580: Cloudflare Power Grab, OpenAI Open Models, DeepMind Genie 3, Amazon + Microsoft + Google + Meta Capex, Cardiovascular Victory, 8 Years of Déjà Vu, and Shrinking

"Trust me, bro"

Nobody ever gets credit for fixing problems that never happened.

—Nelson P. Repenning & John D. Sterman

🗓️📋✅ Once in a while, I like to look back and see how things turned out.

I don’t want this to be 12 pages long, so I’ll keep it to bullets and split it into parts (next part on Friday):

• 🍳 I still love my cast iron skillet, but I’m using the carbon steel and stainless steel ones much more now. For grilling meat, I feel like I get most of the benefits of cast iron from carbon steel, and for everything else, stainless steel is just so much lighter and easier to clean.

• 👟🥾 I also still love my minimalist/barefoot-style shoes and boots. I have three pairs that I cycle through with the seasons (you can see the winter boots and water-resistant trail shoes here). My daily ones are the Xero HFS, and three years later, the soles are starting to wear out near the heels. As soon as my size gets back in stock (12.5), I’m getting the new HFS II. I’ll report whether v.2 is a worthwhile upgrade over v.1.

• 🌬️🫁🚭🧼🧹 I’m still pleased with both of my Levoit air purifiers (more details on the bedroom-sized and the living-room-sized models I got). Their sensors are very sensitive: every time there’s smog from a forest fire, or I grill something on the stove, I can see the rise in particulate matter and how they automatically spin up to clean the air.

• 💧🚰 🍵 It hasn’t been a year yet, but the countertop reverse osmosis filter I got in February is now fully integrated into our daily lives. The instant-hot water for my green tea and my wife’s coffee is very convenient. While it’s a bit annoying to have to refill the reservoir, I’m not yet in a hurry to upgrade to a model that is hooked up directly to the water line. Someday.

• 🎶🎸 Having so much fun with my acoustic guitar, and still waiting for my electric (it wasn’t in stock, so I’ll have to wait a few more months). I play almost daily, even if just for 10 minutes. As a project, I changed the bridge saddle, the little piece of plastic where the strings rest, and swapped it for a bone saddle, which supposedly sounds better. I think I can hear a small difference, but it could be placebo ¯\_(ツ)_/¯ Anyway, it was a fun, inexpensive DIY project. After a bunch of research, I also decided on an amp: the Boss Katana 50W (Gen 3).

• 🏋️‍♂️ 💪 Basement gym: Huge life upgrade. I try to do a strength workout at least 2-4x per week, but it depends on travel and other things. The lower friction of not having to drive to a commercial gym makes it MUCH more likely that I train. I mostly stick to compound movements like deadlifts, barbell squats, bench presses, pull-ups, barbell rows, curls, etc. I’m not jacked or anything, but I’m stronger than at any other time in my life and slowly improving. I track things using the Hevy app.

I think that’s enough for today. More in part 2!

🪞📖🔮 If you missed the White Mirror limited-edition bundle, which sold out in three days (see Edition #574), fear not, the unlimited version is now available!

You can get yours directly from Infinite Books or on Amazon.

📫💚 🥃 Do it! 👇

---

🏦 💰 Liberty Capital 💳 💴

💪🌩️🚔 Cloudflare’s Power Grab: The Internet’s Self-Appointed AI Gatekeeper & Tax Collector 🤔

I generally like Cloudflare, as you can probably tell from the dozens of times I’ve written about them over the past 5 years (including a podcast about them with my friend MBI).

But I'm less sure about this one.

I’ve been thinking about it for a month, and I’m still not entirely sure what I believe 🤔

Let’s think it through together:

On July 1st, Matthew Prince published a blog post with the grandiose title of “Content Independence Day”.

Here’s the crux:

The problem is whether you create content to sell ads, sell subscriptions, or just to know that people value what you've created, an AI-driven web doesn't reward content creators the way that the old search-driven web did. And that means the deal that Google made to take content in exchange for sending you traffic just doesn't make sense anymore.

Instead of being a fair trade, the web is being stripmined by AI crawlers with content creators seeing almost no traffic and therefore almost no value.

