575: Apple + Perplexity?, Anthropic's Success & Brittleness, TikTok U.S., Palantir Goes Nuclear, Google Search, Creatine & Alzheimer’s, and Hans Zimmer

"our civilization’s bootloader for colonizing the stars"

I don't say we all ought to misbehave, but we ought to look as if we could.

—Orson Welles

Made with Midjourney 💚 🥃

🏦🚀🌌 The market is literally our civilization’s bootloader for colonizing the stars.

It allocates resources to those who are creating, refining, and scaling the technologies required to climb the tech tree until we reach the level necessary to become a multi-planetary species.

I don’t just mean creating rocket tech directly 🚀👩‍🚀

To get there, we need a certain level of civilizational infrastructure and wealth. That doesn’t happen without a market where labor and capital are channeled to productive uses and where the creation of useful knowledge is incentivized and rewarded over long horizons.

In other words, if the S&P 500 climbs up-and-to-the-right long enough, it reaches outer space. 📈🪐✨⭐✨🌌

🦷😁 I got Invisalign ‘braces’.

I’m very happy I got laser eye surgery about 15 years ago, and I figured I probably wouldn’t regret this other ‘body upgrade’ either.

I never had braces growing up, and while my teeth were never *that* bad, I have a few crooked ones that bug me. I’d been thinking about doing this for years.

I even studied Align Technologies as an investment and bought some back in 2017, though I didn’t end up holding it for that long for various reasons.

Here’s something not many people know: They run the largest 3D printing operation for unique physical products/mass-customized manufacturing in the world.

They have mega-plants in Mexico, China, and Poland. They print over one million unique aligners *per day*.

They don’t print the aligners directly. Instead, they 3D-print unique molds of teeth, and then thermoform biocompatible plastic over them. Each one not only has to very precisely match a person’s teeth, but the plastic also needs to be perfectly smooth along every edge because our mouths are extremely sensitive, and any sharp bit is instantly noticeable.

This means that somewhere in Juárez, a perfect 1:1 replica of my teeth emerged from a machine, among a sea of other people’s teeth… which is kind of strange to think about 🇲🇽🏭🏗️🦷🦷🦷🦷🦷

I’ve only been wearing my first aligner for a couple of days, but I can already feel myself adapting to it. For the first few hours, I felt like I had a lisp whenever I spoke, but within less than 24 hours, I was already starting to sound like myself again.

It’s interesting how fast our brains can rewire themselves to adapt to a change 🤔🧠

I also suspect I’ll be losing some weight in the next year. The added steps every time I want to eat something make it less likely I’ll be mindlessly snacking 😅 🥨🍪

💚 🥃 🙏☺️ If you’re a free sub, I hope you’ll decide to become a paid supporter.

You’ll unlock access to paid editions, the private Discord, and get invited to the next supporter-only Zoom Q&As with me.

If you get just one good idea 💡 per year, it’ll more than pay for itself. If it makes you curious about new things, that’s priceless.

Can't upgrade right now? Share something you enjoyed here with someone who might like it too. Drop a screenshot and a link in your group chat! 💬

---

🏦 💰 Business & Investing 💳 💴

🍎🤔💭 Should Apple Buy Perplexity? Let’s Think It Through

I was a big fan of Perplexity when they first emerged.

I even bought the t-shirt! 👕

I now realize it was mostly because the competition *sucked* at search. Hallucination was a bigger problem back then, and grounding answers to questions with search results made them *a lot* more reliable (via RAG, retrieval-augmented generation), so Perplexity was generally much more reliable than pre-search ChatGPT or Claude.

That was a strong reason to like them, and they executed brilliantly in those early days of the AI wave. But they weren’t able to create a durable advantage.

Now that all the big players have gotten much better at search and tool-augmented generation, Perplexity has kind of lost its relevance, at least to me.

BUT

They’re still excellent at AI product design and shipping good UX, which is a scarce skill that shouldn’t be underrated and something that Apple could really use.

It also aligns with Apple’s strategy of being a platform for other people’s models, and combining those with some of their own smaller, specialized models for specific tasks (running on-device).

Apple executives have held internal discussions about potentially bidding for artificial intelligence startup Perplexity AI, seeking to address the need for more AI talent and technology.

Adrian Perica, the company’s head of mergers and acquisitions, has weighed the idea with services chief Eddy Cue and top AI decision-makers, according to people with knowledge of the matter. The discussions are at an early stage and may not lead to an offer

Take this for what it’s worth. Big companies explore lots of deals, but most of the time, nothing happens. It was recently reported that Meta also looked at Perplexity.

