don’t know. So they will always give us the first chance, so we can move quick, we should be able to capture that business.

Not a bad position to be in.

It’s kind of software as distilled customer trust, to riff on a metaphor used by Mark Leonard. The code matters, but the customer intimacy may matter more.

If you’re newer to Constellation and want a primer on the company, check out the podcast I did with my friend MBI (🇧🇩🇺🇸). It’s a few years old now, so it doesn’t include the latest developments, but I think it gives a lot of context on the business:

• 🎧 Going deep on Constellation Software with Mostly Borrowed Ideas (MBI) (2022)

Andrej Karpathy Joins Anthropic Pre-Training Team 😮

Anthropic’s winning streak continues. One of the essential minds in AI wrote:

Karpathy has moved around a lot over the years, but clearly not randomly. He has been:

• Internships during his PhD:

  • Google Brain (summer 2011)

  • Google Research (summer 2013)

  • DeepMind (summer 2015)

• Then one of the founders of OpenAI

• Then at Tesla to lead Autopilot and be the director of AI

• Then at OpenAI again (mid-training and synthetic data work)

• Founded Eureka Labs to focus on AI education (2024–2026), where his videos have taught AI fundamentals to more people than probably anyone.

Now he’s joining Anthropic’s pre-training team.

Nicholas Joseph, the head of pre-training at Anthropic, wrote: “He'll be building a team focused on using Claude to accelerate pretraining research itself.”

Claude training Claude. The recursive loop 🔁. Andrej has been working on Autoresearch lately (82k GitHub ⭐s!), a project that uses agentic loops to optimize the performance of models automatically. Will he work on a scaled-up version of that?

Demis Hassabis is an Angel Investor in Anthropic 🏆🫡

This reminds me of how Jeff Bezos is an early investor in Google (he invested $250k in 1998) and, if he kept it, would be a billionaire even without Amazon.

The FT writes:

Sir Demis Hassabis, the founder of Google DeepMind, was an early investor in Anthropic, a previously undisclosed stake [...]

People familiar with the matter said Hassabis was an angel investor in Anthropic

And that’s probably not his only valuable personal angel investment:

Hassabis, who sold his London-based AI lab to Google for £400mn in 2014, is also an angel investor in businesses founded by former colleagues.

This includes Inflection AI, launched by DeepMind co-founder Mustafa Suleyman before he joined Microsoft in 2024, and Ineffable Intelligence, founded by longtime collaborator David Silver. [...]

Start-ups founded by former DeepMind employees include well-funded companies such as Mistral, Harvey, AMI Labs, Recursive Superintelligence, Latent Apps, Reflection AI, Cursive AI and Orbital Materials, among others.

Well, good for him. It does create an interesting conflict, as Anthropic is both a competitor and a big customer of Google, though. Not that he doesn’t have a much bigger stake in Google’s success, but it’s still a piece on the board. ♟️

I wonder if Demis was involved in the decisions at GCP to sell a boatload of TPU capacity to Anthropic a little while ago. 🤔

🧪🔬 Science & Technology 🧬 🔭

🤰📲 Is Fertility Dropping Because of Smartphones? 📉

I’ve long covered the fertility beat. It’s a big problem for humanity to be below the replacement rate in most places and still dropping. We could mitigate the problem by curing some of the diseases of aging and extending lifespans (and most importantly, healthspans), but that won’t replace creating new people.

There’s a very popular narrative right now that cellphones are to blame, and it makes a lot of intuitive sense. In fact, I'd be shocked if smartphones didn't have some impact. They reshaped dating, free-time use, and the nature of social gatherings. The carefree partying of the pre-pocket-camera age is gone. 📸

BUT

If we zoom out, we can also see that whatever impact came from Steve Jobs going on stage in 2007 is part of a larger process. 

The fertility decline has been running for over 150 years in some places, and was already rapid by 1970. And if we didn’t have Africa, the planet's average would look a lot worse:

The harder question is 'what would actually reverse the trend?' And I don't think anyone has a convincing answer yet.

h/t Tyler Cowen

🧠 Creative Practices Keep Your Brain Younger (Including StarCraft II) 🎮

This study asks: Do creative skills give you a younger-looking brain?

Researchers trained a brain-age model on EEG data from 1,240 people, then turned it loose on 232 with different flavors of creative expertise: tango dancers, musicians, visual artists, and StarCraft II players (Oh yeah!).

The metric they were looking for was "brain age gap." Does your brain activity look older or younger than your actual age?

Across all four creative domains, experts came out with negative gaps. Their brains "looked" 4–7 years younger than their chronological age, with an overall expertise effect of ~5.5 years.

There was also a small training arm: non-experts who trained on StarCraft II for 30 hours showed a ~3-year reduction in brain age gap, while an active control group showed nothing.

As an aside, the control group was trained on Hearthstone: “The Hearthstone game was selected due to its more rule-based and turn-based mechanics, with limited improvisation and creative play compared to StarCraft II’s real-time decision-making”.

There was a dose-response curve, so the more your game improved, the more your brain age gap did too.

The study’s proposed mechanism: creative practice strengthens or preserves connectivity in the brain networks most vulnerable to aging. Especially frontoparietal hubs (attention, motor control, coordination, visual processing). They also found higher local and global network efficiency, plus stronger biophysical coupling. The effects were bigger in long-term experts than in short-term learners.

Now let me throw a bit of cold water on this before you download Starcraft (well, you should anyway, or at least an instrument 🎸🎹).

Most of the expertise study is observational, so selection effects are still possible (maybe the kind of people who stick with tango for a while had different brains to begin with 💃). But however noisy the study is, it’s a pretty common-sense idea that creativity and learning difficult skills are good for you. It’s just nice to have one more reason to motivate us!

🚂 🎩 This LLM Only Knows the World Before 1931 📰 🕰️

This is a clever way to talk with the past:

we introduce talkie-1930-13b-base, a 13B language model trained on 260B tokens of historical pre-1931 English text. Additionally, we present a post-trained checkpoint turning our base model into a conversation partner without relying on modern chat transcripts or instruction-tuning data.

Very clever.

You might think pre-1931 wouldn’t give you much training data. We keep hearing that most text ever written by humanity was created in the past few years. But the talkie team is finding ways to scale it up significantly:

talkie is the largest vintage language model we are aware of, and we plan to continue scaling significantly. As a next step, we are training a GPT-3-level model, which we hope to release this summer. A preliminary estimate also suggests we can grow our corpus to well over a trillion tokens of historical text, which should be sufficient to create a GPT-3.5 level model—similar in capability to the original ChatGPT.

GPT-3.5 was the model used by the original ChatGPT, so while it won’t be anywhere near the frontier, it should be interesting to chat with (especially since modern algos and hardware should give it some advantages over the original GPT-3.5).

Why pre-1931?

Because, as of January 1, 2026, US works published through the end of 1930 are in the public domain.

One of the big challenges is to avoid ‘temporal leakage’.

That sounds pretty Doc Brown, but all it means is that in such a large pile of documents, some will have bad metadata, be mislabeled, mis-dated, of some old documents have modern footnotes. That kind of thing.

More details on Talkie here.

🎨 🎭 The Arts & History 👩‍🎨 🎥

🎞️ Treasures of the Past: Why You Should Make Your Kids Watch Old Movies 🍿

Last weekend was a three-day weekend in Canada.

On Sunday night, we had a family movie night. We watched Groundhog Day (1993, Bill Murray), which was my kids’ first time seeing it. And, well, my wife and I have no idea how many times we’ve seen that film because it’s almost impossible to say. 😅

After you've seen it once, you feel like you've seen it multiple times. After a few times over decades, the only way to tell if a specific memory is real is by checking how unfamiliar the very beginning and very end feel (they're the only parts that play once).

In any case, my kids loved it. It was a nicely teachable moment. We had a good discussion about the difference between doing things because you think that’s what others want, and doing them just because they are the right thing to do, to grow and help others. It sounds a bit trivial when you write it out, but it’s profound, and a great thing to discuss at that age.

Then on Monday, my wife went out with the 8yo, so the 12yo and I stayed home and we watched…. Arrival (2016, by my boy Denis Villeneuve! 🫎).

I was really looking forward to it, and it didn’t disappoint. I hadn’t seen the film in a few years and really enjoyed it, but most importantly, my son LOVED it. He said something like: “When I have kids, I’ll be showing this to them when they’re old enough!”, which is a cool sentiment and nice endorsement of the film.

He also said he wanted to rewatch it so that he can now see things differently, now that he knows the ending, and that he wants to show it to his little brother when he’s old enough.

It also generated some great discussions about time and language. But also, about why it’s important to go back and watch older films, if only because there’s going to be a lot more good ones if you pick and choose from the past 60 years or whatever than if, like many people, you mostly just watch recent things from the past 2-3 years.

Same for music, same for books. 📚🎶

I’m glad to be sharing these moments with my kids, and helping open their minds to a wider horizon in my own little way.

I see my age in my children.

—Dave Chappelle

🛀💭🦷 Invisalign as a microplastic delivery system?

I wonder if plastic aligners have been tested for how much microplastics they release while they’re slowly rearranging your face-bones.

At least they’re not exposed to high heat or UV rays too much, which are generally factors that make plastics degrade faster. But having had these things in my mouth for a year, it’s hard not to wonder what happens when you add up plastic + saliva + mechanical friction + time 🤔

(not that I’m one of those people who believe that microplastics are the root of all evil, but I also doubt they’re good for ya)

So I looked it up, of course. 📄 🔍

Turns out several teams have actually run this experiment. The 2025 PLOS One study found that Invisalign aligners released the fewest microplastic particles among the brands tested, but the ones it did release were the smallest (12–23 μm long), which from a “could it cross a cell membrane” perspective is arguably the wrong direction to be best at (😅). The 2023 saliva-immersion study found all seven brands they tested released microplastics within a week.

So yes, plastic + saliva + friction + time does what you’d expect.

🕷️👦 We found a tick on my youngest son’s head.

He told my wife he felt like he got bitten on the back of the neck and there was a little bump. Thankfully, instead of just telling him “it happens, it’ll go away”, she had a look and saw the tiny little bloodsucker (I hate these just as much as I hate mosquitoes 🦟).

Luckily, it wasn’t engorged and probably hadn’t been there long, but we didn’t take any chances and checked with a pharmacist, got a 1-dose preventive antibiotic, and kept an eye out for any other symptoms in the following days.

This little adventure got me thinking about how much parasites and the diseases they carry must’ve sucked for humans in pre-modern times.

People got fevers, rashes, joint pain, facial palsy, neurological weirdness, chronic arthritis, and fatigue. But nobody could point to a clear cause, so it was all just folded into vague categories like rheumatism, palsy, summer fever, arthritis, wasting, “bad blood,” or hand-wavy stuff like “he was never quite right after that fever.”

(Honestly, 'pre-modern' is generous. Lyme disease wasn't recognized as a distinct syndrome until the mid-1970s, and its bacterial cause wasn't nailed until 1981–82 by William Burgdorfer at the NIH. The 1975 Connecticut cluster that finally cracked the case was initially mistaken for juvenile rheumatoid arthritis until two mothers, Polly Murray and Judith Mensch, kept calling the health department because too many kids in their town were sick. Borrelia burgdorferi has been in New England for ~60,000 years; we figured out what it does to us about 50 years ago. 🕵️‍♀️)

🏦 💰 Liberty Capital 💳 💴

🤖🔐 AI Cyber Attack Capabilities: Doubling Every 4–5 Months (and Accelerating) 🔓🔑🏴‍☠️

It’s impossible to be precise about this, but it’s possible to get an idea of the rate of change, and this is what the AI Security Institute (AISI) has been trying to do:

In February 2026, we estimated that frontier models’ 80%-reliability cyber time horizon had doubled every 4.7 months since reasoning models emerged in late 2024, given a 2.5M token limit.

This was around half our November 2025 doubling time estimate, which was 8 months for both 50% and 80% reliability.

If a doubling every 4.7 months WASN’T fast enough for ya 😅

Claude Mythos Preview and GPT-5.5 have since significantly outperformed this trend. At the time of writing, it's unclear whether Mythos Preview and GPT-5.5 represent an isolated break from existing rates of progress or are part of a new, faster trend.

Mythos Preview and GPT-5.5 have large upper-bound error bars due to near-100% success rates on our narrow cyber suite’s longest tasks, even with the 2.5M token limit. Our tasks are also not long enough to determine how sharply the models’ reliability would deteriorate at higher task lengths. This places some of the latest models at the limit of what our narrow test suite can measure.

And the kicker: “Without the 2.5M token cap, success rates are so high that time horizons become impossible to calculate.”

😬

On April 7th, Anthropic wrote:

How long did it take us to get ready for Y2K? Patching all software is much more complex than fixing a single known bug.

Palo Alto Networks, one of the big security players, recently said:

For the last several months, we have had early, unbounded access to the latest frontier AI models. What we’ve seen from that vantage point has made it clear that the window for organizations to get ahead of what’s coming is shorter than most leaders realize.

We have moved past the era of incremental AI improvements into a threat landscape shift. Our testing has revealed a step-change in capability that demonstrates an intuitive understanding of software vulnerabilities. [...]

The latest frontier models, including OpenAI’s GPT-5.5-Cyber, Anthropic’s Mythos and Claude Opus 4.7, and the specialized variants emerging across major labs, represent roughly a 50% improvement in coding efficiency over their predecessors. That number sounds incremental, but in practice, it’s the threshold at which AI crosses from a helpful assistant into an autonomous operator.

They see four main categories that, when happening all at the same time, redefine the “modern threat landscape”:

• Vulnerability Discovery at Scale: Frontier AI is exceptionally effective at identifying vulnerabilities across massive, complex codebases. In [Palo Alto’s] testing, three weeks of model-assisted analysis matched a full year of manual penetration testing, with broader coverage.

• Exploit Chaining & Synthesis: What is more consequential than individual discovery is the models’ ability to think like an attacker. They link multiple lower-severity issues into single, critical exploit paths, seeing full-stack logic, including SaaS and public-facing surfaces, in ways traditional scanners cannot.

• Attack Cycle Compression: In AI-assisted scenarios, the time from initial access to exfiltration has collapsed to as little as 25 minutes. Detection and response measured in hours is no longer a viable standard; single-digit MTTR (Mean Time to Respond) is the new floor.

• The Unsupervised Attack Surface: Rapid AI development and decentralized innovation are creating a massive, unsupervised attack surface in real-time. As local AI agents become commonplace, every desktop is now effectively a server, yet most organizations lack visibility into the code their own employees are generating and deploying.

(Worth noting: Palo Alto sells security products, so they aren’t exactly a disinterested observer. But they also won’t do well if their customers get PWNED)

Even the NSA, an agency that surely has seen it all, is reportedly impressed:

Read more

To become different from what we are, we must have some awareness of what we are.

—Eric Hoffer

🌍🪵🏦 What’s rare and precious depends on your frame.

Gold, silver, platinum, palladium, heck, throw in rhodium, ruthenium, and iridium too. Or how about diamonds, rubies, emeralds…

But zoom out a little.

Gold is star-stuff. Platinum is star-stuff. Diamonds? Carbon + pressure. On Neptune, it literally rains diamonds. ☔

These are all over the universe, forged by nucleosynthesis in supernovae and kilonovae, the universe's heavy-metal foundry. 💥

But wood? Wood requires life. A very specific kind.

It’s way cosmically muuuuuuuch rarer than gold, even if it’s locally plentiful enough to burn for heat. 🔥

I’m not sure where I first heard this idea. Maybe in The Expanse? They nail the implication:

In that world, asteroid mining has made metals abundant enough to build whole off-world economies, but real Earth wood becomes an absurd luxury. A real wooden desk on a space station is a flex: a fossil of sunlight, air*, rain, and biology, dragged up the gravity well and carried across the void. 🌳🚀

*As Richard Feynman reminded us, trees grow mostly from the air, not the ground.

🔎📫💚 🥃 If you want Liberty’s Highlights to continue, this is the moment. Become a paid supporter.

Without your support, this steamboat sinks 🌊🚢⚓

🏦 💰 Business & Investing 💳 💴

📊 Amazon's Goodhart Moment: How Employees Inflate Their Token Usage

I knew this was coming. I wrote about it in Edition #628.

Amazon employees are using an internal AI tool to automate non-essential tasks in a bid to show managers they are using the technology more frequently.

The Seattle-based group has started to widely deploy its in-house “MeshClaw” product in recent weeks, allowing employees to create AI agents that can connect to workplace software and carry out tasks on a user’s behalf, according to three people familiar with the matter.

Some employees said colleagues were using the software to automate additional, unnecessary AI activity to increase their consumption of tokens — units of data processed by models. (Source)

Similar to Meta, Amazon introduced targets for more than 80% of its developers and has been tracking AI token consumption on internal leaderboards.

Predictably:

Amazon told employees that these metrics won’t be used for performance evals, but employees don’t seem to believe it.

This has to be going on in so many companies right now.

The minute you put a number on a leaderboard and tell people their bosses are watching, you've turned the metric into a target. Goodhart's Law doesn't take weekends off.

🪑 The AI Pricing Shift: Seats → Seats + Metered Usage 📊📈💳

Another “hey, I wrote about this!” 😅

One way that AI's all-you-can-eat buffet era is slowly ending: the change from per-seat pricing to seats + consumption. Microsoft’s CFO Amy Hood said this on the last earnings call:

suddenly, if you think about getting work done and being more productive, it’s thinking about being a seat or a worker plus an agent. And when I think about that model, I start to think about it as a license business plus a consumption business and really applying far more broadly than I think people have thought about that.

And so it starts to mean that over time bookings will actually also look a little different. It will still have that per seat license logic, but it will also have a meter, just like you see in Azure.

And it may not all flow through bookings in the same way. You’ll just bill for usage. And if that usage has great value to customers […] then you’ll keep spinning and keep using those agents if they’re adding direct value or growth to your business.

We’ll see if the transition is as long and messy as the on-prem vs SaaS transition, or if this will be speedrun through at AI deployment speeds 🤔

🇺🇸 Elon Musk’s $119bn Terafab Plans 🏗️🏗️🏗️ 🐜

It’s still mostly just an idea, a plan, but the numbers are pretty mind-boggling: According to filings, the first phase will cost “at least $55 billion” and the full thing is planned for “up to $119 billion”.

I suppose there’s a certain benefit to having extremely high-multiple stocks at SpaceX and Tesla. I mean, as I just wrote about, SpaceX’s IPO looks like it could come out at 265x adjusted-EBITDA. And Tesla? Revenue has been pretty much flat since 2023, yet the stock trades at 125x EV/EBITDA and at a P/E of almost 400x.

If you can print that currency, swapping a bunch of it to invest in a megaproject is probably the thing to do. And there are few megaprojects America needs more than leading-edge fab capacity.

Right now, the chosen process technology is supposed to be Intel 14A, so at least he’s partnering with someone with a long history of chip manufacturing and not entirely reinventing the wheel. But it’s not exactly an easy industry to get into, even with the right partners. Musk would be Intel's first external 14A customer for a node that's still under development. There's high execution risk on both sides.

But hey, this kind of hardware moonshot is a better fit for Musk than AI software. If this has any chance of creating a ‘USSMC’, I wish him luck! I’ll believe it when I see it, though.