Sounds pretty good so far, no?

Cloudflare stepping in to protect the little guy, restore the web’s quid pro quo, etc.

The hero Gotham needs! 🦇🦸

That changes today, July 1, what we’re calling Content Independence Day. Cloudflare, along with a majority of the world's leading publishers and AI companies, is changing the default to block AI crawlers unless they pay creators for their content.

Cloudflare’s original business was being a content delivery network/reverse proxy, so their servers sit in front of about one in five websites and handle trillions of requests daily.

They were purely infrastructure, providing technical services. Now they’re making internet policy. The web’s self-anointed AI gatekeeper and tax collector 🤴⚖️

The “on by default” blocking is just step one. The more ambitious next step is to bootstrap a content marketplace:

Next, we'll work on a marketplace where content creators and AI companies, large and small, can come together. [...]

We believe that if we can begin to score and value content not on how much traffic it generates, but on how much it furthers knowledge — measured by how much it fills the current holes in AI engines “swiss cheese” — we not only will help AI engines get better faster, but also potentially facilitate a new golden age of high-value content creation.

Here are my primary concerns

1. Most people don’t change defaults, especially not the least sophisticated, smaller creators. So whatever Cloudflare decides effectively becomes the de facto standard.

2. Cloudflare has very little leverage over the mega-giants like Google and Meta, but it can make life *hell* for small players and startups. This could further entrench the largest players and reduce competition.

3. Creating a system that ranks the value of content has the potential to reshape a large fraction of the world’s intellectual output. That's not something to mess with lightly. It’s another algorithmic black box that, even if it begins with the best of intentions, can evolve in all kinds of ways over time. Even if you trust the current founders of Cloudflare to be benevolent dictators, they won’t be around forever.

4. Incentives shape outcomes on all sides. As the tax-collector-in-the-middle, Cloudflare would be shaped by incentives to design a system that maximizes its revenue. They may be magnanimous while their business is thriving, but they could become more ruthless if growth slows.

5. While it sounds good to want to reward the content that is more valuable to AI, creators who make things that humans find valuable but that AI doesn’t may get a raw deal.

6. Users of AI tools will likely get a degraded experience. If fewer sources are accessible and AIs keep hitting locked doors, answers to queries won't be as good as they could be. When Cloudflare blocks automated agents & tools, everyday users become collateral damage.

7. AI products will get more expensive for users and builders. This added cost and complexity means smaller AI labs and product startups become even less competitive with Google and Meta.

8. By definition, the most valuable content will be very esoteric and difficult to produce. If it wasn’t, it would be widely represented in training data. This means that only *very few* creators will benefit from high rates. Most will probably make pennies, if that. I don’t see a “creator middle class” emerging from such a system. Sophisticated players who understand all the finer points of Cloudflare’s algorithm will find ways to capture the market for valuable content and flood each niche until the price collapses before moving on to the next one. It’ll be the new SEO arms race.

9. The market will be global, so the arbitrage between high-cost and low-cost regions will likely set the ‘value’ of content at levels too low for the vast majority of people living in richer countries to make a living through this kind of system. There are plenty of smart people in India, Vietnam, the Philippines, etc. Language is less of a barrier now that AI tools are great at translation. This system could create a kind of global “Scale AI”-style content sweatshop, whoever can produce what the AIs want at the lowest cost will end up doing it.

10. But even that probably won’t be able to compete with synthetic data for very long. Besides, frontier labs already have plenty of incentive to spend billions on high-quality synthetic data. Cloudflare isn’t really changing those incentives.

Bottom line: Cloudflare may be degrading access to information, making people’s AI assistants dumber, reshaping the content ecosystem in ways based on AI needs rather than human preferences, and making it harder for new entrants to compete with tech giants in exchange for… basically zero benefits to the median creator.

To be fair, the underlying problem Cloudflare is trying to solve is real and significant. My issue isn’t with Cloudflare’s intentions, which I assume are good, but with the potential unintended consequences of their approach.