This deal could be useful to Apple not only for search, especially if the courts strike down the long-standing partnership with Google, but I bet the Perplexity team could also help with Apple’s broader AI efforts.

The value would be mostly from product design, UX expertise, AI infrastructure skills, and a talent infusion.

This could help make Apple a better front-end/platform for other people’s models, something perplexity has experience with. If that user-facing front end is really polished and offers a better user experience, it could differentiate the whole platform. Just look at how the superior ChatGPT app drives its adoption, it’s not just about the model.

Considering what Meta has been doing lately and the AI talent gold rush, Apple buying Perplexity wouldn’t be the craziest acqui-hire. And because of Apple’s massive installed base of a billion+ high-value users, plugging Perplexity’s assets into its distribution pipe would immediately make them be worth more (or at least, it wouldn’t hurt the business the way Meta did to Scale AI).

I think they should probably do it.

It may also be a win-win for Aravind Srinivas compared to trying to remain independent and risk falling further behind the foundational model labs, which are increasingly integrating various products into their offerings and disintermediating wrapper companies.

Search is just one of many. Anything that gets a lot of traction risks becoming one more product integrated into ChatGPT or Claude (eg. Claude Code/Codex).

🧑‍💻🚀 Anthropic Hits $4bn Run Rate in July, But Is It More Brittle than OpenAI?

While Anthropic has largely ceded the chatbot space to OpenAI, there’s a lot of demand for tokens in the coding space:

[Revenue] reached a pace of $4 billion annually, or $333 million per month, up almost four times from the start of the year, according to people familiar with the finances.

This is largely driven by Claude being considered a top model for generating code and agentic tasks, and the explosion in usage of both Cursor (where it’s the most popular model) and Anthropic’s own Claude Code product.

While it doesn’t match OpenAI, the pace of revenue growth at Anthropic still makes it one of the fastest-growing companies at this scale, ever.

Here are the reported numbers:

• December 2024: $1B

• March 2025: $2B

• May 2025: $3B

• July 2025: $4B

I’m sure AWS is enjoying this too!

I’ll be curious to see if this is called out during Amazon’s next quarterly call, and whether they’ll have to increase capex to keep up with inference demand. I’ve seen reports that Claude’s latency and tokens/seconds have been degrading, likely because they’re supply-constrained and have to batch more users per GPU.

While focusing on this vertical is great for Anthropic right now, it also makes them more brittle than OpenAI or Google if they ever lose leadership in that niche.

Just imagine if Google or OpenAI releases a coding/tool-using model that is significantly better than Claude, and they maintain that lead for an extended time. Developers are pragmatic.

I doubt they have much sentimental loyalty to any specific model 💔 They’d likely switch en masse… that could wipe out a significant chunk of Anthropic’s revenue 💸

🇨🇳 📱⏳ Remember TikTok? 🤷‍♀️🇺🇸🧑‍⚖️

Such a weird saga.

One of the most important consumer companies in the world has been in purgatory for months, as a law banning it in the U.S. is not being enforced, and new ‘extensions’ to the ban keep getting announced on an ad hoc basis.

Here’s the latest on what is going on:

[ByteDance] has developed a plan to launch the new TikTok app, known internally as “M2,” to U.S. app stores on September 5, according to two people with knowledge of the matter. (The existing TikTok app is known internally as “M.”)

Under the plan, [U.S.] TikTok users will eventually have to download the new app to be able to continue using the service, although the existing app will work until March of next year.

They’re forking the whole thing!

The effort to migrate the app’s 170 million U.S. users to a new app comes as the Trump administration says it is getting close to an agreement for the sale of TikTok’s U.S. operations, allowing the app to continue operating in the U.S. Under that deal, a consortium of non-Chinese investors including Oracle is expected to buy TikTok’s U.S. business, while the app’s Chinese parent company, ByteDance, is expected to retain a minority stake. The Chinese government has to approve any deal, however, which remains the biggest hurdle. [...]

A person familiar with the matter said the group of investors involved in the deal has been finalizing the legal and financial details with ByteDance in recent months, and the group expects it will be announced when the Chinese government approves the sale

It’ll be interesting to see who this group of investors is, and whether they just all happen to be politically connected…

And the longer-term repercussions remain uncertain. As I wondered out loud a few months ago:

• What precedent is this setting? Could France force Meta to do surgery and amputate Facebook and Instagram’s French operations at some point because they got in a fight with the company? Or maybe the EU as a whole? How about Turkey? India? Indonesia? Brazil? 🤔

• We've built planetary-scale applications that remain vulnerable to divergence in national interests and geopolitical tensions. The TikTok situation might just be the first of many.