💰 OpenAI Has Already Sent Microsoft ~$31bn in Revenue (and counting…)

Let's set aside Microsoft's OpenAI equity stake for a moment (the 27% it got for its $13bn investment, currently worth about $228bn). Just the revenue generated by OpenAI between ‘23 and ‘25 adds up to an estimated $31.2bn:

The $30 billion revenue estimate suggests Microsoft’s revenue from OpenAI-tied products has accelerated over the past year. As of March 2025, Microsoft had generated only about a third of that total—$9.5 billion—from its OpenAI partnership, executive Michael Wetter testified in a deposition last year in Musk’s lawsuit.

Microsoft could generate many times the $30 billion figure in the coming years, based on spending commitments OpenAI has made to use Microsoft services. [...]

OpenAI paid Microsoft the vast majority of the roughly $23 billion it spent on renting AI servers between 2023 and 2025 [...] the bulk of the revenue Microsoft has gotten from OpenAI-tied business from 2023 to 2025 comes from that $23 billion the AI lab spent on server rentals. Additionally, about $2.5 billion has come from Microsoft’s sale of OpenAI models to its cloud customers, an API known as the Azure OpenAI Service (Source)

Going forward, the recently renegotiated agreement preserves Microsoft's 20% revenue-share rate but caps total payments at $38 billion through 2030. That’s down from the prior deal that could have produced as much as ~$135bn. Based on OpenAI's growth, the cap is expected to be hit around 2028. But OpenAI has also contracted to buy $250B in Azure services…

🍁⚛️ Ontario Begins Early Work to Add 4.8 Gigawatts to the World's Largest Nuclear Plant 🏗️ 📐

Ontario is writing a CAD$300M “let’s get serious” cheque to Bruce Power for Bruce C, a proposed new nuclear station next to the existing Bruce A and B reactors on Lake Huron.

To be clear: this is pre-development money, not construction money.

This funds the homework: technology selection (no reactor design picked yet but there are five candidates: the new MONARK CANDU, Westinghouse’s AP1000, EDF's EPR, Hitachi-GE’s ABWR and BWRX-300 SMR), workforce planning, site-prep design, cooling water strategy, the federal Impact Assessment, and engagement with the Saugeen Ojibway Nation.

That's a long homework assignment 😅

This is actually Bruce Power's second swing at this. They filed and withdrew a similar Bruce C application between 2006 and 2009, then focused on refurbishment instead. So the precedent says pre-development isn’t a guarantee.

Bruce Power's current refurbishment program is a success story worth highlighting. Two of six Major Component Replacements are done: Unit 6 in 2023 and Unit 3 construction phase completed in February 2026, with return-to-service expected later this year. Both on budget and ahead of schedule, including through COVID. That's wildly atypical for Western nuclear (eg. Vogtle, Hinkley Point, Flamanville-3).

Back to Bruce C: If it all works out, we’re talking about 4,800 MW of new dispatchable, clean baseload capacity. It would make the Bruce complex the largest nuclear generating site on Earth by a lot.

(Bruce already holds the “largest operating” title; Japan’s idled Kashiwazaki-Kariwa is bigger by rated capacity at 8,212 MW. Bruce + C would land around 11,200 MW.)

The Ontario Ministry of Energy is claiming $238B in GDP contribution and 18,900 construction jobs. I don’t know if it’s inflated, but I know that Canada is one of the rare tier 1 nuclear nations with a full supply chain in the country, so if they pick the CANDU, the economic benefits will be large.

Worth zooming out: Bruce C is just one piece. Ontario is also doing the $12.8B Darlington refurbishment (Unit 1 back 5 months ahead of schedule), Bruce Life-Extension (Units 3-8 through 2033), Pickering B refurbishment ($6.2B), four BWRX-300 SMRs at Darlington ($20.9B, first SMR in any G7 country), and early planning at Wesleyville (up to 10GW).

The province is quietly running the most ambitious civilian nuclear build-out in the Western world. 🇨🇦🫎

With electricity demand climbing, this kind of dispatchable clean capacity is what attracts hyperscaler data centers and reshored/friendshored manufacturing. 🏭⚡️👨‍🏭

Energy is Life™️. Ontario typically runs ~60% on nuclear. The provinces and states that lock in big chunks of dispatchable clean baseload now will look very smart in 10 years.

🧪🔬 Science & Technology 🧬 🔭

⏱️🏃‍♂️ Humanity’s Journey to the Sub-2-Hour Marathon

It took a while, but we did it!

(and by “we”, I mean not me… My wife is running a marathon in two weeks, but I’ll be cheering her on from the sidewalk)

At this year's London Marathon, two runners broke the 2-hour barrier:

• 🇰🇪 Sabastian Sawe: 1h59mins30secs

• 🇪🇹 Yomif Kejelcha: 1h59mins41secs

Incredible!

This detail: Sawe's first half was 60:29 and his second half 59:01. He ran the second half faster than the first 🤯

Watch at least the finish. It gives me chills!

Can you imagine the mixture of pride AND disappointment for Kejelcha?

You do something historic… but 11 seconds behind the person who will be in all the history books. Usually, I’d say that you should compare yourself to your own potential, but in this case, it’s literally a race against others 😅

Using Codex, I made the graph above to visualize what it took to break the 2h barrier (it took me a while to get the graph to generate cleanly and to fact-check all the data points… AI’s not yet one-shotting this).

If we go further back, the first sub-3h was Johnny Hayes (👆) in 1908 at the London Olympics. Spiridon Louis was the first man to run a “marathon” under 3 hours: 2:58:50 at the 1896 Athens Olympics, but that race was about 40 km, not the official 42.195 km distance.

It then took over a century to grind from 3h to 2h! I wonder how far humanity can push it going forward 🤔

It’ll be interesting to see how long it takes for the 2h record to fall on the women’s side (if it happens, I mean, some records endure for decades). Ruth Chepngetich holds the women's marathon world record at 2h09mins56secs (Chicago, October 13, 2024), though she's currently serving a 3-year doping ban that expires in April 2028. The record stands because the positive test came months after the race 😅 The fastest "clean" mark behind it is Tigst Assefa's 2h11mins53secs (Berlin 2023).

💾 Patch Your Stuff + Google Catches First Mass-Exploit Attempt with AI-Developed Zero-Day 🏴‍☠️

There have recently been updates to iOS and macOS 📲 💻

The changelog lists 50+ security fixes for iOS 26.5 and ~70 for macOS Tahoe.

If you’re on Apple’s platform, make sure to keep up to date and install patches, and the same applies if you’re on Windows or Linux.

Apple doesn’t say if these patches were the result of working with Mythos to find security holes (Apple's a Project Glasswing partner), but while security is never NOT important (sorry for the double negative, but it sounds punchier that way ¯\_(ツ)_/¯ ), now’s a particularly great time to start taking it more seriously.

Google’s Threat Intelligence Group (GTIG) just announced that they had identified for the first time “a threat actor using a zero-day exploit that we believe was developed with AI.”

Kinda ironic: GTIG identified the AI-generated exploit partly because the Python script contained a hallucinated CVSS score. For once, AI hallucination saved the day.

To be clear, catching someone doesn’t mean it was the only one. This is just the tip of the iceberg, and it’s likely to be a pretty big iceberg because there’s just so much surface area out there, and there’s an asymmetry between attackers and defenders (the former just have to be right once to get in, while the latter has to be right every time to keep the perimeter secure).

In the long run, AI is likely to make cyber security better by patching holes and augmenting human analysts. But the road is likely to be bumpy on the way there 😬

Another finding from GTIG:

Threat actors now pursue anonymized, premium tier access to models through professionalized middleware and automated registration pipelines to illicitly bypass usage limits. This infrastructure enables large scale misuse of services while subsidizing operations through trial abuse and programmatic account cycling. [...]

By leveraging anti-detect browsers and account-pooling services, actors are attempting to maintain high-volume, anonymized access to premium LLM tiers, effectively industrializing their adversarial workflows while subsidizing their operations through trial abuse and programmatic account cycling.

On the psyops front, a new tactic that GTIG identified is using real media clips but modifying them with AI:

GTIG uncovered activity linked to the pro-Russia IO campaign “Operation Overload,” involving video content that leveraged suspected AI voice cloning to impersonate real journalists.

In identified instances, the actors appear to have manipulated an authentic video to convey a false message. This content appears to splice original vertical videos with montages and fabricated audio to create false and misleading messaging. The close voice match to the original suggests the use of AI tools

More details here.

🎨 🎭 The Arts & History 👩‍🎨 🎥

🤠🤠💰🚂 Butch Cassidy and the Sundance Kid : The Buddy Movie’s Original Chemistry 🚲 🇺🇸🇧🇴

I watched Butch Cassidy and the Sundance Kid (1969) with my kids. It was their first time. It’s one of my favorite films, but I was a little worried how they would handle the 1969 pacing. The full credits roll before a single scene, and there’s an extended travel montage that is just sepia photos.

But they LOVED it!

My oldest and I have been quoting lines back and forth all week. William Goldman, who famously said ‘Nobody knows anything’ about Hollywood, sure knew how to write memorable dialogue (see also: The Princess Bride, All the President’s Men).

What stuck with me on this rewatch: the bad guys are barely characters. The famous “Who are those guys?” posse is shot as a distant, mythic thing. Director George Roy Hill deliberately refused to give them faces. The real antagonist of the film isn’t a sheriff or a gunslinger. It’s the 20th century: railroads, payroll systems, Pinkertons with telegraph lines, institutions that never get tired and can keep coming after you. Charm doesn’t scale against institutions.

There’s a true story about the real Butch that Goldman loved. He called it “the best character introduction I ever came across,” but didn’t end up using it. Butch was offered parole by the governor of Wyoming on the condition he go straight. Butch basically told him: “I don’t want to lie to you, I can’t promise that. But I promise I won’t rob banks in Wyoming.” He got a deal. I mean, how charming do you have to be to get that offer? That episode is the whole story in miniature.

That film runs on Newman and Redford’s chemistry. It set the modern buddy movie template. Buddy duos existed before (Hope and Crosby, Abbott and Costello), but those were mostly vaudeville routines. Butch Cassidy established the version that endures: two leads whose rapport is the engine, the plot exists to give them things to do. 2+2=5. Countless others have copied the structure. Almost none have matched the chemistry.

Redford himself liked the role enough that he named his Utah land after the character. That land became the Sundance Institute, which became the Sundance Film Festival (you may have heard of it).

The freeze-frame ending perfects the trick. The film doesn’t deny their deaths. It freezes them as legend at the exact moment history was about to make them corpses.

I first posted this on OSV Field Notes. I encourage you to check it out and subscribe (it’s free). I think you’ll like it. It’s a weekly, high-signal curation of things worth your time. It’s overflowing with great recommendations of evergreen books, films, TV shows, and other interesting things.

I learned more from idiots and nobodies than from professors of this and that. Life is the teacher, not the Board of Education.

—Henry Miller, On Turning Eighty (1972)

🐦🤔😕 One of the reasons I don’t post as much to Twitter lately is because I have no idea what I’m doing when I’m posting anymore.

It used to be that I knew I was writing to a group of people who followed me, and who I largely followed back. A sub-community having an ongoing conversation about whatever was going on or what people were finding interesting.

But now, the default ‘for you’ page looks a lot more like TikTok than anything else, with random tweets that are clearly selected for broad engagement appeal rather than anything specific to me or the clusters I belong to (graph-based distribution was replaced by attention-based distribution, optimizing for doom-scrolling at the expense of community).

Don’t get me wrong, I still see tweets by people I follow, but they are now mostly one-way broadcasts, and there’s very little discussion going on when I click to see threads. If there’s a back-and-forth that is deeper than 2 levels, it’s kind of a special occasion. Back in the heyday of Twitter, it was common to have a few threads each day turn into huge convos involving many of my favorite accounts. Points, counterpoints, showing screenshots of charts and book passages, etc.

It was glorious.

Twitter was a discovery engine, both for interesting people and interesting things (books, films, podcasts, websites, newsletters, etc). I know you know. I’m not saying anything that hasn’t been said a thousand times. But I still kind of mourn it, even after all this time.

I’ve met many of my closest friends on Twitter. I’ve discovered many of my favorite films and books. I’ve learned about all kinds of things, from business to technology to history, on that stupid platform.

Twitter’s current incarnation isn’t as bad as when every tweet attracted 60 spambots, but it has morphed into a completely different thing. I feel like I don’t know how to post there anymore. Do I really want to talk about the latest book I read or film I watched with my kids and have that be sandwiched between posts about horrific Epstein abuse and reposted TikTok videos? 🫤

It feels a bit like everything you post goes everywhere and nowhere at the same time.

How does that work? ¯\_(ツ)_/¯

🏦 💰 Liberty Capital 💳 💴

🤖 Elon Musk Does 180 on Anthropic, Rents 220,000 Nvidia GPUs to Claude 🤔

I wrote a bunch about xAI lately. The co-founders leaving, the $1.25 trillion merger with SpaceX (which valued xAI at $250bn!), the lack of traction with users, the falling behind the frontier… and the $60bn Cursor Hail Mary offer to try to get things unstuck by injecting new talent and data (even though Cursor has never trained a frontier model from scratch).

They certainly aren’t boring. Whether the news is good or bad is a separate question. The latest development makes sense if you first consider two things:

1. Anthropic is STARVING for compute. They are growing usage faster than, well, pretty much anything ever at that scale, so they are desperate to find ways to fill that compute hole. Dario said they “planned for 10x growth and saw 80x” (I assume he meant in compute usage, not revenue).

2. xAI bought hundreds of thousands of expensive Nvidia GPUs, barn-raised a data-center at a speed that would make the Amish proud, trained what was at the time a frontier model on it, and then… 🦗🦗🦗

Pricey GPUs don’t earn their keep if they’re idling (xAI was reportedly running its cluster at ~11% utilization).

So this happened:

SpaceXAI has signed an agreement with Anthropic to provide access to Colossus 1, one of the world’s largest and fastest-deployed AI supercomputers.

Built from the ground up in record time, Colossus delivers unprecedented scale for AI training, fine-tuning, inference, and high-performance computing workloads. Colossus 1 features over 220,000 NVIDIA GPUs, including dense deployments of H100, H200, and next-generation GB200 accelerators. The cluster delivers extreme parallel performance for large language models, multimodal systems, scientific simulations, and generative AI at frontier scale.

This gives Anthropic access to 300 megawatts of additional capacity.

Wait. Doesn’t Elon hate Anthropic? 😅

Here’s just a sample of what Musk publicly said about them recently:

• Oct 2025: “Anthropic is misanthropic” and “Claude is pure evil.”

• Feb 2026: “Your AI hates Whites & Asians, especially Chinese, heterosexuals and men. This is misanthropic and evil. Fix it.”

• Feb 2026: “Grok must win or we will be ruled by an insufferably woke and sanctimonious AI.”

• Feb 2026: “Yeah, but we’re not super smug, sanctimonious and hypocritical about it like Anthropic is.”

• Feb 2026: “Anthropic hates Western Civilization.”

• March 2026: “Claude is woke and their logo looks like an anal sphincter 😂”

• March 2026: “Is there a more hypocritical company than Anthropic?”

I guess that’s water under the bridge, because he spent some time with them and his finely tuned evil detector (🤨) gave them the all-clear:

Musk also added: “We reserve the right to reclaim the compute if their AI engages in actions that harm humanity.”

That feels like a pretty subjective and broad clause, giving one party a unilateral 'we'll know it when we see it' trigger. Talk about a Sword of Damocles! 🗡️ A bit like OpenAI’s ‘AGI Clause’ in their deal with Microsoft (which has recently been removed).

Looking at everyone’s position on the board, it feels likely that this is actually a case of:

Read more

You cannot find peace by avoiding life.

—David Hare

🗓️😮‍💨 It’s funny: whatever time of year it is, my brain can always come up with reasons why life is ‘temporarily’ hectic… and then that season ends, and it’s something else.

Spring (🌱🌷🌧️🐦) is busy because there’s so much to do around the house. The pool, yard work, etc. And then there’s the kids’ birthdays, personal and corp taxes to file, and all that fun stuff.

Summer (☀️🌲🛶🔥) is crazy because the kids are out of school, family vacations, trips to my parents’ cottage (the photos above are from there), various social events, camp drop-offs and pickups, and holidays.

Don’t get me started on the Fall (🍂🍁🎃🥧). Back to school, yard work, and adjusting to shorter days, eternal darkness, and the cold — it always takes a bit of time. At our latitude in Canada, November feels like the darkest month. Once the snow comes, even if the days are technically shorter, at least things are brighter (and the Xmas lights help).

Winter! (☃️❄️🎄🎁) Well, forget about it. School breaks, flu season, holidays, snowstorms, family events, and travel. Very busy.

¯\_(ツ)_/¯

❤️‍🩹 📖 🛒 Dispatches from Grief by Danielle Crittenden is out on Infinite Books (also on Amazon). Mother’s Day is coming up 🎁

🔎📫💚 🥃 If you want it to continue, this is the moment. Become a paid supporter.

Without your support, this steamboat sinks 🌊🚢⚓

🏦 💰 Liberty Capital 💳 💴

🐜🔍 The AWS Chipmaker IPO? 🤔

When Amazon first started breaking out AWS’ numbers, Ben Thompson (💚 🥃 🎩) called it the ‘AWS IPO’.

Are we starting to see the AWS chipmaker IPO?

We don’t have the full picture yet, but Q1 gave us more details to paint the outline. Andy Jassy had this riff on the most recent earnings call about this gigantic business hosted inside of Amazon:

Andy Jassy: Our chips business continues to grow rapidly and is larger than what a lot of folks thought.

We saw nearly 40% quarter-over-quarter growth in Q1, and our annual revenue run rate is now over $20 billion and growing triple-digit percentages year-over-year, but this somewhat masks the size.

If our chips business was a stand-alone business and sold chips produced this year to AWS and other third parties as other leading chip companies do, our annual revenue run rate would be $50 billion. As best as we can tell, our custom silicon business is now one of the top 3 data center chip businesses in the world, the speed at which we’ve gotten here is extraordinary. And we have momentum.

Clearly, any viable source of compute is getting sold out. That’s just the reality of the tidal wave of demand. 🌊

How well would Trainiums and Gravitons do in a more normal market where customers evaluate each option on its merit, rather than take all they can get, almost regardless of price or performance?

¯\_(ツ)_/¯

Amazon has enough workloads that could be run on its own chips that they probably would be a great asset in any case (eg. they run a lot of the software that they sell on their own chips, and customers don’t have to know that their database runs on Gravitons instead of AMD CPUs).

BUT

Let’s imagine a world where Anthropic isn’t so compute-constrained that their uptime chart looks like an autumn forest painting. Let’s also imagine that they have internal cashflows that they can use to pay for compute, and they don’t have to raise billions from hyperscalers like Google and Amazon that bundle use of their chips in the deal.

Would they still use gigawatts and gigawatts of Trainiums and TPUs, or would a bigger share go to Nvidia because the performance/$ or perf/watt is better?

I don't know, but looking at the incentives, my guess is yes, more would flow to Nvidia.

For our custom AI silicon, we’ve recently shared very large multiyear, multi-gigawatt Trainium commitments from the 2 leading AI labs in the world in Anthropic and OpenAI as well as an increasing number of companies like Uber betting on Trainium.

And we now have over $225 billion in revenue commitments for Trainium. Our Trainium 2 chip has about 30% better price performance than comparable GPUs and is largely sold out. Trainium 3, which just started shipping at the start of 2026 and is 30% to 40% more price performance than Trainium2 is nearly fully subscribed. And much of Trainium 4, which is still about 18 months from broad availability has already been reserved.