There may be a world where it all works out: millions of creators get rewarded nicely, AI Labs are super happy because they get good training data that wouldn’t exist otherwise and what they pay for content isn’t that much compared to their other expenses (compute, billion-dollar AI researchers), and users get better and smarter products.

BUT

I see so many ways this could go wrong and throw a bunch of sand in the gears ⚙️ of an already weakened web without generating enough benefits to be a net good.

As long as they frame it as “Small creators vs Big Tech & AI Labs” it sounds pretty good, but if creators and AI users don’t really get much out of this, then the reality is more like “Cloudflare vs Big Tech & AI Labs” with everyone else as potential collateral damage.

The functional outcome could be a tax on AI that benefits Cloudflare and is a net negative for basically everyone else.

Perplexity is already getting blocked because they were reportedly trying to circumvent Cloudflare (which Perplexity is denying, and accusing Cloudflare of bad behavior/incompetence in turn). To be clear, that decision was made unilaterally by Cloudflare, the website owners didn’t decide to block anything. There are also reports that OpenAI’s ChatGPT Agent is getting blocked.

Cloudflare is kinda saying: “Trust me, bro. You thought I was infrastructure, but now I’ll be the internet’s border patrol & customs officer, and I’ll be grading my own homework.”

That’s a little much for me to be comfortable with 😅

Especially since Cloudflare is an ambitious company. I could easily see them move into other AI-related areas that would create even more conflicts of interest.

What if they release their own AI agents and bots? 🤔 I bet these bots won’t have problems accessing content. What if they create other kinds of tools and services for creators to use to “rank” better in the new AI content marketplace? More friction & fees.

As a business analyst, I see a way for Cloudflare to use its unique position on the infrastructure layer to generate revenue.

But as a Citizen of the Internet, I don’t like seeing the infrastructure layer do things like that. I’m glad ISPs and telecos don’t have free rein to inject their own ads on other people’s websites or collect a tax on all commerce on their pipes or whatever (and some of them have tried all kinds of stuff, but so far it hasn’t polluted the internet too much).

Cloudflare may be solving a real problem, but the cure risks being worse than the disease if it creates new chokepoints, favors the giants, and centralizes control at the infrastructure layer.

💾🏋️‍♂️ OpenAI’s Open-Weight Reasoning Models are STRONG

OpenAI finally lives up to its name (a little).

Many in the AI community were wondering how strong the open-weights model from OAI would be. My first reaction when I saw the benchmarks was that they must be very confident in GPT-5 if they’re releasing something that does almost as well as o3 and o4-mini on text (worth noting these new models aren't multi-modal, and benchmarks should always be taken with a grain of salt until confirmed by real-world use).

It’ll be interesting to see the ‘raw’ chain-of-thought of an OpenAI model for the first time, like with DeepSeek.

Also impressive is how small the models are for that level of performance (they are quantized in MXFP4, though, not the dense, base model). The 120b-parameter version can run on a single H100, and the 20b one can run on a laptop with just 16GB of RAM 💻

Each model is a Transformer which leverages mixture-of-experts to reduce the number of active parameters needed to process input. gpt-oss-120b activates 5.1B parameters per token, while gpt-oss-20b activates 3.6B

They created a ‘gpt-oss playground’ to test the models. You can even pick a reasoning level from low to high.

If they’re as good as they seem, they will reclaim open-source leadership from China, unless they leapfrog it soon. Most non-Chinese companies that want to go for self-hosted open models now have very little reason to pick a Chinese model.

They are also undercutting their closed models on price. Serving GPT-oss will likely cost a fraction of what they charge per token for o3 and o4-mini. But the pressure will likely be *even higher* on competitors, so it should be good for OAI.

I will be using gpt-oss-120b in the coming weeks and will report back on how much I like its “flavor” and how it compares to other models. Stay tuned.

🏗️ Capex: Amazon, Microsoft, Google, and Meta 🚧

I think in this case, a single image says more than a thousand words.

Note: This is QUARTERLY. These four have almost reached $100 billion in spending on capex every 90 days!

Think what this means out in the real world, the supply chains, the armies of construction workers, the heavy machinery, the concrete, the steel, the energy!