⚛️ Palantir gets into… Nuclear Energy (!?) 🏗️👷📲🚧🔩

Palantir has partnered with the aptly named The Nuclear Company to help reboot gigawatt-scale nuclear construction in the U.S.:

Together, the companies will co-develop and deploy NOS, the first AI-driven, real-time software system built exclusively for nuclear construction. NOS will transform the construction of nuclear reactors into a data-driven, predictable process, enabling The Nuclear Company to build plants faster and safer for less.

China is proof that it’s possible to build nuclear power plants quickly and relatively on budget:

The partnership comes as China continues to outpace the United States in new nuclear, announcing 10 GW of reactors annually while America has built just 2 GW in the last three decades.

I guess Palantir would like to be the SAP of nuclear construction and position itself strategically for the coming nuclear renaissance. 🤔

They claim that NOS, which is built on Palantir’s Foundry, will help provide:

• Schedule Certainty: With NOS, construction teams will receive instantaneous, context-aware guidance — from the availability of certain parts and materials to the weather — that adapts to real-time constraints, so teams can work rather than wait.

• Cost Savings: A supply chain will track and verify all parts, as well as prevent shipment errors, material shortages and lost documentation. And when delays appear imminent, NOS will initiate backup options or prioritize other work in its place.

• Problem Prevention: Sensors placed across construction sites can feed data in real-time to a digital twin model of the site, allowing leaders to track progress with precision and compare what’s actually happening to the original plans. By using predictive analytics, teams can spot potential problems early, catching issues before they become expensive mistakes.

• Regulatory Confidence: AI will turn a traditionally labor- and time-intensive task to a process that becomes nearly instantaneous. Large language models can rapidly review tens of thousands of documents, while AI agents trained on regulatory requirements will help validate the data recorded automatically at construction sites.

The partnership is estimated at around $100 million. Palantir will embed a dedicated engineering team inside TNC.

If they can actually increase construction productivity and avoid delays and cost overruns, this service could more than pay for itself and provide a great service to the world.

*If* it works, it should be expanded to all mega-projects.

🌐🔍 Google Search: “The single greatest business ever created” + How Excite Almost Won (Podcast)

Great episode once again by friends-of-the-show Ben and David (💚 🥃) at Acquired.

They put it well: “We tell the story of the single greatest business ever created: Google search.”

This one is self-recommending, and I’m looking forward to the next one:

• 🎧 Acquired: Google

Here’s a great highlight about BackRub, the early implementation of PageRank:

they hammer out a deal that Excite is going to license this BackRub search technology from the two of them for about a million dollars. [...]

Larry and Sergey are going to come work at Excite that summer, implement BackRub for their search [...]

They get so far that they run a side-by-side test of Excite’s search results, the original algorithm and then the BackRub algorithm. The legend goes that they’re demoing this test to Excite’s CEO as a final step to finalizing this deal. [...]

the usual Excite search is bad. You have to click around, you go forward, you come back, you spend a lot of time on the site.

The CEO is like, why on earth would we move to your algorithm? I want people to stay on my site. I make money when people stay on my site. I don’t want them to leave my site. You guys are crazy. Get out of here. I’m killing the whole deal.

Can you imagine the alternate history where Excite licenses this tech, figures out CPC monetization, acquires the whole thing...?

Larry and Sergey tried to sell to Yahoo for $1m and were rejected, so it’s not like they weren’t interested in selling the company.

Maybe today we’d all be ‘Exciting’ things up (yikes, not as good a verb) 😬

---

🧪🔬 Science & Technology 🧬 🔭

🤖💭💡 Optimizing Intelligence Per Token

Here’s a metric that I hadn’t thought much about until I saw this on SemiAnalysis:

Anthropic models require significantly less tokens than other models to answer a question. […]

While this can depend on workload, Gemini 2.5 Pro and DeepSeek R1-0528 are more than 3 times as wordy as Claude. Gemini 2.5 Pro, Grok 3, and DeepSeek R1 used significantly more tokens to run Artificial Analysis’ intelligence index, which aggregates several varied benchmark scores together. Indeed, Claude has the lowest amount of total output tokens for leading reasoning models and showed an impressive improvement over Claude 3.7 Sonnet.

This aspect of tokenomics shows that there are many dimensions on which providers are working to improve models. It is not just more intelligence, but more intelligence per token produced.