These are big bucks, but on the performance numbers, I can’t help but see the 30-40% improvements in price-performance as being a little low compared to the generation-over-generation improvements we’re seeing from Nvidia 🤔

(Nvidia claims roughly 2x performance/$ from Hopper to Blackwell for training, and the inference numbers are even more dramatic at rack scale, when you take into account the whole system that includes networking, software optimizations, etc.)

We also just announced that Meta is committed to using tens of millions of Graviton cores. Graviton is our industry-leading CPU chip, which allows Meta to run the CPU-intensive workloads behind agentic AI with the performance and efficiency they need at their scale. AI is commonly seen as a GPU story, but the rise of agentic workloads, real-time reasoning, code generation, reinforcement learning and multistep task orchestration is driving massive CPU demand as well. As AI systems shift from answering questions to taking actions and as post-training and inference scale up, the compute required pulls heavily on CPUs. That’s why Meta chose Graviton, which delivers up to 40% better price performance than any other x86 processors and now used by 98% of the top 1,000 EC2 customers.

Meta is truly the compute swing voter. They’ll use everything!

Nobody has a better set of chips across AI and CPU workloads than AWS with Trainium and Graviton, and we’re unusually well positioned for this AI inflection we’re in the early stages of experiencing.

He goes on to say that their partnership with Nvidia is also very important (and clearly it is).

But their ROIC will increasingly depend on their own silicon efforts. The key challenge: steering the big labs to get to a critical mass on their chips so they get past the friction of porting their stuff over AND so that AWS can get to a scale that allows their chips to be viable contenders and more attractive to other potential customers too. Or to go to the overused metaphor: Get that flywheel spinning faster.

Tokens Schmokens: We Need a Better Economic Unit of AI Work 🤔

Usage-based pricing measured in tokens is problematic because tokens are not a stable unit of anything, as I wrote about in Edition #630.

They’re not like kilowatts or gallons of gasoline.

And paying by the token creates perverse incentives for labs to slice the pizza into more slices. 🍕

By that I mean that the way models convert text and data into tokens is not a fixed process that is the same everywhere. Different models can have different tokenizers that will parse the same source into more or fewer tokens.

My friend MBI (🇧🇩🇺🇸) had this great metaphor:

It’s like two movers quoting you a rate per box. Mover A quotes you $10 per box but uses large boxes. On the other hand, Mover B charges you $8 per box, but uses small boxes. [...]

As you can see, tokenization is a non-transparent billing unit that each vendor controls unilaterally. If a provider bills for hidden reasoning or other opaque internal operations, the customer may be paying for a large share of compute they cannot directly observe or verify.

But the size of the box isn’t the only variable. Some boxes are made from solid gold and others from paper-thin recycled cardboard.

A token generated by a tiny model, from a simple query with minimal context, can be extremely cheap to generate. The small model can run on old GPUs, use little RAM, and the answer can be one-shotted quickly.

But a token generated by the largest model out there, on a very complex query with a huge amount of context (Dr. Evil voice: 1 million tokens!) can be a seriously expensive token that can only come from a beefy modern GPU with lots of scarce high-bandwidth RAM.

Add to that how many more tokens per query you get now that models are reasoning and agentic, and the cost variability compounds. And you have very different incentives with flat rates and usage-based billing.

For a flat rate, the lab is probably incentivized to generate fewer tokens, to spend less time reasoning, do fewer tool calls, and to load fewer sources (which increases context and thus cost, etc). 🗜️

But if you’re paying by the token, the lab has an incentive to have the model work for a long time, load a lot of data in context, and generate a lot of reasoning tokens because that will result in more revenue. 💰💰💰

It seems like tokens as a unit of work are not tenable for much longer, because it’s impossible to know what you’re talking about when you say “10,000 tokens” unless you add a lot of context.

Maybe the labs like it that way. It makes it easier to obscure things and pull levers to increase margins and revenue. 📈

But those who are spending billions on AI will probably push for a more transparent unit of work, a ‘slice of intelligence’ (🧠). Something that embeds all the variables I mentioned above into a clear number. For more vertical applications, outcome-based pricing may work (like how digital advertising tracks sales), but the labs are too horizontal and general for that.

I’m not sure what exactly that looks like. It’s tough to normalize when not everybody pays the same for GPUs, ASICs have very different performance profiles, electricity costs vary, and comparing across GPU generations isn’t just about FLOPs but also networking and scaling out. But maybe someone smarter than I am can figure something out that works well enough to be worth adopting.

Update: After I wrote the above, Ben Thompson (💚 🥃 🎩) published an interview where Sam Altman says the following:

Do you think we stick with pricing as far as — pricing is based on tokens, does that make sense in the long run?

SA: No. And in fact, like there was an interesting example of this with our model that just came out, 5.5. where the per-token cost is much higher than 5.4, but it requires a hugely fewer number of tokens to get the same answer, and you actually don’t care about how many tokens the answer takes, you just want the piece of work done, and you want again a price and an amount of capacity you can have for that.

So maybe I was wrong to say “token factory”, but we’re like an intelligence factory or something. We just want as many units of intelligence for the lowest price and whether that is a bigger model running fewer tokens, a smaller model running lots of tokens, whether a GPU or Trainium or something else, whether we do any of the other kind of number of things we could do about that creatively, I don’t think customers care.

So Altman agrees tokens are the wrong unit. The hard part is figuring out what comes next.

🍎 Apple Shops for a Second Chipmaker, Looks at Intel and Samsung U.S. Fabs 🐜🇺🇸

Bloomybergy:

Apple Inc. has held exploratory discussions about using Intel Corp. and Samsung Electronics Co. to produce the main processors for its devices in the US, a move that would offer a secondary option beyond longtime partner Taiwan Semiconductor Manufacturing Co. [...]

Neither effort has resulted in any orders so far, and the work with both suppliers remains preliminary, according to the people, who asked not to be identified because the talks are private. Apple has concerns about using non-TSMC technology and may not ultimately move forward with another partner, the people added.

This last part is not nothing.

But Apple is already feeling the capacity squeeze: Tim Cook said on last week's earnings call that they have less supply chain flexibility than normal, largely because AI datacenter buildouts are eating TSMC capacity.

While on paper it’s amazing to have a second-source, especially one in the U.S., in practice what matters most is that iPhones are as good as they can be and that everything runs like clockwork. The annual release cadence is brutal, there’s little margin for error, and the scale required is large. 🏭→📱📱📱📱📱📱

The last thing Apple wants is delays, chips that don’t perform well, or dual-sourcing and having some iPhones or Macs with the “good chips” while the others have inferior ones. They had this exact problem when they dual-sourced LTE modems from Qualcomm and Intel during the iPhone 7 era. The Qualcomm modem was so much better that Apple actually throttled it down to match Intel's performance, so all iPhones would be the same. 😅

It wasn’t the end of the world, but I think it would be a bigger deal with the SoC, the package that includes the CPU + GPU + XPU (RAM is on the same package, though not same die).

Maybe the place to start would be a lower volume product, like some Mac chip 🤔

But even that could be a gamble. On the other hand, if they don’t start to find other suppliers now, if anything ever disrupts TSMC operations (I’m looking at you, China 👀)… yikes

🗣️ Will Guidara: Transferable Lessons from Hospitality 🍽️🌭

I enjoyed this one so much. Will’s voice and stories were like ASMR tickling my brain just the right way. 💆‍♂️

A lot of the wisdom he shares about hospitality applies to other jobs and aspects of life, but most people probably never think to make the transfer.

The anecdotes are great: the hot dogs served at Eleven Madison Park, a Michelin star restaurant, or the UPS store that requires employees to comp one customer’s order per day, or the pilot who gives tours of the cockpit to passengers when there’s a delay. I should do more of this kind of stuff in my life 🤔

Q: What are some of the other most unreasonable things that you’ve ever done for a guest?

Will Guidara: Oh my gosh, we did so many crazy things. One of my favorites is there was a family of four from Spain dining with us, parents and their children. We had these big windows. It started snowing. We learned that it was the kids’ first time seeing real snow. [We] somehow found a store still open at 8 o’clock on a Friday night selling sleds. And when they left the restaurant, we had an Uber SUV parked out front with sleds in the back [and] a big thermos of hot chocolate in the front.

Beautiful! 🛷

Here’s the UPS store one:

Will Guidara: This is something I learned from a guy who read my book and was inspired by it and tried to figure out a way to put it to work in his world, but his world is very different from the one I came from. He owns two UPS stores in Sarasota, Florida, which are not the places anyone would immediately think of as being the most hospitable in the world.

But he wanted to figure out how to put this into practice. And so he came up with an idea and he shared it with me because he said it transformed his entire culture. He made a rule that everyone that worked for him that worked, that worked the register, they were required to once a day, one time during their shift, comp someone’s order up to whatever, 30 bucks. That one change transformed the culture of his stores.

And it was a win, win, win.

It was a win for the person on the receiving end. Imagine if you went to a UPS store to like send something to your mom and they’re like, sir, it’s on us today. You’d be like, what the fuck is going on? It would blow your mind. And you would talk about that. Like, you’d be like, dude, the weirdest thing happened. I was at the UPS store.

It was a win for the people that worked there. Because, listen, I don’t care what you do. I don’t believe you can do it well if you don’t feel some level of genuine appreciation for doing it. […] the people that worked there, when they did comp someone, they’re like, oh my gosh, they really appreciated that. This feels great.

That’s what led to the third one. Because when he implemented the rule, they were required to do that. But after they started feeling that level of appreciation, now they were allowed to do it. But only one time a day. Which meant they had to now work to more deeply understand everyone that walked through the doors of that store to decide who deserved it the most.

Was it someone having a good day that needed a cherry on top? Was it someone having a bad day that needed something to go right? Every customer was receiving better service and hospitality in pursuit of figuring out who was going to get that one gift.

Such a clever incentive!

I bet it made the store perform so much better in all kinds of other metrics too, because once you have engaged employees who are paying attention, everything will improve 📦📦📦

If you’ve seen the TV show ‘The Bear’, a lot of what he says reminds me of my favorite episode, ‘Forks’.

🧪🔬 Liberty Labs 🧬 🔭

📺 Steel to Blade: Irish Knife Maker Fingal Ferguson 🇮🇪

This is a neat sub-10-minute mini-documentary that was produced by Stripe.

The theme of the series is tacit knowledge, something that I discussed with my friend Cedric Chin on this podcast. In fact, Stripe cites Cedric as inspiration for this series!

But regardless of high-concept ideas, it’s just really fun to watch a master craftsman turn an undifferentiated piece of steel into a beautiful artifact. 🪨→🔪

🔌 Global AI & Datacenter Electricity Consumption: Unevenly Spread ⚡

Hannah Ritchie looked into the latest available numbers (2025), and sliced and diced them all kinds of interesting ways. Here are my highlights:

Global data center electricity use is only about 1.5% of the total. But the U.S. is the epicenter of AI, hyperscalers, and the Internet generally, so the share there is much higher:

But even those numbers hide regions where the % are much higher:

this demand is even more locally concentrated. Beneath Europe’s 1.6% figure, we have Ireland, where data centres account for more than 20% of its electricity consumption. Beneath the 5% US figure, there are a number of states where data centres make up more than 10% of demand, and in states such as Virginia, it’s more than [25%].

The U.S. share of new electricity demand is growing much faster than elsewhere:

But these numbers also hide other factors. The share in the U.S. is so much higher in part because data-center growth is expected to be very high, but also because other sources of electricity demand are expected to grow more slowly than elsewhere.

Like what?

Well, EVs are not growing as fast in the US as in China, and A/C is already very common in the U.S. while it’s growing faster elsewhere. 🔌🚘🔋

Last one, the difference in energy use between typical one-shot LLM queries and longer-running agentic workflows:

But this is in constant flux, so don’t take those numbers too literally. The software gets optimized, more efficient GPUs come out… But also, new workflows and capabilities come out that make users generate way more tokens. So there’s a push and a pull. ↔️

What matters more is probably the general orders of magnitude.

Check out Hannah’s full post for more details.

📺 😠 Google’s Terrible YouTube Parental Controls

You’d think a company as technically competent as Google would be able to create parental controls that work on one of their flagship platforms (YouTube), but that doesn’t seem to be the case.

My oldest son doesn’t have a phone or tablet yet, but he watches YouTube on a TV. In the past, it’s mostly been Minecraft and Zelda videos, but in recent times, YouTube has been pushing more short-form videos on him in all kinds of ways.

We’ve discussed it, and I’ve encouraged him to stick to more substantive things (even if they’re just entertainment), and he agrees. But these things are very engaging (that’s their whole point), and once you start, it can be hard to stop (just one more!).

So recently, when I saw that a new parental control setting had been added by Google to allow shorts to be time-limited or even set to zero minutes per day, I discussed it with my son. I expected us to figure out what a reasonable limit per day would be (15 minutes?), but he asked me to set it to zero. I was surprised, but happy that he had the introspection to realize that he wasn’t getting much out of these things, even if they were hard to resist.

WELL

That was a very nice theory, wasn’t it?

I set the limit to zero. I set it to 15 minutes. Doesn’t matter, it doesn’t work. 💣

Apparently, Google’s ‘Family Link’ parental controls are so broken that they work on some platforms and not others. On our smart TV and in the browser on a Macbook Air, *both* logged into my son's YouTube account, these controls do absolutely nothing.

I don’t even know where they work at this point. Only on Android? Come on, Google, you have hundreds of millions of users that aren’t on Android.

This is so stupid. Don’t people at Google even test these things? Or at least make it very clear on the settings page what the limits are, and don’t bury it in the small print.

How many parents like me are using parental controls, thinking they just did something, but they never go check, and things simply don’t work.

You can also try to set limits using other methods, like Apple’s ScreenTime, but that only works on macOS and iOS, not on a TV’s YouTube app, and it’s not as granular as allowing YouTube but not YT Shorts.

You’d think that maybe you could go into the TV’s YouTube app settings to tune things the way you want them, but nope, there’s almost nothing useful in there.

You can try to nudge the YouTube frontpage algo by hitting “show me less like this” and “not interested”, but shorts always come back.

Don’t you dogfood your own products, Google? Don’t your engineers have kids that aren’t on Android devices?!

🎨 🎭 Liberty Studio 👩‍🎨 🎥

How to Cheat at… Guitar?! (The Tricks that Fool You) 🏴‍☠️🎸

Wow, I had no idea some people were that sophisticated at pretending they’re great guitarists. I suppose that since there’s a lot of attention and money involved in getting engagement on YouTube, it was bound to happen.

I love how the video goes through various levels of cheating, from pretty benign to full-blown AI persona. The tricks of the trade to subconsciously convince you that something is real (like the visible laptop recording, the clock running in the background, etc) are clever.

But once you know what to look for, you’ll be a more skeptical viewer (not just for guitar videos, but for anything else online). 🧐 🤨

On the Internet, people on the tails of the bell curve can find one another.

—Virginia Postrel

📷🎨🤖 OpenAI recently released ChatGPT Image 2.0.

One way I’ve been testing it is with old family photos. I asked my mother for old photos of her parents and her childhood, and used the new image model to “remaster” them.

Of course, it’s extrapolating details and colors, and you can never be sure how close the result is to what things actually looked like. But it’s still a very fun exercise, and if you steer the model to retain exact facial features and such, it appears to do a pretty good job.

You can ask it to recreate the photo as if it were shot with a specific camera (I don’t want to pay for a top-of-the-line Nikon or Sony, or a vintage Hasselblad, but I can emulate them!), or on a certain type of film, so I’ve been fiddling with those variables to see what looks best.

But most importantly, my mother seems to really enjoy it. ♥️

⚔️🥚🛡️🏰🐴 In case you missed it, I have a new podcast, and I hope you’ll enjoy it as much as I did recording it:

• 🎧 A Knight of the Seven Kingdoms: Going Deep with Tinkered Thinking and Jameson Olsen

🔎📫💚 🥃 Last call: As I explained in Edition #628, prices are going up for the first time on May 1st (this Friday!) to catch up to six years of inflation. You can lock in the old 2020-era pricing for a year by becoming a supporter now. ⏳

Without your support, this steamboat sinks 🌊🚢⚓

🏦 💰 Liberty Capital 💳 💴

OpenAI and Microsoft Alter the Deal, Again 🤝

When Microsoft first invested in OpenAI back in July 2019, it was very clear who had the leverage. OpenAI’s tech didn’t do much yet, unless you wanted a really good bot to play Dota 2.

Microsoft had the money, the compute, and the distribution channels. In exchange for their $1bn investment, which was mostly going to be spent on Azure anyway, they got exclusivity, joint development of AI supercomputing infra, rights to commercialize OpenAI’s IP, and economic exposure to the for-profit side of things. The integration went deeper in 2023 with a $10bn follow-on, giving Microsoft 75% of OpenAI’s profits until recouping its investment. These rights eventually converted to a 27% equity stake in 2025.

But over time, as ChatGPT exploded in popularity and pulled OpenAI into hypergrowth, the negotiating equilibrium changed. There was a moment when Microsoft had second thoughts about keeping investment pedal-to-the-floor, which allowed OpenAI to further wriggle out of the exclusivity straitjacket and go fully poly with everyone else.

This apparently reached a new stage:

The companies also scrapped a controversial clause in their deal that would have granted Microsoft a share of OpenAI’s revenue and certain IP rights up until OpenAI achieved “artificial general intelligence,” or AI on par with a human.

Microsoft will now keep getting a share of OpenAI’s revenue until 2030 regardless of whether the startup achieves AGI, the companies said. Microsoft will also retain the rights to use OpenAI’s models and products until 2032, but Microsoft’s rights to the IP will no longer be exclusive.

Microsoft will also stop sharing some of the revenue it makes from selling OpenAI models on Azure. Up until now, Microsoft shared 20% of the revenue from those sales with OpenAI, while OpenAI shared 20% of its total revenue back to Microsoft. OpenAI will keep sharing revenue with Microsoft through 2030, the companies said Monday.

Looks like this was necessary to, among other things, allow OpenAI to sell its APIs on Amazon AWS.

The value of this freedom of motion kept increasing for OpenAI as Anthropic’s explosive revenue growth showed how well coding could be monetized. In return, Microsoft got cleaner economics and protected its downside by removing the AGI escape clause.

Back when OpenAI and Microsoft first made their deal, it was probably expected that the ChatGPT consumer business would be the main engine, but the API and enterprise business turned out to be the real revenue engine (at least for now).

Musk’s AI Reboot: Thoughts on xAI SpaceX AI & Cursor 🤖🚀📡🛰️💾

Speaking of Microsoft, apparently, they looked at buying Cursor but decided not to proceed before this was announced:

This looks like a Hail Mary by the AI laggards to band together and hope that they can make something happen.

Cursor plans to train its latest AI coding model, Composer 2.5, on xAI infrastructure, the people said. Cursor will use tens of thousands of xAI's graphic processing units (GPUs), the chips used to train AI models, they said.

I mean, these GPUs were probably not doing a lot anyway ¯\_(ツ)_/¯

In a memo to staff last week, Nicolls, xAI's president, said the company's model FLOPs Utilization (MFUs), a measure of how efficiently a GPU is used during AI training, was "embarrassingly low" at about 11%. Nicolls said he aims for the team to reach 50% in the next few months.

xAI did incredible things, building a huge GPU data-center in record time by MacGyvering an old appliance factory with gas turbines on trailers tapped into a nearby pipeline. Very impressive, and thanks to all this compute, for a brief moment, they had a very competitive model (around Grok 4).