---

🧪🔬 Liberty Labs 🧬 🔭

🧞‍♂️ Genie 3: DeepMind’s Real-Time 3D World Simulation

I’ve known real-time, interactive 3D world generation was coming, but it’s happening faster than I expected.

If it lives up to the demo video, it’s extremely impressive. More details here.

🫀 The Silent Victory: Heart Disease Deaths Have Plummeted 75% Since 1950 ❤️‍🩹

Great piece by Saloni Dattani.

Here are a few of my highlights:

For much of history, heart disease was a mystery. Middle-aged adults often collapsed without warning, and doctors usually blamed “dropsy”, “apoplexy”, or simply “old age.”

It’s hard to make progress when you don’t even know the problem exists!

In 1945 — at just 63 years old — President Franklin D. Roosevelt was sitting for a portrait when he raised a hand to his head and whispered, “I have a terrific pain in the back of my head.” Minutes later, he lost consciousness and died from a massive brain hemorrhage — a consequence of uncontrolled high blood pressure and heart disease, which doctors at the time couldn’t treat.

This was the president of the U.S. with access to the very best medical care at the time!

In the United States alone, the age-standardized death rate from cardiovascular disease has fallen by three-quarters since 1950. [...]

What made this progress possible?

There wasn’t just one discovery or intervention that did it all. Medical breakthroughs — in detection, treatment, surgery, and emergency care — have made surviving cardiovascular disease much more likely. In addition, public health measures and lifestyle changes have stopped many people from developing it in the first place.

The graph above shows a timeline of various inventions and discoveries.

⏰ 8 Years of Déjà Vu 😬 ⏰ 8 Years of Déjà Vu 😬 ⏰ 8 Years of Déjà Vu 😬 ⏰ 8 Years of Déjà Vu 😬 ⏰ 8 Years of Déjà Vu 😬 ⏰ 8 Years of Déjà Vu 😬

Real-life Groundhog Day:

A student was forced to drop out of university after a bizarre case of chronic déjà vu left him unable to lead a normal life.

The 23-year-old even stopped watching TV, listening to the radio, or reading newspapers or magazines because he believed he had seen it all before.

He told doctors that he was "trapped in a time loop" and said he felt as if he was reliving the past moment by moment.

Like a panic attack that makes you panic more, feeding on itself, apparently the same can happen with deja vu:

Rather than simply the unsettling feelings of familiarity which are normally associated with déjà vu, our subject complained that it felt like he was actually retrieving previous experiences from memory, not just finding them familiar," said Dr Wells.

Most cases like this occur as a side effect associated with epileptic seizures or dementia.

However, in this instance it appears as though the episodes of déjà vu could be linked to anxiety causing mistimed neuronal firing in the brain, which causes more déjà vu and in turn brings about more anxiety.

🔄

"In relation to our case, distress caused by the déjà vu experience may itself lead to increased levels of déjà vu: similar feedback loops in positive symptoms are reported in other anxiety states e.g. panic attacks.

It’s a case from 2014, but it was the first time I heard about it.

I tried to look for an update, but none seems to be public. I hope he was cured and could get back to his life!

---

🎨 🎭 Liberty Studio 👩‍🎨 🎥

📺 ‘Shrinking’ is Quite Good 🎭

I had zero expectations for Shrinking (2023, Apple TV+).

I just wanted something light to watch before bed. No guns, no explosions, no car chases. Just likeable characters and a few jokes that made me smile.

I was pleasantly surprised. This is more than that.

It’s about 2/3 comedy and 1/3 drama, but both sides carry real weight.

In fact, I’ve enjoyed it so much that after watching S1 and most of S2 on my own, I restarted it from the beginning with my wife. She likes it too, it’s one of the rare shows that makes her laugh out loud, which I always enjoy!

I can tell it’s not for everyone, and if you can’t stand Jason Segel, it’s not for you. But I enjoy the father-daughter stuff, and Harrison Ford is funnier than I’ve ever seen him.

If you check it out, commit to at least 3 episodes. The pilot has some elements that don’t really carry forward, so don’t let it turn you off.