This is very good for Anthropic, because as mentioned earlier in today’s Edition, they are supply-constrained on inference and their latency and tokens/seconds have been getting worse.

Semianalysis reports that Claude 4 Sonnet went from about 80 tokens/second in May to a little under 50 tokens/second in July. This slowdown would be *much worse* for users if Claude 4 was as verbose as, say, DeepSeek or Gemini 2.5 Pro. Conciseness is a feature, not just a stylistic choice.

The improvement over Claude Sonnet 3.7 is particularly striking (check the graph above).

I wonder what kind of techniques and architectural improvements they use to increase “intelligence per token.” And I also wonder if Anthropic had more inference capacity, would they tune the models to be equally as verbose as the competition, which would make them even smarter 🤔

🧠🔋🪫 Creatine Might Improve Cognitive Function for Alzheimer’s Patients (and it probably can help you too)

A pioneering trial study showed that giving high doses of creatine (20 grams per day) to patients in the early stages of Alzheimer's increased their brain's creatine levels by 11%, which improved their mental performance by 4.4% overall.

Improvements were highest in areas like working memory, problem-solving, focus, and reading.

From the paper:

Because the CABA study was a small, single-arm pilot trial of short duration, the preliminary secondary outcome results should be interpreted with caution, yet they provide additional evidence that bioenergetic intervention may be beneficial in the treatment of AD [...]

The cognitive improvements observed in this study are also promising, as AD is a progressive disease with expected decline over time. We hypothesized that memory and executive function, the most affected domains in AD,34-37 would benefit from CrM supplementation. In addition to these domains, participants also improved in oral reading recognition performance, a test that assesses the participant's ability to properly read and pronounce a visually-presented word. [...]

These studies suggest that CrM supplementation improves brain mitochondrial function and cognition

This was a small pilot study, so while the results were promising, it’s not enough to draw any firm conclusions yet. I hope that they will do a larger study for AD, but also keep investigating the brain energy metabolism benefits of creatine for healthy individuals too.

Personally, I had been taking 5g of creatine per day in my green tea for a few years, but a few months ago, I increased that to 10g/day.

💾 🎨🤖 Google’s Gemini Flash-Lite is So Fast it Can Simulate a Computer OS in Real-Time 😮

Ok, it’s kind of a useless demo for now, but it shows the potential of AI eventually doing real-time generation for games and all kinds of UI and software.

Here’s how DeepMind describes it:

Our prototype simulates an operating system where each screen is generated on the fly by a large language model. It uses Gemini 2.5 Flash-Lite, a model whose low latency is critical for creating a responsive interaction that feels instantaneous. Instead of navigating a static file system, the user interacts with an environment that the model builds and rebuilds with every click.

You can try it for yourself here.

This is a bare-bones, Pong-level implementation of this idea, but someday we’ll be using “software” that didn’t exist until we needed it. It’ll be created on the fly based on our needs and the context.

This will be particularly used with AR glasses because usage patterns and context will vary a lot more than with laptops or phones, making it harder to design a UX that fits everyone and every situation.

It’ll be the same for virtual worlds (for games and VR), where the cost of asset generation by artists will be replaced with massive real-time compute costs.

The game won’t just make stuff up randomly, of course. The art will be to create the scaffolding around which the generative AI will build: the storyline, characters, locations, and some visual and sound assets that the AI will use as a starting point to build on.

---

🎨 🎭 The Arts & History 👩‍🎨 🎥

👩‍🚀 Project Hail Mary, Film Edition ☀️🐜

I really enjoyed the audiobook of this sci-fi novel (Ray Porter is such a good narrator that I recommend the audio over the text). I wrote about it back in Edition #370.

I hope the film does it justice.

It’s directed by Phil Lord and Christopher Miller, the guys who made the excellent Spider-Verse films (and I don’t even care about Spider-Man 🕷️).

I trust their taste and story sense, so I expect it to be good 🤞

🎬 🎟️ Hans Zimmer Live 🎸🎶🎹🎺

I got tickets to see Hans Zimmer live with a couple of my friends!

The show is in September. I’ll report then, but based on the videos I’ve seen from previous tours, it seems very fun.

This isn't your typical stodgy orchestra playing film scores in tuxedos 🤵‍♀️🤵‍♂️🎻

It’s more like arena rock in spirit, a big celebration of the music 🎉

Hans banters between songs and plays multiple instruments: guitar, piano, bass, even banjo 🪕

A bunch of his studio collaborators join him on stage, and they look like they’re having a blast.

I can’t wait 😀