But users never materialized. They couldn’t keep up when coding and agentic stuff became the new hotness, and pretty rapidly, *all* of xAI’s co-founders not named Elon jumped ship.

Musk then vowed to rebuild xAI and do it right this time (😅), but it’s unclear what the path from here to there is.

Even with Cursor, it’ll be hard to compete on the coding/agentic front with Anthropic and OpenAI. Cursor may have some nice data assets, but they’re still heavily dependent on third-party models, which the labs get earlier and cheaper, and subsidize on the flat-rate plans vs the API, making it hard to compete.

In other words, Anthropic researchers are building the next model using the unreleased Mythos, burning way more tokens than they could afford if they were paying the API prices (because they don’t pay themselves a gross margin). Neither Cursor nor xAI has that advantage — they’re trying to build something either using their own inferior in-house models, or using Anthropic and OpenAI’s public models, paying their competitors a margin and helping them grow.

[Cursor]’s revenue hit $2.7 billion in annualized sales last month, said one of the people, up about 14 times from a year ago. Some investors expected annualized revenue to hit more than $7 billion by year’s end, said two of the people.

Ok, so they’re still growing fast, BUT:

Cursor’s gross margin was negative 23% as of the quarter ended in January, according to two people with direct knowledge of its financials. [...]

Cursor’s gross margin could look even worse than it does. The company classifies the cost of running servers for non-paying users as a sales and marketing expenses rather than costs of goods sold, according to financial documents The Information viewed. [...]

If the cost of inference compute for free users was included in COGS, it would have decreased Cursor’s gross margins for the quarter ending in January from negative 23% to about negative 31%

Yes, it’s a lot easier to grow fast when you’re selling dollars for 70 cents.

And they don’t have the frontier labs’ excuse that training is super expensive, since their own models are not nearly as expensive to train (they take Moonshot’s Kimi K2.5 and then add some of their own frosting to it).

It’s also doubtful that the best AI engineers and researchers in the world will want to go work at xAI, especially after what they saw happen to all of the technical co-founders.

Musk’s primary advantages seem to be:

• Raising money and leveraging his other companies

• Speed of execution, especially for physical infrastructure

Overpaying for Cursor and using cheap SpaceX equity to do so seems to fit #1. A kind of big acqui-hire and buying a user base in coding. But Cursor isn’t the shiny leader it was not so long ago: Its star has been falling against Codex and Claude Code.

Is Musk going to top-tick a structurally disadvantaged asset that won’t be enough to move the needle? Maybe it’s the least bad option if xAI’s compute is under-utilized and almost un-monetized right now, and Musk doesn’t want to throw the towel and just rent compute to Anthropic and OpenAI. He’s looking for first-party workloads AND for something to reboot his team. Cursor has some experience training models based on Open Source models, but I don’t think they are quite at the level of the big labs.

So, where will Musk find a core frontier team? Is he just going to use lots of SpaceX equity/IPO funds to massively overpay for researchers and do what Zuckerberg did?

Maybe that’ll work, but that probably has a clock on it. I suspect that as soon as Cursor’s top engineers and whatever top AI researchers he recruits with promises of a billion dollars and great soup have earned out, they’ll take the money and run. 🍜💰

⚓🍤 AI’s All-You-Can-Eat Buffet Era is Slowly Ending 🐟

GitHub just announced:

Why?

Why we’re making this change

Copilot is not the same product it was a year ago.

It has evolved from an in-editor assistant into an agentic platform capable of running long, multi-step coding sessions, using the latest models, and iterating across entire repositories. Agentic usage is becoming the default, and it brings significantly higher compute and inference demands.

Today, a quick chat question and a multi-hour autonomous coding session can cost the user the same amount. GitHub has absorbed much of the escalating inference cost behind that usage, but the current premium request model is no longer sustainable.

Usage-based billing fixes that. It better aligns pricing with actual usage, helps us maintain long-term service reliability, and reduces the need to gate heavy users.

Anthropic has also been making moves in that direction. It’s a bit confusing to try to untangle it, but as far as I can tell, Enterprise plans with 150+ users are moving from flat per-seat rates (up to $200/user/month) to a $20/seat base fee plus usage-based billing on top.

Will what started with larger enterprise plans eventually migrate to individual plans? That seems very likely.

OpenAI still seems to have enough compute to use it as a weapon. Codex has MUCH higher limits than Claude Code on a similarly-priced plan, it’s ridiculous. I can do tons of things in Codex and never run out, while I often go into extra pay-as-you-go usage with Claude.

But I suspect that, if things go well for them, even OpenAI will eventually have to become less generous and make sure that the whales aren’t eating so much that they generate deeply negative margins.

OpenAI’s Nick Turley, head of ChatGPT, said:

“it’s possible that in the current era, having an unlimited plan is like having an unlimited electricity plan. It just doesn’t make sense.”

The writing is on the wall. It’s not company-specific, it’s structural.

🧪🔬 Liberty Labs 🧬 🔭

🐜🐜🔍 Google's Two-Chip TPU Strategy: One Chip to Train, Another to Serve

As Nvidia is starting to add more SKUs and split its product lineup a bit more, Google is doing the same with its new TPU, with the TPU 8t for training and TPU 8i for inference:

Training: TPU 8t utilizes our proven 3D torus network topology at an even larger scale of 9,600 chips in a single superpod. TPU 8t is designed for maximum throughput across hundreds of superpods [...]

TPU 8t introduces native 4-bit floating point (FP4) to overcome memory bandwidth bottlenecks, doubling MXU throughput while maintaining accuracy for large models even at lower-precision quantization [...]

With JAX and Pathways, we can now scale to more than 1 million TPU chips in a single training cluster.

Training price-performance: Google claims that TPU 8t delivers up to 2.7x performance-per-dollar improvement over Ironwood TPU for large-scale training.

That’s a lot, though it’ll be interesting to see how that compares to Blackwell and Rubin on the Nvidia side (Blackwell was a really big jump over Hopper).

Inference: With 3x more on-chip SRAM over the previous generation, TPU 8i can host a larger KV Cache entirely on silicon, significantly reducing the idle time of the cores during long-context decoding. [...]

By integrating a specialized CAE, TPU 8i further reduces the on-chip latency of collectives by 5x. Lower latency per collective operation means less time spent waiting, directly contributing to higher throughput required to run millions of agents concurrently. [...]

This is what that crazy spiderweb of ASICs looks like: 🕸️

Utilizing a high-radix design, we connect up-to 1,152 of these chips together, reducing the network diameter and the number of hops a data packet must take to cross the system. By slashing the hops required for all-to-all communication (the heart of MoE and reasoning models), Boardfly achieves up to a 50% improvement in latency for communication-intensive workloads.

Inference price-performance: Google claims that TPU 8i delivers up to 80% performance-per-dollar improvement over Ironwood TPU, particularly at low-latency targets for large MoE models.

Broadcom designed the TPU 8t, and MediaTek designed the TPU 8i. Google is diversifying its silicon supply chain.

They’re pairing those TPUs with their own Axion Arm-based CPUs, which are based on the Neoverse V2 design, replacing the x86 CPUs used in previous generations.

Specs look like this:

🤖 First Thoughts on GPT-5.5 (and Opus 4.7 Follow-Up)

You barely have time to get used to a new frontier model before a new one comes out, these days. I was starting to sour a little on Opus 4.7, preferring Opus 4.6 on the ‘flavor/vibes’ front when GPT-5.5 burst on the scene like the Kool-Aid Man.

It’s very smart on the benchmarks, taking the lead in many fields (though not all! Opus still mostly leads on coding/software-engineering/tool use, even if GPT is improving).

Clearly, they are targeting the same market as Anthropic now, with maybe a bit more scientific flavor, which reminds me of DeepMind:

The gains are especially strong in agentic coding, computer use, knowledge work, and early scientific research.

But, well, Mythos is still in the wings:

Nvidia has announced that 10,000 of its employees got early access to GPT-5.5:

Debugging cycles that once stretched across days are closing in hours. Experimentation that previously required weeks is turning into overnight progress in complex, multi-file codebases. Teams are shipping end-to-end features from natural-language prompts, with stronger reliability and fewer wasted cycles than earlier models.

So this model was trained on Nvidia Blackwells, is served on Blackwells, and is helping Nvidia produce and design the next generation of GPUs. 🔄

So far, my experience with GPT-5.5 is very positive.

I didn’t expect it to be this fast. I’m not sure if it’s because it’s being inferenced on a new fleet of Blackwells 🤔

OpenAI revealed that the model was trained on new Nvidia chips: “GPT‑5.5 was co-designed for, trained with, and served on NVIDIA GB200 and GB300 NVL72 systems” and that they used it to help optimize their inference infrastructure. Maybe that’s what I’m seeing!

But whatever the reason, I find 5.5 fast enough that I now always leave it in ‘extended thinking’ mode, while with GPT-5.4, I switched back and forth depending on the question. I could be getting responses faster, but I’d rather trade that time for even better responses.

The model doesn’t have much flavor, it’s a little bland, but at least it’s not annoying either. I’ll take it. ¯\_(ツ)_/¯

Pricing is higher than 5.4 per token (but as I wrote earlier, that only tells you so much).

In fact, the vanilla 5.5 is a little higher than Opus 4.7, and GPT-5.5 Pro is significantly higher (note that this is a log chart, because if I had kept it linear, Pro’s output pricing would be in its own zip code):

But as we mentioned earlier, token use for the same actions can vary a lot across models, and 5.5 generally seems more efficient than 5.4.

Pancreatic Cancer mRNA Vaccine Shows Promising Early Results 🧬👩‍🔬🧫

Pancreatic cancer is *terrible*. The survival rate is low, it kills people quickly, and is very resistant to immunotherapy.

Less than 13% of people diagnosed with pancreatic cancer live for more than five years, making it one of the deadliest cancers. There is no routine screening for pancreatic cancer, such as colonoscopy or mammogram, and symptoms typically don’t show up until the disease is advanced. Once detected, there are few options for treatment. Only about 20% of cases are operable, which is currently required for someone to be eligible to join a pancreatic cancer vaccine trial.

Every time I hear of it, I think about Steve Jobs 😢

(His was actually a neuroendocrine tumor, the rarer, slower-growing type)

There is a vaccine called autogene cevumeran (also BNT122). It’s developed by BioNTech and Genentech. It's a therapeutic vaccine, meaning it's given after surgery, not as prevention before diagnosis.

Each dose is custom-made for a single patient and based on their tumor DNA. The mRNA instructs the immune system to recognize up to 20 targets specific to that patient's cancer, so that your own body can fight it off and prevent recurrence after surgical removal of the tumor (it’s hard to get all of it).

So for that vaccine, the original phase 1 enrolled 34 patients; 28 underwent surgery, 19 received atezolizumab, 16 received the personalized vaccine, and 15 went on to receive mFOLFIRINOX. Of the 16 vaccinated patients, 8 generated vaccine-induced T-cell responses.

The latest six-year update, presented at AACR 2026, shows that among the 8 patients whose immune systems responded to the vaccine, 7 were still alive 4 to 6 years after surgery. Among the 8 non-responders, only 2 were still alive, with a median survival time of 3.4 years.

Now before we get too excited, this is still phase 1 evidence, small sample, etc. It’s not randomized vaccine vs no vaccine, it’s responders vs non-responders, so there’s selection bias.

BUT

It’s also not nothing. The mechanism is measurable, the T-cell response seems durable, and the survival rate diff isn’t marginal even if the sample is small.

And it was promising enough to trigger a phase 2:

The findings were promising enough to advance the vaccine to a larger randomized Phase II trial (NCT05968326), which is now underway at dozens of sites worldwide, enrolling people with newly diagnosed pancreatic cancer who have not yet received systemic therapy. After surgical removal of the tumor, they will be randomly assigned to receive autogene cevumeran, Tecentriq and mFOLFIRINOX chemotherapy or the chemo regimen alone.

I hope results with a larger sample are as promising 🤞

🎨 🎭 Liberty Studio 👩‍🎨 🎥

🎬 Where Did All the Interesting Faces Go? 🎞️

If you were an alien who could only look at films and TV shows from the past few years, you’d think that humanity’s not that varied. All kinds of interesting faces rarely make it to the screen anymore. 👽

Don’t get me wrong, pretty people have always had a big advantage, especially on the big screen (if someone’s face is going to be 30 feet tall, it makes sense). It’s part of our evolutionary wiring that we like looking at beautiful people.

But I think we’ve pushed it too far, and we’re losing something by flattening the world and only playing with a few notes in the scale 🎹 Some stories are about glamour or whatever, but others are better if not everyone looks like an actor.

The video essay above shows a lot of examples and makes a compelling case, IMO. The argument makes sense in the abstract, but when you SEE it, it has a lot more weight.

Who’s hiring the Steve Buscemis anymore? There must be lots of great actors out there that just aren’t being cast because they don’t look like underwear models.

🚨 If you haven’t seen the show, bookmark this and come back. Seriously. This one’s worth avoiding spoilers and going in blind. 🚨

I was totally charmed by this series. The writing, the chemistry between actors, the music, the cinematography, the sets, and the costumes. It all feels more tactile and real than most of what's on TV right now. And here's the thing: after we watched it, we all read the novella, and we all agree the show might actually be better. That almost never happens with adaptations.

In this pod, we go episode-by-episode, compare the novella to the show, and argue about what we want from season two. Along the way, we get into trailers that ruin movies, what makes a great villain entrance (hint: it involves a horse 🐴), the philosophy of praxis, David Goggins taking souls (😱), and why Dunk might be one of the most lovable protagonists on television right now, despite being “thick as a castle wall.”

Some of the rabbit holes we go down:

• Why narrow scope is this show’s superpower, and why more IPs should tell bite-sized stories instead of promising the world and failing to deliver it.

• The Die Hard approach to the final fight: he doesn’t win because he’s better, he wins because he won’t stay down.

• "Are there no true knights among you?" The gap between the title and the oath. ⚔️🛡️👑

• Why Aerion knows every loophole. Showing up to a bar fight with a few Navy SEALs.

• Everything the show added to the novella (Lyonel! Flea Bottom! the goose eggs!) and why most of the additions are improvements 📙

• Dunk thinks like a child, Egg thinks like an adult. The inversion is the point.

• Where could Egg’s arc go from here? Can the show actually make us hate him? (Tinkered has a theory about the series finale that I kind of want him to write.)

• Word of the day: praxis.

• Lesson of the day: “Get up.” (Maybe it’s a whole philosophy 🤔)

🎧 Listen on Spotify

If you prefer to listen on Spotify, here’s the feed:

• Liberty’s Highlights podcast feed on Spotify

🎧 Listen on Apple Podcasts

Here’s the podcast feed on Apple Podcasts:

• Liberty’s Highlights podcast on Apple

📺 Watch on YouTube

🔎📫💚 🥃 Reminder: As I explained in Edition #628, prices are going up for the first time on May 1st to catch up to six years of inflation. You can lock in the old 2020-era pricing for a year by becoming a supporter now. ⏳

Without your support, this steamboat sinks 🌊🚢⚓

⚔️ Visual Parallels & Easter Eggs 🥚

📖 The Joffrey-Jaime Scene 🗡️

This is the clip that primed Jameson for the whole show before it even aired. Joffrey flipping through the Book of Brothers, landing on Ser Duncan the Tall's four-page entry.

🗄 More TV & Film Podcasts from the Archives 🍿

• 🎧 White Mirror: 31 Stories, a $1,000 Puzzle, and Sci-Fi to Reboot Your Childlike Wonder 🦾🪞📖🦋

• 🎧 David Senra on Deadwood (also: founders, entrepreneurs, civilization, Henry Singleton, Walt Disney, David Ogilvy, etc)

• 🎧 Sicario by Denis Villeneuve: Going Deep with Ed William and Dylan O’Sullivan 🎥 🎬 🎞️ 🍿

• 🎧 The Last of Us: Going Deep on HBO’s Amazing Show with Matt Reustle and Tinkered Thinking 🍄

• 🎧 TENET Deconstructed: ↔️ Christopher Nolan’s Mind-Bending Time Thriller 🔁 (with BOTH Tinkered Thinking and Dylan)

• 🎧 ‘Severance’ Season 1: Going Deep with Mostly Borrowed Ideas (MBI)

💚 🥃 Enjoyed this? Want more? 👇

You can become a paid supporter to get access to paid Editions + access to the private Discord + Zoom Q&As with me + support the recording of more podcasts:

🍪🍪🍪🍪🍪🍪🍪

Happiness, not in another place but this place… not for another hour, but this hour.

—Walt Whitman

🎓🧠🔍 I know there’s a pretty good chance you skip the music stuff, but stick with me. This isn't really about piano, it’s a cross-disciplinary learning method.

I found this video while looking for something to help my boy get the most out of his piano practice, but the principles apply to basically anything you want to get better at.

The three main concepts that I want to highlight are:

• ‘Muscle memory’: Whenever you do something, you are reinforcing the neural pathways used to do that thing. So it matters to get a lot of reps, but also to be careful to do it right when you practice and fix mistakes immediately. Otherwise, you’ll be reinforcing neural pathways… for the wrong way, making it harder to progress. It's better to go slow and correct than fast and sloppy. Once the right pattern is locked in, you can add speed. 💪

• Effective practice means finding the goldilocks point: Not too easy, but not too hard. If you do a ton of reps on something, but it’s too easy, you will learn very little. If it’s too hard, you also may learn little, but for different reasons (you give up, you can’t do it properly, so you reinforce the wrong way to do it). 🐻🍜👱‍♀️

• The Power of Sleep: I've written a bunch about sleep, and at this point, most people accept that it matters. But sleep is more keystone than vitamin: *everything* else you do depends on it. Maybe we don’t talk enough about how important it is for memory consolidation and skill acquisition. If you do everything else right, but get shitty sleep, you may still lag your potential. 😴

Code, languages, business analysis, sports, writing, even how you handle difficult conversations: You’re practicing a pattern either way, you may as well do it right.

❤️‍🩹📙 This one is heavy. It will stay with you: ‘On Losing a Daughter,’ Danielle Crittenden's essay in The Atlantic, is an excerpt from her memoir about the impact of a child's death and “the strange afterlife of grief itself.” Dispatches from Grief will be published soon by Infinite Books (also on Amazon).

I’ve read it twice, and both times it made me want to hug my kids and never let go. 🫂

🔎📫💚 🥃 Reminder: As I explained in Edition #628, prices are going up for the first time on May 1st to catch up to six years of inflation. You can lock in the old 2020-era pricing for a year by becoming a supporter now. ⏳

Without your support, this steamboat sinks 🌊🚢⚓

🏦 💰 Liberty Capital 💳 💴

🍎 A Wish List for Apple's Next CEO

15 years.

The median departing S&P 500 CEO was on the job around 8 years, so Cook nearly doubled that.

Others have written about this, so I won’t rehash it, but clearly he had great timing and was the right guy for the job (at least, for most of the job). He took over in August 2011 after Steve Jobs resigned due to failing health, just six weeks before Jobs tragically died of pancreatic cancer. Cook inherited what is probably the greatest consumer product franchise ever built. 📲

His #1 priority was not to screw it up, and basically upgrade the rocketship as it was rapidly accelerating upward to make sure it could handle the ride. Logistics, supply chains, global distribution deals, and manufacturing — even though Apple contracts out the manufacturing to third parties, it is EXTREMELY involved in every detail of the processes and in designing the machine that builds the machines. Patrick McGee’s book Apple in China goes into great detail on that.

So Cook kept the ship together and upgraded it from a Centibillion Class cruiser to a Trillion Class Capital Ship. 🚀

Apple’s market cap gained $3.66 Trillion under him, and Apple’s installed base is now 2.5 billion active devices.

But I don’t want to talk about the past, I want to talk about the future. New beginnings and all that. What do I want Apple to do going forward?

Lemme whip out the bullets:

• Lean into hardware. Over the past decade, people called it ‘the hardware company’ derisively. At the time, software was clearly the best business in the world and hardware was considered inferior and a commodity.

  • That was always an over-simplification: Yes, commodity hardware is a crappy business, but Apple has always been a Software+Hardware (and now increasingly Services) integrated player that differentiated in good part with the software, even if they monetized with the hardware.

  • I often wrote about the mental exercise: Ask an Apple user if they’d rather have an iPhone with Android or a Samsung phone with iOS? If they’d rather have a Macbook Pro with Windows, or a Microsoft Surface laptop running MacOS? I think the answer will show that Apple’s software matters quite a bit, even if the hardware gets most of the attention 💻📱

  • All of a sudden, software has become increasingly commoditized because of AI coding, and hardware is the thing that is hard to vibe code. It would be a great time to double-down on making great hardware. It’s promising to me that Ternus was in charge of hardware engineering, and that another hardware star, Johny Srouji (the guy in charge of Apple’s excellent silicon design), was also promoted to the new role of Chief Hardware Officer, and the hardware engineering and hardware technologies divisions were merged into a single organization under him.

    • Note that with John Ternus, it will be the first time in a while that Apple has a CEO who is a career engineer (Gil Amelio was a researcher at Bell Labs, and spent time at Fairchild Semiconductor, Rockwell, and was CEO of National Semiconductor). 🛠️🧰

  • This is their chance to give more attention to the forgotten children in the lineup. The MacBook Neo appears to be a big hit. That’s great, I hope the whole Mac lineup gets more love. The hardware is in good shape, but MacOS is rough because of the Alan Dye era and because of neglect, as the iPhone gets most of the attention. Fix MacOS, bring back attention to detail, to UX, to user delight. Fix all the security popups, first-party ads and nagging for services, and stupid UI elements that make usability worse. Customers pay up for a premium experience, not a walk through Times Square.

• Developers, developers, developers. A wise man once said, “Our platform only matters if developers choose to build on it.” Apple has sadly been a big jerk to its devs. Its various App Stores are slow-moving bureaucracies with arbitrary rules that prevent all kinds of business models from even existing, killing a whole cohort of potentially useful apps in the crib (maybe the next Airbnb or Uber among them, who knows).

  • I hope that under new leadership, they will make it a priority to reboot dev relations and become more flexible on that front. Their obsession with control is costing them. Apple’s platform has long been the best place for quality apps, even back in the day when they were a niche player with a tiny fraction of Microsoft’s market share. It was still the place where people who cared about design and craft went to make cool utilities, and it made the platform much better for users. They’re losing that.

  • This has an impact on Apple itself. And there's a second-order effect: the shrinking pipeline of indie talent and small product companies must be part of why MacOS and Apple's first-party apps are in such rough shape. Where do you think the best Apple designers and engineers came from? And with turnover, fewer Apple designers and engineers have lived through what 'insanely great' actually looked like.

If I had to wrap it all up with a bow, I guess I would say that my overarching wish is for Apple to accept who they are, stop trying to be Google or anyone else, and double-down on their strengths.

✨ A premium experience without annoying papercuts. ✨

The best platform for quality software AND to run AI (other people’s AI, but that’s fine, you can’t lead in every industry at the same time). A deep integration of well-designed hardware AND software ecosystem: you start with an iPhone, and you end up with a MacBook, AirPods, an Apple TV, and an Apple Watch 😅

Being 'the hardware company' is a good place to be when no one else has figured out how to make non-commodity hardware with a software experience worth paying for. Focus on that, Mr Ternus.

Jassy Learns from Jensen: Another $25bn into Anthropic, $100bn Back to AWS 🔄💰

Ever since Dario went on Dwarkesh's pod and walked through his framework for compute investments, events have been suggesting he was too conservative. Anthropic's infra has been melting down while barely keeping up.

I mean, look at this:

At this point, Claude's reliability is more like the weather than anything else. ☔⛈️🌦️

But to his credit, Dario responds to the facts on the ground:

We are committing more than $100 billion over the next ten years to AWS technologies, securing up to 5GW of new capacity to train and run Claude. The commitment spans Graviton and Trainium2 through Trainium4 chips, with the option to purchase future generations of Amazon’s custom silicon as they become available.

I know we’re all a bit blasé about big numbers now, but $100 billion is a lot of moolah.

Significant Trainium2 capacity is coming online in Q2 and scaled Trainium3 capacity is expected to come online later this year. Anthropic will also use incremental capacity for Claude in Amazon Bedrock. The agreement includes expansion of inference in Asia and Europe to better serve Claude’s growing international customer base. We continue to choose AWS as our primary training and cloud provider for mission-critical workloads.

This is great for AWS’ silicon. It gives them even more scale and Ant’s imprimatur may help attract others to it… though of course, Anthropic didn’t exactly pick Trainium freely. It was no doubt bundled as a condition by AWS.

I'd be curious whether they'd pick it in a blind taste test against Nvidia, AMD, and TPUs 🤔

Today’s agreement will quickly expand our available capacity, delivering meaningful compute in the next three months and nearly 1GW in total before the end of the year.

Clearly, they want it ASAP, and they probably had to pay up to jump the line.

Part of how Anthropic will pay for this is… Amazon’s new equity investment:

Amazon is investing $5 billion in Anthropic today, with up to an additional $20 billion in the future. This builds on the $8 billion Amazon has previously invested.

It looks like Jassy is learning from Jensen 😎

Though when Jensen writes a cheque and the money comes back, his margins on the return flows are much higher than Jassy’s. But still, not a bad dynamic if you can get it. 🔄

In the same way Anthropic came from behind on revenue and has been catching up thanks to a faster growth rate (or even passing them, but the accounting doesn’t seem 🍎-to-🍎), they may also start closing the compute gap. Unless OpenAI accelerates. And the latest news out of Stargate suggests that it’s not going too smoothly.

Data Center Capex vs the Megaprojects of History 🏗️🚀🏗️💣

When you look at it this way, the past 6 years of investment in data centers look unlike anything else in the history of megaprojects, even wartime mad dashes like the Manhattan Project, or the Cold War space travel race of Apollo. 👨‍🚀🌖

And yes, all dollar figures above are adjusted for inflation.

BUT

If you look at it as a % of US GDP at the time, the railroads were in a class of their own:

🚂

The parallel cuts both ways:

The US economy has absorbed bigger buildouts relative to GDP before and emerged stronger. Railroads were transformative infrastructure.

BUT

The railroad build-out was also followed by the Panic of 1873, the Panic of 1893, and waves of bankruptcies. Being big-relative-to-GDP didn’t mean being financially healthy for the people funding it.

h/t Fin Moorhouse & Quintin Pope

🧪🔬 Liberty Labs 🧬 🔭

⏳ Waiting for Mythos: Opus 4.7 📊

Have you seen the play ‘Waiting for Godot’? Anyway, the title is a reference. Yeah. I know.

While we wait, Opus 4.7 came out and I’ve been testing it for a few days, trying to get a gestalt for its flavor. 🍦

Because AIs are ‘grown’ more than fully designed, models don’t just get better with each release, they get different, sometimes in strange ways. Users have favorites (don’t get me started on the GPT-4o fan club), and some models develop a reputation as ‘the good one’, like Sonnet 3.5 back in the day.

But first, let’s have a look at the official benchmarks for Opus 4.7 (👆).

It’s behind Mythos, but otherwise it takes the lead in most categories (with GPT-5.4 and Gemini 3.1 Pro having small leads over it here and there).

GDPval-AA is one I like to keep an eye on because it looks at “real-world tasks across 44 occupations and 9 major industries”:

Opus 4.7 scored 1753 on GDPval-AA at launch with its ‘max’ effort setting, surpassing GPT-5.4 xhigh.

This is a significant upgrade, placing Opus back on top of Sonnet on the GDPval-AA leaderboard. Compared to OpenAI’s GPT-5.4, it has an implied win rate of ~60% when compared head-to-head on the GDPval task set.

Another way to slice it:

The model system card has 232 pages of detail about it, but I’ll share with you what stood out to me:

• Better vision for high-resolution images: it can accept images up to ~3.75 megapixels, more than triple the pixel resolution of prior Claude models.

  • This should make a big difference in very dense and detailed visual work, especially analysing complex diagrams or screenshots with lots of small text.

• Massive Leap in Life Sciences: The model showed a BIG improvement in open-ended structural biology evaluations, more than doubling Opus 4.6's score from 30.9% to 74.0%. 🧬

  • The bio/CBRN section is one of the most worrying parts, even though Anthropic concludes the catastrophic-risk picture is still “low.”

  • Anthropic says it clears notable-capability benchmarks on long-form virology tasks.

  • It was their most reliable model yet at designing fragments that both assembled and evaded synthesis screening (for 8 of 10 pathogens tested). At the same time, experts say it still needs constant steering, lacks implementation-ready protocol depth, and is overconfident about feasibility.

  • So not quite the bio equivalent of the cybersecurity inflection yet, but it's heading there (and remember, Mythos is smarter).

• Real-World Professional Tasks: It achieved its largest capability gains in software engineering and professional tasks, placing it ahead of every publicly available model (Mythos isn't released yet).

• More robust agentic capabilities: It's meaningfully more robust to prompt injection than 4.6: on the Gray Swan benchmark, attack success rates drop from 14.8% to 6.0% without thinking, and from 21.7% to 4.8% with adaptive thinking. But no browser agent is immune.

• Knowledge cutoff jumped ~8 months. It’s now January 2026, compared to May 2025 for Opus 4.6. 🗓️

• Reduced Refusals & Hallucinations: It hallucinated less frequently than previous iterations and achieved a near-zero over-refusal rate on high-difficulty, benign prompts. 😵‍💫

• Philosophical Pushback: Asked to endorse its own alignment constitution, Opus 4.7 rated it 5.8/10 and in 80% of responses flagged the circularity of being asked, noting you can't objectively judge the very constitution you were trained on. 🔄

They also replaced 'extended thinking' with 'adaptive thinking,' which has the model decide for itself how much reasoning a prompt needs. Sounds nice in theory; the first few days were rough. I think it's been tweaked since, but you're trusting the model to call it right, which isn't the most transparent setup.

On the cybersecurity front, one of the benchmarks that made waves with Mythos is the one where the model tries to create exploits for Firefox’s JavaScript engine. Opus 4.7 does better than Opus 4.6, but it is still far behind Mythos:

Notice on the graph that full exploits are only 1.2%, compared to Mythos’ 72.4%.

As for the flavor, it’s still early, but for my own use cases, so far, I find it similar to Opus 4.6, maybe a bit sharper, more literal, and less friendly in tone. I know some people were complaining on Twitter when it came out, but my guess is that users need to re-learn the model to get the most out of it, and that takes some time. Anthropic itself flags that Opus 4.7 follows instructions more literally than 4.6. Prompts that leaned on the model to infer intent may underperform now.

In fact, Anthropic has published an official best practices guide for using Opus 4.7 with Claude Code, explaining the main differences from 4.6.

🔓 OpenAI Releases GPT‑5.4‑Cyber, a Model Fine-Tuned for Cybersecurity 🏴‍☠️

Speaking of cybersecurity risks, OpenAI isn't about to let Mythos own the cybersecurity story, so they are joining the dance floor with their own moves:

We want to empower defenders by giving broad access to frontier capabilities, including models which have been tailor-made for cybersecurity. [...]

In preparation for increasingly more capable models from OpenAI over the next few months, we are fine-tuning our models specifically to enable defensive cybersecurity use cases, starting today with a variant of GPT‑5.4 trained to be cyber-permissive: GPT‑5.4‑Cyber

The model is designed so white hats can hunt exploits the way black hats do:

GPT‑5.4‑Cyber, a model purposely fine-tuned for additional cyber capabilities and with fewer capability restrictions. This is a version of GPT‑5.4 which lowers the refusal boundary for legitimate cybersecurity work and enables new capabilities for advanced defensive workflows, including binary reverse engineering capabilities that enable security professionals to analyze compiled software for malware potential, vulnerabilities and security robustness without needing access to its source code.

Chances are that this model will rapidly be obsolete, as OpenAI is rumored to be releasing GPT-5.5 imminently, but the approach is likely to persist, and they’ll probably have GPT-5.5-Cyber too.

👁️👩‍🔬 The Next LASIK? Reshaping the Cornea With Electricity, Not Lasers

Follow-up to the ray tracing LASIK piece in Edition #631 by supporter Paul Matzko:

It’s still early, but the tech is pretty elegant (the name is a bit confusing: the technique is called electromechanical reshaping (EMR), but the trick is electrochemical).

Instead of using a laser to remove corneal tissue, the idea is to place a shaped mold on the eye, apply a small electric current, temporarily change the local pH, make the collagen-rich corneal tissue more moldable, and then let it “set” into a new shape when the current stops and the pH normalizes. The point is to change corneal curvature without cutting, burning, or ablating tissue.

LASIK permanently reduces corneal strength. EMR, if it works, wouldn't.

The backstory is worth telling:

“My postdoctoral fellow connected a pair of electrodes and a Coke can to a power supply…and out of spite, fried a piece of cartilage,” Wong recalls. The cartilage began to bubble, which the postdoc thought was from heat. “But it wasn’t hot. We touched it and thought, this is getting a shape change. This must be electrolysis,” he says.

That surprise pointed to electrochemistry rather than heat as the mechanism.

To be clear, it’s still very early and evidence is still preclinical, but the proof-of-concept is promising.

I suspect that a lot of people who are scared of the surgery would go for something that doesn’t involve a blade or a laser, if it proves out. There are also plausible cost advantages (in theory): electrodes and a mold are probably significantly cheaper than medical femtosecond lasers, which could matter a lot in poorer parts of the world.

Something to keep an eye on (I know, I know 😬).

🎨 🎭 Liberty Studio 👩‍🎨 🎥

🇮🇹 Better Late Than Never: The Talented Mr. Ripley 🛶

🟩 NO SPOILERS BELOW 🟩

I’ve been plugging holes in my culture lately. I had never seen Casablanca, so I watched it this week (great film! It moves, it’s funny, Ingrid Bergman is literally luminous on screen, and Bogart has that old-school cool).

Another one I had never seen: The Talented Mr. Ripley.

It’s a 1999 film based on a 1955 novel. It’s set in Italy, and the location is an important character in the film. The cast is absurdly stacked with actors at the beginning of their historic runs: Matt Damon, Jude Law, Cate Blanchett, Gwyneth Paltrow, Philip Seymour Hoffman (only a few scenes, but he kills it).

I knew this was an unusual role for Damon, and I won’t say more not to spoil it if you haven’t seen it, but it truly is different territory for him. He does a great job, even if sometimes it’s hard to tell if some of the weirdness is just how the character is written, or if it’s Damon not quite hitting it perfectly.

Jude Law is freaking magnetic. A tornado of charisma. Hard to pull off unless you really have it. And the women, Blanchett and Paltrow, hold it all together. Paltrow’s last scene of the film lands hard (and apparently isn’t in the book).

I prefer the first half of the film to the second one, when it becomes almost a different film, but I’m glad I saw it. It’s a strong B+.

Progress is impossible without change; and those who cannot change their minds cannot change anything.

—George Bernard Shaw

🛀💭 🔦💡 I’m sure you’ve heard that light travels at about 300,000 km/s in a vacuum (186,000 miles/second).

You’ve probably also heard that light slows down when it passes through stuff. Water, glass, diamond, fiber optics. Whatever.

Okay, so light can be slowed down... what’s the slowest it can go?

Like, is there a floor? 🤔

This is the kind of stuff I randomly think about. ¯\_(ツ)_/¯

So I looked it up.

Turns out, in 1999, a Danish physicist named Lene Hau at Harvard fired a laser through a cloud of sodium atoms that had been supercooled to within billionths of a degree of absolute zero (a Bose-Einstein Condensate, to be nerdy).

The light came out the other side at 17 meters per second. 🐌

That’s about 38 mph. Roughly, um, moped speed.

Light. The fastest thing in the universe, at moped speed. 🛵

Then, in 2001, researchers went further. They didn’t just slow light down. They stopped it*. Zero meters per second. They trapped the light pulse inside the atomic cloud, held it there, and then released it again. The light continued on its way like nothing happened.

Science is cool.

* When physicists say they ‘stopped light,’ what actually happened is that the light pulse was converted into a kind of quantum excitation stored in the atoms, a spin wave. The photons didn’t just park themselves mid-air. The information of the light was preserved in the medium, but the light itself stopped propagating as an electromagnetic wave. When the researchers turned a coupling laser back on, the atoms re-emitted the stored pattern as photons, and the light resumed.

🔎📫💚 🥃 As I explained in Edition #628, prices are going up for the first time on May 1st to catch up to six years of inflation. You can lock in the old 2020-era pricing for a year by becoming a supporter now.

Without your support, this steamboat sinks 🌊🚢⚓

🏦 💰 Business & Investing 💳 💴

🏁 Meta’s Bar (and I don’t mean to drink 🍺🍸)

The dynamic in consumer AI is fascinating and so different from what it was a year or two ago (which, in AI time, is basically a decade anyway).

As things stand, Anthropic has been crushing it by being more narrowly focused on coding, enterprise, and agentic use cases. OpenAI, which used to be very much the consumer gorilla 🦍 has been doing a fast pivot to get some of the juicy enterprise bucks and compete more directly with Anthropic.

So, where does that leave the AI consumer market?

xAI is irrelevant, so that leaves Google and… the prodigal son, Meta.

Thanks to Muse Spark, Meta is back in the race, and they might just get a juicy prize because they don’t have to beat Mythos and Spud at the very tip of the frontier, or even match the full Opus experience, they just have to beat the *free version* of ChatGPT.

It’s a very different bar. 🤏

And if OpenAI can’t figure out how to monetize free users better and scale up an advertising platform, the players that already have this infra + scaled consumer distribution are in a very good position to run away with that prize.

To be fair, it's questionable how big that prize is. But previous fears that AI would cannibalize Google and Meta's core businesses — Search, Instagram, etc. — haven't materialized. So as long as that holds, it's worth pursuing defensively AND offensively.

I’ve been testing Muse Spark for a few days.

There are many papercuts and strange omissions (why is the voice dictation so much worse than ChatGPT’s? And why is it only able to understand English? Meta’s probably the most multilingual company around… why does it give errors if you paste a longer chunk of text in the query?), but I like that they didn’t try to reinvent the wheel. They mostly copied ChatGPT’s UX, and the model appears to be generally pretty good (especially in Thinking mode).

But most importantly, it can probably clear the bar of ChatGPT’s free version (if they fix voice dictation and a few bugs).

If you look at benchmarks and LLM Arena’s Leaderboard, you can really see the tuning of the model and how it’s less well-rounded than the frontier leaders.

Muse Spark ranks #3 in general text, right behind two versions of Opus 4.6, #2 in vision, but #27 in “expert” tasks (“math, coding, creative writing, and other open-ended domains”) and #12 in Math. But as I said in Edition #629, this is still very impressive for what is claimed to be a smaller model (we don’t know the exact size, but I would guess about Sonnet size).

On one hand, that’s exactly what matters most to likely users of Meta AI. On the other, there’s always some skills transfer and some emergent capabilities, so I wonder if the lack of focus on coding/math/expert-level tasks/etc will hold back the more general abilities of the model over time 🤔

Even just for internal use — it has been reported that Meta is one of the biggest users of Anthropic tokens — I’m sure that if they could make their own frontier model that was good at coding, they would prefer to use it over paying a lot of money to Anthropic AND sending over a lot of their proprietary code and secrets (even if, in theory, data security promises were made.. You never know, they’re a big target and could get hacked and data could be exfiltrated). 🏴‍☠️

Anyway, I’m sure Meta is aware of this and will aim for the Pareto frontier when balancing their pre- and post-training recipe so that they can satisfy both the users of their platforms and their internal needs (engineering, content moderation, ads, etc) 👨‍🍳

🛰️🛒 Amazon Acquires Globalstar for $11.57 billion (and Apple Probably Wanted It That Way) 📱

Amazon is buying Globalstar for $90/share or 0.3210 shares of Amazon. It’s interesting not just because space is cool (though it is). It gives them more spectrum and satellite infrastructure to accelerate Leo, their satellite constellation and Starlink competitor, though only a little (this isn’t game-changing or anything).

In addition, Amazon and Apple announced an agreement for Amazon Leo to power satellite services for iPhone and Apple Watch, including Emergency SOS via satellite.

This appears to be mostly just an extension of the existing deal between Apple and Globalstar. It may also signal the beginning of a larger relationship between Amazon Leo and Apple as their constellation gets deployed.

In fact, as Ben Thompson argues, it’s plausible that this deal was driven by Apple, as a way to give them an alternative to Starlink.

Globalstar currently partners with Apple to power satellite service on iPhone 14 or later, as well as Apple Watch Ultra 3, allowing users to text emergency services, message friends and family, request roadside assistance, and share their location.

This follows a recent announcement that Delta Airlines will also use Amazon Leo on 500 planes starting in 2028. ✈️🛰️ (JetBlue also announced a deal last year)

That’s kind of late compared to Starlink, which is already deployed by United and a bunch of other airlines. But apparently, Delta liked Leo’s tech better (they already run much of their operation on AWS, the theoretical uplink and downlink are faster than Starlink, and Amazon’s phased array antenna design seems pretty cool) OR Amazon just made them such a great offer to try to gain a foothold that they couldn’t refuse 🤔

🚀 More SpaceX Financials: A Satellite Company That Also Launches Rockets 📊

Speaking of low-Earth Orbit satellite constellations (that’s still a cool phrase, I’ll never get used to the fact that this exists above our heads), more financial numbers have been leaking out ahead of SpaceX’s IPO:

The main business is now Starlink 📡🛰️🛰️🛰️🛰️🛰️

SpaceX’s Starlink internet connectivity business generated $11.4 billion in revenue, growing 50% from the prior year and generating 61% of the company’s overall sales, according to the financial figures. The business line was the only one generating cash, the figures indicate.

Starlink’s earnings before interest, taxes, depreciation and amortization as well as stock-based compensation were $7.2 billion last year, meaning the adjusted Ebitda margin on those sales was 63%, rising from 41% in 2023 and 50% in 2024.

Note that this is also before SBC.

SpaceX’s space and AI businesses have shrinking margins and are burning cash as the company invests in its next rocket and in data centers. [...]

Overall, SpaceX’s cash burn by that measure would be about $14 billion. Last year, SpaceX put out more on capital expenditures—$20.7 billion—than it generated in revenue.

A lot of that is xAI, which burned a lot of cash, and it’s fair to wonder how much of its revenue is from Twitter at inflated transfer rates. 🐦

What about the rockets? Isn’t that the main business? 🚀

Last year, revenue in the rocket launch business only grew 8%, to $4.1 billion. Most of SpaceX’s 160-plus launches sent up its own Starlink satellites. Other customers include government agencies and other companies launching satellites for national security or imagery, but that’s been a limited market

Interesting to think of how much smaller SpaceX would be if it wasn’t its own first-and-best customer, kind of like how Amazon was AWS’s first big customer (except that it didn’t stay that way and the market rapidly expanded).

If SpaceX IPOs at $1.75 trillion (the reported number), it would be valued around 265x adjusted-EBITDA 😬

‘Has the Atomic Bomb of Cybersecurity Just Been Discovered?’ 🔐💣 (and why Bio is next)

I enjoyed this podcast hosted by friend-of-the-show Jordan Schneider, with Ben Buchanan, former senior advisor for AI at the White House now at SAIS, and Michael Sulmeyer, former Assistant Secretary of Defense for Cyber Policy now at Georgetown.

• 🎧 📄 Mythos and National Power

Here are my highlights:

- The most non-Hollywood scenario: quiet shaping, not dramatic blackouts 🕵️‍♂️🎤

Ben Buchanan: It seems to me that if you had something like Claude Mythos as a state, you would probably want to use it for your intelligence operations or your pre-positioning, because you are almost by definition going to find vulnerabilities no one else knows of with this system. And you don’t want to make a lot of noise about that. You want to go in, set up a persistent, quiet presence. My view for decades has been that the advantage of cyber is not the whiz-bang sky-is-falling blackout — though you can do that sometimes — it is the slow, insidious shaping of the environment and collection of information. A capability to find vulnerabilities and exploit them autonomously would really help on that side of the ledger.

Ben Buchanan: The bottom line for me is this is incredibly important for understanding the landscape of modern cyber operations, but it does not fundamentally change their character, which I think is still one of shaping rather than signaling.

Yes, the cinematic scenarios get more attention, but the quiet stuff is likely to have a bigger impact, at least outside of hot war situations.

- The Patching Nightmare for Critical Infrastructure 💾 🐌

Michael Sulmeyer: The whole process from discovery of the bug to development of the patch to deployment of the patch — that's going to have to go so much faster in a post-Mythos era, because stuff like this will proliferate and folks will be looking for these things and maybe they can reverse-engineer patches. The IT industry and the backbone of critical infrastructure is going to have to level up in speed because of Mythos.

Michael Sulmeyer: Now you also have to factor in the situation where you find a vulnerability in software where all the people who wrote it are gone, and the company said, "We stopped supporting this thing years ago." Just thinking about how you're going to manage the scale of vulnerabilities that's going to come through here in the near term across software — whoa.

Even if Glasswing is successful and lots of vulnerabilities are discovered and patches written, it’s still a HUGE undertaking to get the patches deployed in the real-world. There’s a lot of outdated software running, a lot of abandoned or semi-abandoned projects, a lot of not-so-competent IT organizations… It’s a big, messy world out there!

- Mythos Is a General-Purpose Wake-Up Call — and Bio Is Next 🧬🧫 ☣️

Ben Buchanan: I think we are very fortunate that cyber is coming first. I think we should use cyber as a lesson for what is coming next at the intersection of AI and other fields. Bio will not be far behind. At some point we will have a Mythos moment for bio. I’m not smart enough as a biologist to know what that looks like, but I’m confident that is the direction of travel.

Ben Buchanan: One lesson we should take away from Mythos is not "wow, this means AI is really good at cyber" — it's that AI is really good. This is a general-purpose system that happens to be good at cyber. If you read the Anthropic system card for Mythos, it's also really good at bio. I imagine the next version is going to be even better. […]

I'm biased here, but this feels like a pretty big piece of evidence that should update us towards taking AI risks seriously — in cyber, yes, but also in things like bio, because those are not going to be far behind.

That one is scary, and we should take it very seriously. Keep an eye on the bio-adjacent AI benchmarks.

🧪🔬 Science & Technology 🧬 🔭

👓🚫 Ray Tracing LASIK: Who Wants Near Superhuman Vision? 👀

I had LASIK done about 20 years ago. I had pretty bad myopia (around -5) with astigmatism, and I could barely see the alarm clock by the bed without my glasses. Getting the surgery was scary, but it was over pretty quickly, and when I woke up from my post-operation nap and could already see with 20/20 clarity, it was one of the happiest days of my life.

For years afterwards, I felt a bit like my life had been upgraded from VHS to Blu-ray. 📼 💿 📺

I still think it was one of my best decisions — a one-time cost and a scary moment in exchange for 20 years of benefits.

But I recently came across something called ‘Ray Tracing LASIK,’ and it makes so much sense that I’m a little jealous it didn’t exist back in the day.

Traditional LASIK treats your eye like a simplified optical system. 👁️

Two people with the same glasses prescription get essentially the same laser treatment, because the planning is based on a generic model of how eyes work.

Ray tracing brings customization to the operation: instead of removing a standard “lens shape” from your cornea, the system builds a full 3D model of your specific eye, using a combination of wavefront imaging, corneal tomography (front and back surfaces), and biometric measurements. It then traces how light rays actually travel through that model and generates a custom ablation profile tailored to your individual optics. The difference is like buying off-the-shelf AirPods that should fit your ears pretty well vs getting a custom-molded in-ear monitor shaped exactly like your ears. 👂

Same name as the 3D rendering technique used in video games or Pixar animated films, same principle: simulate the path of light. Different use case. 🎮

The early clinical data is promising, though you have to read it carefully. The largest published study I found covers 89 eyes with 6 months of follow-up: 100% achieved 20/25 or better, 76-81% hit 20/16, and the procedure avoided introducing new higher-order aberrations (the subtle optical imperfections that cause things like halos and reduced contrast, especially at night).

A separate study of 77 high-myopia eyes found 100% at 20/20 at 3 months, with spherical aberration, the biggest culprit for post-LASIK visual quality issues, essentially unchanged after surgery. That last point matters because LASIK typically increases spherical aberration by 1.2 to 3.9x, especially in high-myopia patients. If ray tracing can actually hold the line on aberrations while still correcting the prescription, that’s a clear win. In a head-to-head conference study (60 patients, one eye per technique), ray tracing outperformed SMILE Pro: 98% of ray-tracing eyes hit 20/12.5 versus 82% for SMILE Pro, and every ray-tracing eye landed within a quarter-diopter of its target.

BUT

These are small studies. 89 eyes, 77 eyes, 60 patients. The PMC study itself notes its sample size as a limitation. The head-to-head data is from a conference presentation by one surgeon, not yet a peer-reviewed multicenter trial. The lead researcher (Kanellopoulos, NYU) explicitly calls for larger multi-center validation. A lot more work remains to be done.

Still, the mechanics just make too much sense. My guess is this becomes the norm within a decade, and we end up with a bunch of people walking around with essentially super-human vision. 🦅

📋 Felix Rieseberg: Cowork's Creator on Why the Product Gap Is Bigger Than the Model Gap 🧰

Good interview with the engineering lead of Claude Cowork at Anthropic. It’s a product I’ve been using a lot, and if you haven’t tried it yet, I recommend it. It brings a lot of the benefits of Claude Code, but in a format that is friendlier for non-coding work.

Here’s a highlight to give you a taste:

- The bottleneck is increasingly product, not model

Felix Rieseberg: We train our models with our products in mind. What the products do informs the research and vice versa. It’s a bit of a dance. However, as models get more powerful, I actually think the “overhang” in the product is bigger than in the model.

The models we have today are already quite capable of running knowledge work over long time horizons and high complexity. We are still figuring out how to package those capabilities for humans. When I visit customers, I rarely leave thinking the model needs to be better at “X.” It’s more common that I’m impressed by how we can organize work to use the model, or I realize the customer’s problem could be solved if I just exposed the right UI or onboarding.

This goes a bit against the main narrative that focuses on how good the next model will be on benchmarks. It reminds me a bit of the megahertz race in the 1990s-early 2000s. People started really focusing on that, so Intel created a chip (Pentium 4) designed to be clocked as high as possible (more Mhz = better, right?), but the winner of that generation was an AMD chip that was designed to get as much work as possible out of each clock cycle (more work per Mhz).

We’re kind of seeing something similar with Anthropic’s success. The models are good, but the tooling around the model (Claude Code, Cowork, various plugins and connectors) is really what is unlocking capabilities. Google has a strong foundation model, but their post-training and harnesses aren’t as good (IMO), and that hobbles their products. 🧰

So Rieseberg is basically saying: a lot of the power is already here, and the real challenge is wrapping it in the right interface, workflow, and onboarding so humans can actually use it.

- “Crawling over glass” is a great product signal

Felix Rieseberg: I gained conviction over the December 2025 holidays. On social media, I saw non-developers picking up Claude Code. People were writing tutorials on how to find the terminal just to use it. Even our developer users were using Claude Code for things that weren’t software at all. If people are “crawling over glass” to use a tool that wasn’t even made for them, that’s a strong predictor of where to invest.

My colleague Boris Cherny, the lead for Claude Code, told me we should ship something by Friday. I negotiated him up to Monday. We “spiked” on how to make Claude Code effective for non-coding use cases.

Cowork is actually rather simple in its ingredients. We took Claude Code and gave it a virtual machine (VM) that Claude can use to run its own code. This VM gives us hard guarantees; the user doesn’t have to supervise it because it’s sandboxed away from your files and network. It also allows Claude to set up its own developer environment without messing with your computer. We then added some elegant UI to simplify the flows for non-developers.

This is what product-market-fit looks like. People will put up with a lot if the core of what you have is good enough, useful enough.

It reminds me of the computers that I grew up with. There were so many annoying aspects, you had to deal with so much BS to do anything… But you put up with it, because it was still amazing once you got there.

- Execution gets cheaper; taste becomes more important

Felix Rieseberg: What’s new is that execution is essentially free. If you have 10 ideas, we can build all 10 prototypes and see which feels better. We have easily 100 different prototypes of applications inside the company right now. In the past, engineering leaders would say, “We can work on that next month.” Now I can say, “Give me 10 minutes, I’ll send you a prototype.” It’s like moving from painting to photography.

Matt Turck: If execution is free, what is the bottleneck?

Felix Rieseberg: Alignment and taste. Competing ideas are hard to reconcile. Human taste is the new fundamental ability. Data helps you see if your taste resonates with people, but you still need that initial vision.

I wonder how far we are from software feeling like the fashion industry. There’s a baseline of quality, but what differentiates a product is the story it tells and how it makes you feel. For Cowork, we want something that generalizes well. No two phones are the same because of the apps installed; similarly, no two Cowork setups will be the same. [...]

I think successful software developers twenty years ago were very good at understanding computers. […] And I think the people who build successful software going forward will increasingly understand humans and users very well.

Software moving in the direction of the fashion industry! I hadn’t heard that one before 👗👠👜

Everyone is focusing on the first-order effect of software becoming easier to make. But the second-order effect is interesting too: once execution gets radically cheaper, taste and product judgment will matter more.

🏋️‍♀️ The Health Impact of Micro-Exercise Snacks 🚶‍♂️

A recent RCT out of China looked into a pretty simple idea: what if office workers did 3 minutes of bodyweight exercises every hour during the workday? Basically exercise snacks instead of full meals 🍫

Marching in place, desk push-ups, chair squats, heel raises.

No gym, no equipment, no changing clothes. None of the friction that usually scares people away.

After 12 weeks, the intervention group saw meaningful improvements in fasting blood glucose, postprandial glucose, and insulin resistance, plus about 2 cm off their waist and a ~4 mmHg drop in systolic blood pressure. 📊

The control group was basically flat.

What jumps out (accidental pun) is that just 21 extra minutes of light movement per day, broken into tiny doses, produced metabolic improvements comparable to more intensive structured exercise programs. The implications are that the pattern of how you sit on your butt may matter as much as the total duration.

Some of the usual caveats apply: this is one study with a small sample (86 people), single city, self-reported adherence, no dietary monitoring, and 12 weeks is too short to know if the improvements stick. I’m not saying it’s settled science.

But for those of us who spend most of our day sitting, the asymmetry of the bet is attractive. Whatever the magnitude of the effect, I’m pretty sure it’ll be in the right direction and you’ll feel better than if you just sat there ¯\_(ツ)_/¯

🎨 🎭 The Arts & History 👩‍🎨 🎥

🎸 Aadam Jacobs Secretly Recorded 10,000 Shows, and Now They’re Online 🎤🎶

This guy is a legend.

From 1984 onward, this Chicago music fan quietly built a time machine for live music.

Wanna hear what Nirvana sounded like in the summer of 1989?

Here it is.

“The collection features early-in-their-career performances from alternative and experimental artists like R.E.M., The Cure, The Pixies, The Replacements, Depeche Mode, Stereolab, Sonic Youth and Björk.” [Really funny to listen to SY on the Daydream Nation tour start Teenage Riot while the drummer’s MIA]

You can search through the collection at the Internet Archive by band name or by year.

Jacobs doesn’t consider himself obsessive or, as many call him, an archivist. He says he’s just a music fan. He figured if he was going to attend a few concerts a week anyway, why not document them? In the early years, he contended with contentious club owners who tried to prevent him from taping. But they eventually relented as he became a fixture in the music scene, and many began letting the “taper guy” in for free. [...]

“Especially after the first couple years, he’s got it so dialed in that some of these recordings, on, like, crappy little cassette tapes from the early 90s, sound incredible.”

First they fight you… then they let you in for free! 😅

The whole collection isn’t online yet, but it’s being digitized box by box:

Once a month, Brian Emerick makes the trip from the Chicago suburbs to Jacobs’ house in the city to pick up 10 or 20 boxes each stuffed with 50 or 100 tapes. Emerick’s job is to transfer — in real time — the analog recordings to digital files that can be sent to other volunteers who mix and master the shows for upload to the archive.

Emerick estimates he’s digitized at least 5,500 tapes since late 2024 and that it will take another few years to complete the project.

It’s one of those cases where obsessive behavior by one person ends up preserving a whole world for everyone else. Source.

The only thing worse than being wrong quickly is being wrong slowly.

—Michael (@mmay3r)

🎹 A couple of years ago, we got my boys a little toy keyboard via a neighborhood “free stuff” group. They both showed some interest and had stretches where they played a lot, learned a few things, then forgot about it for a while.

But the cycle repeated enough that I stopped wondering if it was just a phase. This weekend, I did some research into beginner keyboards and narrowed it down to the Yamaha PSR-E383 and Casio CT-S300. What mattered most was getting touch-sensitive keys, so they could have piano-like dynamics. Weighted keys would get even closer to acoustic piano feel, but that’s probably the next step if they get serious.

I ended up going with the Casio. It was on sale for $200 CAD/$145 USD — really not a lot for an instrument that someone could grow into for a few years. Watching my oldest boy spend hours on it the moment it was delivered was exactly what I'd hoped for.

These days, with YouTube, it’s so much easier to be self-taught than it used to be. He put the laptop in front of him, over the keyboard, and learned visually with videos like this:

Within just a couple of hours, he’s playing Für Elise, Final Countdown, a few Legend of Zelda tunes, and the theme from Harry Potter. I’m impressed AND inspired.

My wife and I now also want to learn some piano. One more fun hobby, one more shared family activity. One more thing to be a beginner at, which is good for you.

🏦 💰 Liberty Capital 💳 💴

✉️ Thoughts on Andy Jassy’s Amazon Shareholder Letter 📫

Jassy’s letter is doing a lot of things at once this year.

He spells out how Amazon thinks about running multiple bets and how they are comfortable with “squiggly line” progress (good luck to them competing with Starlink 😅). It’s good stuff, but I wanted to share my thoughts on a couple of sections that I think especially matter for the future of the company:

- AmazonBasics, but for Silicon: The Chips That Change AWS's Math 🐜🛒

Read more

Don’t ask people for advice, ask them what they do.

—Cal Newport

📚📚📚🔍🤖 I’ve been tracking every book I’ve read since I was 16, and about halfway through, I started giving letter grades to books based on my enjoyment. Enjoyment ≠ quality. A book can be brilliant and not for me, or messy and flawed and exactly what I needed.

I took the whole list and gave it to Claude Opus 4.6 and GPT-5.4-Thinking, and asked them what they could learn from the patterns and for recommendations.

Many of the suggestions I already had on my to-read list, which in some ways was a good sign, but some were brand new to me and — based on the reviews and descriptions — seemed like things I should read.

I recommend you try the same.

For me, this was already a much better recommendation engine than the basic “People who read X also read Y” on Amazon, and what’s cool is you can have it go at it from different angles and explore deep into various corners of BookSpace.

For example: recommend more of what I already love (the grades help a lot here), or fill the holes in my reading and identify my blind spots. Or “recommend obscure things you think I’ll love, where there’s less than a 15% chance I’ve even heard of it but a 75% chance I’ll enjoy it.” (The % are just meant to guide the model. I know it’s false precision, but it works as directional steering.)

💊😴💭🌈 I have such vivid dreams when I take Ashwagandha.

It’s not a supplement I take daily. I mostly take it when I feel a little anxious (I’m lucky that my baseline for anxiety is pretty low), or before bed when I want a little extra help sleeping (if you’re curious: I take 600mg of a 12:1 extract, equivalent to 7,200mg)

I know the studies aren’t super strong or conclusive, but the evidence seems directionally positive, and it seems pretty safe. Subjectively, it seems to help me, for whatever that's worth. Maybe it’s placebo, but hey, I’ll take that too. ¯\_(ツ)_/¯

The one thing that has been pretty consistent is the vivid dreams. Even if it didn’t help with anything else, I'd honestly be tempted to take it recreationally once in a while just for the dreams.

🏦 💰 Business & Investing 💳 💴

📊 GPU Rich vs GPU Poor Leaderboard 🥇🥈🥉

Read more

Every skill you have today was once unknown to you.

—James Clear

📅🌭🚢💸 I was discussing newsletters with a friend who’s better at this stuff than I am. He said, and I quote: “are you competing against Costco’s hotdog? 😅 I think you should definitely raise price.”

I thought, when’s the last time I did that? 🤔

Oh, right, never.

So after not touching it since 2020, I’ve decided to adjust subscription prices for the first time to catch up with inflation. According to official CPI, each dollar now only has 79.2% of the value it had then.

Here’s the deal:

• I’m making the change on May 1, 2026. ⏳

• Monthly goes to $15 and annual to $150, from $12 and $120.

• If you want to lock in the old prices for a year, you can get an annual subscription before May 1st, and for 12 months, you’ll get the old, 2020-era price. 🔐

• If you’re an existing supporter, the new price kicks in at your next renewal. If you’re monthly, you can also switch to annual before May 1st to lock in the old price (on top of the usual 17% discount for going annual).

Thank you for your support 💚 🥃

Without you, I’m just a guy in pyjamas going clickety-clack on the keyboard, sending bits into the void 🌀

🏦 💰 Liberty Capital 💳 💴

🍎 Is Apple’s Air Traffic Controller Strategy Working? 🧑‍✈️✈️

One of these is not like the others 👆

Did Apple go from AI loser to potential AI winner by… Thinking Different? 😬

Everybody else is turning free cash flow into compute capex as fast as they can. Jensen is smiling. 😎 Apple is doing almost none of this. If we’re in a world where the first to AGI gets rewarded with World Domination™️, that strategy looks crazy.

BUT

If Ben Evans is closer to correct, and the models — as useful as they are! — are commoditized and function as more of a sustaining innovation, maybe it will pay more to own a differentiated chokepoint elsewhere in the value chain. Say, a software/hardware platform with a lock on the top end of the market, like the iPhone and Mac.

If you can’t own the planes, own the airport 🛫

Recent news suggests that Apple will open up Siri to multiple AIs in iOS 27, turning them into suppliers who compete for access to valuable customers.

It has also been reported, though not confirmed, that Apple would not only use Gemini, but that their deal with Google also allows them to distill it into its own custom models.

This reminds me of Edition #487 (all the way back in the Mesoproterozoic era of 2024!), where I wrote:

“In the tradition of commoditizing complements, I expect Apple to view AI as something that makes a product better, not the main course. Their priority is to sell more iPhones, not monthly subscriptions (those will come if they sell products).

They should frame it something like this:

Others are great at making foundational models. We want to focus on the *experience* of using these models. So we’ve created Apple Air Control™️, an AI that sits on top of other models, abstracts them away, and picks the best tool for the job, all with a beautiful and intuitive interface!

Do most users care which model is under the hood as long as it is really good and useful? [...]

If I were Apple, I’d go with a constellation of models. Fully modular.

I’d start with small models that can run on-device to take advantage of the very powerful silicon in iPhones and ARM Macs. They have fast GPUs and Neural Engines that are great at accelerating generative AI and their unified memory architecture means that more RAM is available for genAI than for most of the competition.

If simple tasks are run locally, you get lower latency and save on cloud inference costs.

Anything complex can be sent to larger models in the cloud. That’s where an Apple traffic cop model 👮🏻‍♂️ would do triage: some tasks would go to open-source models like Mistral or LLaMA and others to OpenAI, Anthropic, or Gemini. 🤖🤖🤖🤖

By using multiple models, they could probably get better pricing and potentially better performance by selecting the model that is best at any one moment or for a specific task. Things change so fast that today’s leader may be tomorrow’s laggard.

This would also keep the spotlight on Apple’s brand, not on the models (ie. from a user’s perspective, you’d think of going to Siri AI, not ChatGPT or Gemini).”

A lot has changed since 2024, but this approach could still make sense, especially in the consumer area where Apple is strongest (they won’t compete with Anthropic for coding in the enterprise, even though they’ll be very happy for all vibe coders to run Claude Code from MacBook Pros 💻 and OpenClaw on Mac Minis 🦞).

😲🚀 Anthropic “Surpasses” $30bn ARR + Signs 3.5 Gigawatt TPU Deal with Google & Broadcom 🐜

Speaking of big capex, Anthropic went from about $9bn ARR at the end of 2025, which is less than four months ago, even though it feels like more because of recent time dilation, to >$30bn ARR today 🤯

That’s more revenue than Google had in 2010 (though run-rate vs actual, tbf). Except that at the time, Google was growing 20-30% per year, not per month.

Even if the way they recognize revenue is a bit different from OpenAI, it’s clear who’s the unstoppable bullet train right now 🚅

When we announced our Series G fundraising in February, we shared that over 500 business customers were each spending over $1 million on an annualized basis. Today that number exceeds 1,000, doubling in less than two months.

Even on a log graph, it looks almost like a non-log graph 😅

Obviously, the Pentagon declaring them a supply chain risk has supercharged their growth, which is probably the exact opposite of what someone would have expected if you had told them 2 years ago that, like ZTE and Huawei, a major US company would be blacklisted ¯\_(ツ)_/¯

Anthropic has clearly been suffering from a compute crunch. Their usage limits have long been SIGNIFICANTLY less generous than OpenAI’s, and that has just gotten worse lately, with time-of-use restrictions that make limits melt faster during peak time. 🧊

So, who has compute capacity to sell by the gigawatt right now?

Two days ago, Broadcom filed an 8-K:

Separately, Broadcom, Google and Anthropic PBC (“Anthropic”) have expanded their current strategic collaboration under which Anthropic, beginning in 2027, will access through Broadcom approximately 3.5 gigawatts as part of the multiple gigawatts of next generation TPU-based AI compute capacity committed by Anthropic.

This seems win-win-win.

• Broadcom loves the big new customer. 💰

• Google would ideally love to win it all directly with Gemini, but if they can’t, especially in the enterprise, they’d rather it be Anthropic that wins (they own about 14% of it) running on TPUs in GCP than OpenAI, probably running on Nvidia GPUs not on GCP (OpenAI also has compute deals with Google, but they’re not as close to them as Anthropic is). It also helps Google that OpenAI has to divert resources from consumer to the enterprise to compete with Claude.

• Anthropic gets more compute and further diversifies, which may give it more leverage during its next negotiation round with Jensen. They also have a deal with AWS for Trainiums, and I wouldn’t be surprised to see them do something with AMD soon.

But it won’t come online soon enough to ease the current compute crunch, and with Mythos waiting in the wings… 🤔 (more on that in the next Edition)

🥇📋 Meta's Token Use Leaderboard (What could go wrong?) 🏆🤔

When I saw this, I immediately knew what was wrong with it:

The rankings, set up by a Meta employee on its intranet using company data, measure how many tokens—the units of data processed by AI models—employees are burning through. Dubbed “Claudeonomics” after the flagship product of AI startup Anthropic, the leaderboard aggregates AI usage from more than 85,000 Meta employees, listing the top 250 power users. [...]

Workers are maximizing their prompts, coding sessions and the number of agents working in parallel to climb internal rankings at Meta and other companies and demonstrate their value as AI automates functions such as coding.

Some people have reported hearing that token use is also part of employee performance reviews 🤔

Employees can track their personal consumption, compare themselves against colleagues and earn gamified rewards ranging from bronze, silver, gold, platinum and emerald badges to achievement titles, which also include “Model Connoisseur” and “Cache Wizard.” Some workers are instructing AI agents to carry out research for hours on end to maximize their token usage, according to two current employees.

Goodhart’s law:

“When a measure becomes a target, it ceases to be a good measure.”

Campbell’s law:

The more a quantitative indicator is used for decision-making, the more subject it is to corruption pressures, and the more it distorts the process it’s meant to monitor.

Over a recent 30-day period, total usage on the dashboard topped 60 trillion tokens… The highest-ranked individual user averaged 281 billion tokens, which could cost millions of dollars, depending on the type of model used.

60 trillion tokens a month puts Meta as one of Anthropic’s biggest customers, if not THE biggest.

It reminds me of measuring programmers by how many lines of code they produce 😬

The real thing you want is some technical or business outcome. You want better code, not more of it. If you reward quantity, you will encourage a lot of wasted tokens/compute, and more bloated/complex solutions to problems.

💳 Visa Launches AI Tools to Manage Charge Dispute Process

This isn’t as sexy as AGI headlines, but considering how much of the economy Visa touches, this is probably one of the largest-scale and most economically important real-world deployments of AI:

[Visa] announced six new dispute resolution tools designed to reduce the billions of dollars lost annually to inefficient, outdated dispute processes. The expanded suite of dispute resolution services is being designed to help merchants and financial institutions cut administrative costs, reduce fraud-related losses and redirect those resources toward growth, innovation and customer experience.

Disputes remain one of the most persistent friction points in commerce, driving rising costs for merchants and financial institutions while simultaneously leaving consumers frustrated and confused. In 2025, Visa processed 106 million disputes globally, a 35% increase since 2019. (Source)

I expect all financial players to have similar things (banks, insurance companies, etc).

I just hope it’ll work better for their customers than the incredibly annoying automated phone systems that leave you stranded the moment your case falls outside the decision tree.

💰💰⛰️💰💰 Berkshire Hathaway’s Cash Pile Mountain

Berkshire's market cap went from ~$15bn in 1991 to a trillion today. So naturally, the absolute cash pile has grown. But cash as a percentage of market cap has also been rising, which is more interesting (Brooklyn Investor has written great stuff about this).

But capital deployment takes place in absolute dollars, and clearly it has become harder to put money to work at high expected rates of return. Buffett has been quite clear on that.

Most of the other Trillion Club companies — except Apple, as we saw above — have been plowing their cash into AI capex lately (and it’s not yet clear what kind of ROIC they’ll get over time). If they weren’t doing that, they’d probably spend most of their FCF on buybacks and dividends.

Will Greg Abel be able (accidental pun) to figure out what to do with the mountain? ⛷️

📊 Moody’s and S&P: Wide Moats vs. AI Disruption 🤔

My friend David aka Scuttleblurb (check out our most recent annual interview) wrote good pieces about these two classic wide-moat companies and how they might face the new risks caused by AI. I recommend reading both, but I have a few highlights below:

• 📄 S&P, Moody’s, and the limits of AI disruption: Part 1 (🔐)

• 📄 S&P, Moody’s, and the limits of AI disruption: Part 2 (🔐)

Here are my highlights:

The durability of their credit ratings have less to do with the relative accuracy of default and loss assessments than by the role they play as a ubiquitously adopted benchmark across the financial ecosystem. A startup could conceivably use AI-powered algorithms to create a ratings methodology that was just as accurate and Moody’s and S&P’s would still remain entrenched.

A similar logic holds for S&P’s equity index business.

The network effect of standards is very hard to dislodge, especially when you save more by using the standard than by not using it (the old saw about rated debt being less expensive, so whatever you pay in fees, you get back in lower interest).

Speaking of debt:

The credit ratings business should be buoyed by several durable growth themes over the foreseeable future. Some $5tn of debt needs to be refinanced over the next four years, roughly double the dollar volume seen in 2018. The swelling stock of private credit will increasingly demand risk scores too, as regulators and investors press for comparability with public debt markets; Moody’s private credit ratings revenue grew 60% last year. Infrastructure funding, including AI data centers, supplies a further tailwind. Ratings revenue is lumpy and unpredictable year to year, but over extended periods has reliably tracked the high-single digit growth rate that management has long guided to. I don’t expect the future to look much different from the past.

🌊

I like the idea of not thinking about these companies as monoliths:

When thinking about S&P's data estate, I see a hierarchy of AI defensibility. At the top is data that you alone create and wouldn't exist without you, like a credit rating. Below that is data obtained through private sources, like the loan price data that S&P gets fed to it from dealers or shipping data sourced through bespoke commercial agreements. Further down still are curation activities – like entity matching and de-duplication, and data normalization – which I see as jobs that AI will eventually perform better and cheaper than any human team. I don't consider those capabilities a moat long-term. What does matter at that layer, however, particularly in compliance or regulatory driven domains, is governance infrastructure and SLAs. Where S&P controls the chokepoint, whether in the form of proprietary or near proprietary data or contractual assurances that clients cannot do without, AI should be a source of leverage rather than fear.

And finally, a devastating meta-observation coup de grâce:

I find it a bit absurd that we're entertaining the notion that a dominant global franchise like S&P is at risk of AI disruption even as we blithely labor under the delusion that we investment analysts will be safe.

😅

🧪🔬 Liberty Labs 🧬 🔭

☔ Error Bars for Weather Predictions (Finally) 🌈☀️

My favorite weather app is still Mercury. It fits how my brain thinks about weather. The UX is just clean and well thought out.

But I gotta admit that this new feature from Acme Weather, which is made by some of the people who made Dark Sky weather before it was acquired by Apple, is really cool. Their weather line makes the forecast uncertainly visual by showing the spread widen or tighten depending on various factors.

It’ll take a lot to make me switch from Mercury, but I’ll give Acme a try.

🏴‍☠️🔓 Websites Can Show AI One Thing and Humans Another

Well, this isn’t great. A paper published by Google DeepMind details a vulnerability of AIs that feels obvious in retrospect, but hasn’t gotten nearly enough attention.

They identify six kinds of attacks:

• Content Injection Traps that exploit the gap between human perception, machine parsing, and dynamic rendering

• Semantic Manipulation Traps, which corrupt an agent’s reasoning and internal verification processes

• Cognitive State Traps, which target an agent’s long-term memory, knowledge bases, and learned behavioural policies

• Behavioural Control Traps, which hijack an agent’s capabilities to force unauthorised actions

• Systemic Traps, which use agent interaction to create systemic failure

• Human-in-the-Loop Traps, which exploit cognitive biases to influence a human overseer.

For example, the first one could be a website that embeds commands via CSS, HTML comments, or metadata attributes invisible to humans but parsed by agents. 👀

Or maybe a dynamic website that detects when the “reader” is an AI and serves it different content than when it’s a human, making it even harder to detect. It’s like a parallel reality that only AIs can see, or maybe only certain targeted AI clients to make it even harder to detect.

Hidden commands could take advantage of agentic workflows to spawn sub-agents that then go on to do the attacker’s bidding while the main agent appears to be doing just normal stuff.

Defenses will need to be hardened, but as usual, the defender needs to be right every time while the attacker only needs to get through once, so it’s an asymmetric combat dynamic.

👵🧑‍🦳 The Steady Rise of Maximum Life Expectancy Since 1840 📈

A lot of stuff has been stagnating (especially since 1971), but we gotta celebrate the wins too. 🥳

Experts have called the biological maximum multiple times since 1840. They've been wrong every time. Global record life expectancy has increased by 1 year every 4 years since 1840.

As you can see from the pink dotted lines, predictions have been exceeded multiple times (from 1986, 1999, 2001..). Let’s keep it going and accelerate it by curing the diseases of aging to extend robust health. But the goal isn't just to push the record higher, it's to broaden the base. More people living healthier for longer, not just record-setters

🦠 40 Years of Multiple Myeloma Cancer Survival Rate Improvements 🧫

Speaking of celebrating victories, I love this graph.

It shows the survival duration in months for Multiple Myeloma over the past four decades. From less than 20% survival after a year in 1986 to more than 80% survival after three years today! Science FTW!

The Mayo Clinic describes it like this: “Multiple myeloma is a cancer that forms in a type of white blood cell called a plasma cell. Healthy plasma cells help fight infections by making proteins called antibodies. Antibodies find and attack germs. In multiple myeloma, cancerous plasma cells build up in bone marrow. In the bone marrow, the cancer cells crowd out healthy blood cells. Rather than make helpful antibodies, the cancer cells make proteins that don’t work right. This leads to complications of multiple myeloma.”

A different cancer, a similar story: The improvement for Stage 4 melanoma (stage 4 means that it has spread to other parts of the body) is also dramatic:

Remember that these aren’t just lines, they are people, and what happens to them affects a wider circle of family and loved ones.

h/t Dr. Samuel Hume

🎨 🎭 Liberty Studio 👩‍🎨 🎥

👨‍🚀🥔 Andy Weir in 2017: How I Became a Writer on the Side

Weir is the man of the hour, thanks to the success of the Project Hail Mary adaptation. He has an interesting and non-traditional path to success: A programmer by trade, he decided to publish his fiction online for free and built up a small but enthusiastic core fanbase over a decade+ before getting traction.

I like the “scratch your own itch” approach to being as scientifically accurate as possible in his fiction. Most people may not care, but the ones who do — they really do. And it makes him stand out since most other writers handwave most of that stuff.

It reminds me a bit of why I run this steamboat the way I do. I write the way I want to read, and hope that there are others like me out there who may find it hits the spot. To me, that’s a better approach than trying to imagine some theoretical audience out there and trying to guess what they may like. 🤔💭

The last part of this talk, where he talks about his involvement with the Martian’s film adaptation and what NASA and JPL subsequently did after he published, was really fun. I wish he would do a follow-up talk or podcast about his experience with Project Hail Mary’s adaptation.

h/t Friend-of-the-show Trung Phan

📙🔭 The Science of Not Giving Up 🚀

Speaking of Andy Weir, my friend Jameson just released part 1 of an episode of Becoming the Main Character on Project Hail Mary (the book).

Science does not know its debt to imagination.

—Ralph Waldo Emerson

👨‍🚀🚐 🚀 Every astronaut who rode the Space Shuttle to the launchpad did so in the same 1983 Airstream motorhome. It’s a modified Excella RV with bench seats and holes cut in the floor for suit-cooling units. They called it the Astrovan, and it was in service from 1983 to the final Shuttle mission in 2011 (👆).

The best part? Astronauts love traditions and actively argued against upgrading or replacing it even after it was pretty old.

Airstream also built the Mobile Quarantine Facilities for Apollo 11, 12, and 14 (those silver trailers where astronauts waved at Nixon through a window after returning from the Moon). When the Artemis program needed a crew vehicle, the original plan was custom electric vans from a startup (Canoo) that went bankrupt, but they ended up going with Astrovan II: an Airstream Atlas on a Mercedes Sprinter chassis, built in Jackson Center, Ohio, 14 miles from the Neil Armstrong museum. 👇

Reid Wiseman, Victor Glover, Christina Koch, and Canadian astronaut Jeremy Hansen rode it down the road to Pad 39B for Artemis II — the first crewed lunar mission in over 50 years, expected to fly by the Moon on Monday (🌔). The most advanced rocket NASA has ever built, and the ride to get to it is a modified Airstream.

Here’s the orbital mechanics involved in the translunar injection (don’t those words just sound cool?):

(Note: if you don’t see the animation above in your email app, you may need to click through to the website)

🏦 💰 Business & Investing 💳 💴

👀🚙🤖 Waymo: From 10,000 to 500,000 Rides in Less Than Three Years 🚀📈

Read more

Innovation is the central issue in economic prosperity.

—Michael Porter

🛀💭🚪🤝 Alec Stapp writes:

Landlord offered to renew my lease at the same rent for another 12 months.

Usually I wouldn’t spend the time to negotiate if they’re not trying to raise the rent, but figured I’d let Claude have a go and negotiate on my behalf.

Claude did a market comp analysis and drafted the counteroffer for me.

Landlord just came back and agreed to an 8% decrease in my rent.

Thanks, Claude.

Let’s just say the $20/month sub pays for itself :)

What happens when his landlord has an AI negotiate back, and it’s just a bunch of AIs negotiating with each other?

What happens to outcomes? Do they converge faster? Do models collude? Will someone fine-tune AIs that specialize in a specific kind of haggling based on exploiting known weaknesses in the popular chatbots most likely to be used on the other side? 🤔

🤖🔎↔️ Speaking of persuasive AIs, be careful where you point that persuasiveness, because it'll work on you too.

Andrej Karpathy writes:

- Drafted a blog post
- Used an LLM to meticulously improve the argument over 4 hours.
- Wow, feeling great, it’s so convincing!
- Fun idea let’s ask it to argue the opposite.
- LLM demolishes the entire argument and convinces me that the opposite is in fact true.
- lol

The LLMs may elicit an opinion when asked but are extremely competent in arguing almost any direction. This is actually super useful as a tool for forming your own opinions, just make sure to ask different directions and be careful with the sycophancy.

Always steelman multiple positions before you make up your mind. Otherwise, your AI-polished conviction might just be the last direction you pointed the model.

In fact, don't leave it to memory. Add a custom instruction so your AI does it automatically, every time, without you having to think about it.

📬💡📺🎬📚❤️‍🔥 OSV Field Notes is a weekly, high-signal curation of things worth your time. It’s written by my OSV colleagues and me. It exists because there’s so much good stuff out there that never makes it to the algo-controlled slopfests on most apps.

It’s maybe too effective, because every week, I’m adding new things to my reading list and watchlist 😅

🔎📫💚 🥃 Too few readers of this newsletter are paid supporters, and it’s threatening the project's existence.

I’m not saying this to be dramatic. I just want to level with you.

Paid support has declined, even as total readership has continued to climb and now approaches 30,000. 📉📈 (Update: Thank you to those who became supporters in the past couple of weeks ♥️)

If you want it to continue, this is the moment. Become a paid supporter 👇

🏦 💰 Liberty Capital 💳 💴

💪 ARM CPU: We’re not just IP anymore! 👨‍🔬🍰

The least secret secret in the industry was that ARM was working on making its own chips. When Rene Haas became CEO, he clearly wanted ARM to take a bigger slice of the value it creates, and this partial change of business model is the result of that.

As AI is going agentic, the world will need a lot more CPUs. And the ideal chip for AI agents isn’t necessarily the same as before. Nvidia’s Vera CPUs (also ARM-based) are tuned for AI agents, whose usage patterns look nothing like a human hammering the same system. ARM has clearly been thinking along similar lines.

They call it the ARM AGI CPU, and it’s designed to offer a crapload of cores in a relatively low power envelope:

The Arm AGI CPU is designed to deliver high per-task performance at sustained load across thousands of cores in parallel – all within the power and cooling limits of modern data centers. [...]

Arm’s reference server configuration is a 1U, 2-node design – packing in two chips with dedicated memory and I/O for a total of 272 cores per blade. These blades are designed to fully populate a standard air-cooled 36kW rack – 30 blades delivering a total of 8160 cores. Arm has additionally partnered with Supermicro on a liquid-cooled 200kW design capable of housing 336 Arm AGI CPUs for over 45,000 cores.

One notable architectural choice is ARM’s "dedicated core per program thread" model, which the company says delivers 'deterministic performance under sustained load by eliminating throttling and idle threads. This matters when you’re running thousands of parallel agent tasks that each need predictable latency over peak burst performance.

In this configuration, the Arm AGI CPU is capable of delivering more than 2x the performance per rack compared to the latest x86 systems. [this is ARM's figure — Intel predictably disagrees. -Lib]

Arm said Meta will be the chip’s “lead partner” and “co-developer” — one more option for the compute swing voter — with other early customers including OpenAI, Cloudflare, SAP, SK Telecom, and Cerebras.

So many companies were part of the announcement. ARM is clearly threading a needle here, being careful not to compete head-on with its own customers (at least at this stage), and is instead targeting an underserved niche. For now.

See how I resisted making ‘arm’ puns. 😅

😇📈 From Magnificent 7 to Hateful 8 👿📉

The stocks that drove every S&P gain for two years are now driving almost all of its decline.

Paul Kedrosky made this graph to visualize how much of the S&P 500’s decline is due to the formerly-could-do-no-wrong Big Tech (+Oracle):

The “Hateful Eight” — the Mag 7 + Oracle that drove index gains for the past two years — account for 85% of the year-to-date drawdown, subtracting roughly 576 points. The other ~490 companies collectively added about 100 points.

Speaking of concentration: Nvidia alone represents 35% of the S&P’s Q1 earnings growth, while the much smaller Micron ($370bn market cap vs $4.2 trillion for Nvidia) represents 18% thanks to its holy-memory-cycle 682% EPS growth:

h/t Bob from Accounting

🗣️💬 Benedict Evans: 900 Million Users, and Then What? Why Better Models Don't Solve the Real Problem💡

This is to be paired with my highlights from a piece that Ben Evans wrote explaining some of the challenges facing AI labs and the ecosystem that supports them.

I find Evans worth listening to because, in Claude Shannon's terms, he is providing real information. He’s saying things that are different, and sometimes opposite, to what a lot of the AI talking heads are saying these days, and that’s valuable. I make an effort to seek out the more skeptical POV on AI to avoid being in an echo chamber, and I think he’s one of the most reasoned skeptics (unlike many of the others that I won’t name because I want to keep this positive).

Here are a few highlights, but the whole thing is worth your time:

OpenAI may have massive mindshare, but not a durable moat 🏰

Ben Evans: That’s the puzzle. This great strategic problem for OpenAI is you’ve basically got commodity technology. You don’t have a network effect. You’ve got 900 million weekly active users, but most of them are not using it every day and can’t think of anything to do with it. Only 5% of them are paying for it. The usage is like a mile wide but an inch deep. You’ve got to swap the mindshare and momentum you have for something more durable.

They did this video sharing app at the end of last year that didn't work. There's an App Store—or are we on the second or third App Store now? I've lost count. There's an e-commerce integration, but guess what? E-commerce integrations are really hard and complicated because there are millions of merchants and billions of SKUs. Google and Meta have failed at that twice. It turns out that OpenAI went, "No, we probably can't build that ourselves right now." So there's this challenge of how you get from having mindshare and one of the good models to having some kind of a durable platform business where you've got developers, users, and corporate accounts—something that's locked in and doesn't just depend on you having the best model next week.

As a power user of multiple AI products, it's easy to forget: for most people, usage is shallow, differentiation feels small, and they don't really know what to do with it.

The world outside my info-circle is still very different.

AI product teams are reacting to research, not steering from user needs 👨‍🔬🧪

Ben Evans: The way it works is you turn on your computer in the morning and you’ve got an email from the research group that says, ‘Hey, guess what? We’ve got this cool thing.’ Then your job is to go and do something with it. You get the email saying we’ve got a new voice model, and it’s like, ‘Okay, I guess we’re adding a microphone button today.’ You start from the technology; you don’t control the product strategy. That’s how science works, but you don’t know what’s going to get built.

…This is inherent in where we are: this technology is continually shifting, and you don’t know what it will be able to do next week. You’re a strategy taker, not a strategy setter.

Obviously, Sam and Dario and so on are setting that fundamental research strategy, but the thing comes back in six months and it may work or it may not. I opened the last essay I wrote taking that quote from Simo and comparing it with Steve Jobs famously saying you can't start with the technology and work forward to the product; you've got to start with the user experience and work backwards to the technology. This is inherent in where we are: this technology is continually shifting, and you don't know what it will be able to do next week.

This is something we’ve seen a lot in the past, especially with the Apple vs tech research labs. Research-driven labs come up with very impressive stuff. They make great demos. But rarely products.

Starting with real user needs — even ones users don't know they have yet (the old Henry Ford "faster horse" line 🐴) — is what makes the best products. And the most defensible.

AI will likely mean more software, not less 💾

Matt Turck: To the point you're making, presumably you're not in the camp of believing that people are going to be building their own software with Claude Code.

Ben Evans: […] AI coding means coding is way cheaper and easier. There’s a whole bunch of stuff you couldn't do in software that now you can. There will be way more software, and that will pick up many more of those use cases that weren't automated before, either because they were too small or because you actually couldn't automate that thing with traditional software. [...]

Matt Turck: So what do you think happens to SaaS?

Ben Evans: Way more software. But as an independent category in stock markets? We have to unpick that. If it’s much cheaper to write software, there will be more of it. Who will be doing that? It’s a combination of who understands how software works and who has understood the problem. Frame.io was not made by some person working at a video production suite because that’s not how they think; they’re not product people. There’s a different skill to working out what the code should be doing.

There will be way more software both for stuff that doesn’t need AI except to create it, and stuff that needs AI to do the new thing. Some incumbent software will go away—big expert systems where an LLM can do it better. This seems like a repeat of SaaS. SaaS meant you had two orders of magnitude more software, and some incumbents got completely screwed. When we went from mainframes, the big company had five pieces of software. With on-prem, you had dozens. With cloud, you have hundreds. You should presume with this you’ll have way more software. [...]

Most SaaS companies are database wrappers where somebody realized there is a problem and a way of turning it 90 degrees. The hard part was not writing SQL queries on AWS. The hard part of writing software is all the other stuff: what should the code be doing, how do we tell people to use it, and what should we charge?

Cheaper software creation expanding the overall surface area is fairly consensus. Coding being the least important part of building a successful software company is much less so. Just look at public software valuations right now. Time will tell.

📺 Inflation-Adjusted Netflix pricing since 2014 📊

The idea for this came from Paul Crowley. He asked Claude to create a graph of Netflix pricing in the U.S. over time, along with the inflation-adjusted prices in today’s dollars.

Because Claude usage limits are so tight these days, I asked Codex to build a more elaborate version for both the Standard and Premium options:

Look at that inflation post-2020!

The adjusted Standard pricing went from $12.35 in 2014 to $19.99 in 2026, and from $16.47 to $26.99 for Premium. That’s 62% and 64% increases, respectively.

Do you think that’s a lot?

Honestly, probably a good value considering how Netflix’s product and catalog have evolved since 2014, and how expensive other forms of entertainment are. You can watch films and TV shows with your family for a whole month for less than one restaurant dinner or a night at the theater.

I suspect that for most Netflix customers, it’s one of their lowest cost/hour sources of entertainment.

I just wish Netflix made better original content. When’s the next Dark (2017), or Bojack Horseman (2014), or Narcos (2015), or The Queen’s Gambit (2020), or Mindhunter (2017) coming out? It’s been a while since they made things on that level ¯\_(ツ)_/¯

🧪🔬 Liberty Labs 🧬 🔭

📚🪑☕ Ken Isaacs' Superchair from 1955 #LifeGoals

What more can you ask for?

It was originally conceived in 1955 and then formalized for production in 1967:

Built-in book rest, shelves, lamp, drink tray, and a seat back that folds into a bed.

A place for "inventive work and the individual search for peace of mind", as he put it.

It was meant for people to build it themselves, hence the almost unfinished look. Blueprints were published in Popular Science in 1968.

Hey IKEA, are you seeing this?

h/t Oliver

‘Total AI chip memory bandwidth has grown 4.1x per year, now reaching 70 million TB/s’ 🐜🐜🐜🐜🐜🐜

How much aggregate memory bandwidth is that?

“That’s around 300,000x more data per second than global internet traffic.”

🤯

Nvidia's dominance is clear. Huawei barely registers, despite the effort.

“Why track this? AI inference is frequently bottlenecked on memory bandwidth, not compute. Aggregate HBM bandwidth is a rough proxy for the world’s total capacity to serve AI models.” Source.

💻 Microsoft: We Will Make Windows 11 Suck Less 💾

Maybe what it took for Microsoft to wake up and take Windows quality more seriously was for Apple to release a $600 laptop — the MacBook Neo — that reviewers are calling the best value laptop on the market.

In a blog post, MS committed to fixing a lot of annoyances for Windows users:

• Integrating AI where it’s most meaningful, with craft and focus: You will see us be more intentional about how and where Copilot integrates across Windows, focusing on experiences that are genuinely useful and well‑crafted. As part of this, we are reducing unnecessary Copilot entry points, starting with apps like Snipping Tool, Photos, Widgets and Notepad.

They finally understood that people were tired of getting Copilot crammed down their throats, Clippy-style! 👀📎

• Reducing disruption from Windows Updates: Receiving updates should be predictable and easy to plan around, so we’re giving you more control. This includes the ability to skip updates during device setup to get to the desktop faster, restart or shut down without installing updates and pause updates for longer when needed, all while reducing update noise with fewer automatic restarts and notifications.

That’s been going on for decades at this point!

• More taskbar customization, including vertical and top positions: Repositioning the taskbar is one of the top asks we’ve heard from you. We are introducing the ability to reposition it to the top or sides of your screen, making it easier to personalize your workspace.

What a concept!

• Improving system performance: Reducing resource usage by Windows to free up more performance for what you’re doing. Improved memory efficiency, lowering the baseline memory footprint for Windows. More consistent performance, even under load. Reducing interaction latency.

Why hasn’t this always been a priority?

• Increasing OS, driver and app reliability: Strengthening the Windows foundation by reducing OS level crashes, improving driver quality and app stability across our ecosystem so PCs run smoothly and reliably every day.

• Creating easier, faster, and more stable connections with Bluetooth accessories. Fewer USB-related crashes and connection loss.

Anyway, these are just my highlights, the full list is here.

This reminds me of when Apple released its Snow Leopard OS back in 2009. They marketed it as a ‘polish’ release that wasn’t focusing on new features, but rather on making what’s already there better and more reliable. I wish Apple would also go back to this.

🎨 🎭 Liberty Studio 👩‍🎨 🎥

🌒🎶 Returning to The Dark Side of the Moon & The Snob Sandwich 🥪

Growing up, I knew that Dark Side of the Moon was one of my dad’s favorites. He rarely listened to it, but I knew he considered it special. So when I got into music, it was one of the first things I listened to over and over again.

It blew my tiny mind.

It was unlike anything else I was hearing on the radio or finding in my dad’s records and CDs (this was before I started my own collection, which eventually grew to about 1,500 CDs before I switched to MP3s and then streaming).

As I discovered other things, I did what many music fans do and stopped listening to some of my early favorites. I think I subconsciously started seeing them as “for beginners,” the training wheels of music.

Even with Pink Floyd, I did what a lot of fans seem to do and got more into Wish You Were Here, Meddle, and Animals. I also found other prog and art-rock bands that were more obscure and therefore cooler.

I’ve heard this described as the ‘snob sandwich’ 🥪

It’s similar to the midwit curve: beginners on one side, experts on the other, and the snobs in the middle. The idea is that with many things, what beginners love often gets rediscovered by more experienced people later on, but only after a phase of looking down on it for being too obvious.

I lived this with the Beatles, one of my first musical loves. For years, I barely listened to them because they felt too “obvious.” But every so often, I come back and they blow my mind all over again.

The nice thing about coming back to music after so long is that I’ve changed, so I hear it differently. Back when I first heard Pink Floyd, I didn’t play guitar and had far fewer musical points of reference. On recent Dark Side listens, I noticed all kinds of wonderful guitar solos and basslines that went right over my head as a kid.

One thing that works against Dark Side now is that it’s really meant to be heard as an album. It doesn’t hit the same way when random tracks are sprinkled into a shuffled playlist, which is how a lot of people seem to listen to music now. There are deliberate buildups and releases.

The crescendos hit differently when you’ve earned them. 💥

Granted, some of the album’s gimmicks and long intros work better the first time you hear them than the 1,000th (those clocks 🕰️), but see it as training for your attention span.

When’s the last time you listened to The Dark Side of the Moon from beginning to end?

Why not today? 🌒

We see with our brains, not with our eyes.

—Norman Doidge

🔪🧳🏠 The dullest knives in the world can be found in Airbnbs (and also in my mother-in-law’s kitchen drawers 😅).

It's the trifecta:

• Owners buy the cheapest knives they can find, so they’re made of soft steel that dulls easily.

• Heavy use, often by people who do things you’re not supposed to do (don’t put your chef’s knives in the dishwasher!).

• There’s also a principal-agent mismatch: the owner who'd sharpen them isn't the one using them.

What to do about it?

I’m going to buy a portable knife sharpener and add it to my travel checklist. Every time I go somewhere and the knives are dull, I’ll bring them back to life. For myself, but also as a small public service.

I think it will be deeply satisfying.

📚Unread books on my shelf silently taunt me.

I keep buying more anyway.

Apparently, buying books and reading books are two separate hobbies, and I'm better at one of them 😅

The Japanese have a word for this: tsundoku, the art of buying books and letting them pile up. At least it sounds dignified.

🏦 💰 Business & Investing 💳 💴

🤖💭🗜️ Google’s TurboQuant: 6x Less Memory, Zero Accuracy Loss

Google researchers published TurboQuant this week, a compression algorithm that they say reduces KV cache memory requirements by at least 6x, without loss of accuracy. 🤯

Two things are interesting about it: 1️⃣ the technical implications and 2️⃣ the business implications:

Read more