Bookmarks

AI discovered how a material just one atom thick can keep carrying load as its atomic structure begins to break, preventing catastrophic failure. This discovery is based on first-principles atomic scale reasoning integrated with biological principles, cutting across scales and providing deep insights into materials in extreme conditions. The resulting material is extremely lightweight yet strong, far better performing than existing structures. By organizing "simple" carbon atoms into hierarchical graphene architectures, unique materials can be designed that redistribute forces, accommodate deformation, and confine damage, the AI identified design principles for resisting catastrophic failure. The AI built the atomistic simulation instrument itself "from scratch"; and then conducted experiments autonomously that revealed when alignment strengthens a material, when hierarchy protects it, and when an apparently promising design fails. This is a frontier in designing matter at its ultimate thinness - controlling how mechanical failure unfolds through the organization of individual atoms.

Here is what we did:

▶️ We asked an AI to build and use a scientific instrument. It wrote the force engine, structure generators, loading procedures, analysis tools, and experiment database. The independent campaign ran for multiple days without scientific intervention.
▶️ We required the instrument to pass physical and numerical tests. The implementation passed twenty validation tests and reproduced reference energies to approximately 10⁻¹³ eV per atom in the tested configurations. Every proposed design then faced the same reactive interatomic model.
▶️ We required predictions before results, creating a loop of world model building and falsification/verification. The AI had to commit to what unseen designs would do, then run the simulations. Incorrect predictions became opportunities to identify missing mechanisms.

What emerged is a set of physical design principles:

▶️ The atom-scale arrangement of matter controls strength. The AI first showed how and why strength varied by more than sixfold across architectures. Similar amounts of carbon produced very different resistance to failure because they organized the load-bearing connections differently.
▶️ Rotating a pattern can change the mechanism of failure. Angled slit arrays revealed three regimes: neighboring slit tips link, intervening ligaments rotate, or short bridges bend. A rule based only on the remaining cross-section misses these changes in connectivity and motion.
▶️ Hierarchical structuring works under identifiable conditions. At the original scale, much of its apparent strength advantage can be explained by alignment. With greater separation between structural levels, selected hierarchical designs became about 25% stronger than same-mass single-level controls and showed larger integrated stress-strain responses. Veins redistribute load, compartments localize damage, and the architecture changes how cracks propagate.
▶️ A failed prediction is crucial to reveal the next experiment. Some proposed rules survived targeted tests; others failed. Longer loading preserved the broad architectural strength contrasts while revealing additional deformation and, in some cases, later stress peaks. The scientific value lies in identifying both the rule and its boundary as the AI punctures known scientific knowledge.
▶️ The instrument opens an extremely complex design space. The AI was able to expand the design languages into an open atlas of hundreds of thousands of atomically explicit structures.

The deeper implication is that AI can construct an executable connection between equations, experiments, and explanations. Physical reasoning is something they can implement, interrogate, and revise.

A scientific instrument extends what a scientist can observe; and an AI that builds such an instrument extends the experiments it can perform, and the questions it can ask, starting from basic principles of how atoms interact based on quantum mechanical ground truth.

Models building models, with physical evidence shaping recursive reasoning loops.

AMD acquires World Labs for ~$8.2b. But who gets the 💰? My usual breakdown below 👇

From founding to an $8.2b exit in ~2.5 years. A huge value creation event. This is a fantastic exit, especially for the co-founders and the team.

Investors will still share ~$2.3b of profits on $1.2b invested, a ~2.9x blended. Low-ish because most of the capital came in last. But it hides a lot of disparity between the various rounds!

So let's dive in:

1) The real home run: founders and team 🏆🥳

This is one of the best founder outcomes I've modelled.

@drfeifei, Justin Johnson, Christoph Lassner and Ben Mildenhall started World Labs in Apr-24. Less than 2.5 years later, they and the team still own ~56% according to my estimates. At an $8.2b valuation, that's a total prize of ~$4.6b to share between them.

Taking a 15% attributed option pool, that's ~$1.2b for the team and ~$3.4b for the four co-founders. If split equally, that's ~$850m each, in just two and a half years. 💸

And, at close, Fei-Fei Li joins AMD as EVP and Chief Scientist, reporting to Lisa Su! Legendary.

2) The buyer was already on the cap table (again) 🔁

AMD Ventures invested in the Series B and the Series C here. Now they are buying the whole company. It's the same playbook as Nvidia with Hugging Face: strategic corporates are using their venture arms as a free option on future acquisitions.

3) The standout VC winners: a16z and Radical Ventures 👑

Both backed Fei-Fei at inception, in Apr-24 at a ~$300m valuation. That Series A will return ~19x the money, or ~$1.2b of proceeds, in under 2.5 years.

Congrats to @martin_casado and team. Another mega win after Cursor and OpenRouter. The team at @a16z truly is on a generational run.

4) The star angels did well 🌟

The Sep-24 Series B had arguably the strongest AI angel list ever assembled: Geoffrey Hinton, Andrej Karpathy, Jeff Dean, Eric Schmidt, Reid Hoffman, Marc Benioff, Ram Shriram and more.

Priced at $1.3b, it was never going to be a 1,000x. But ~5x in two years is a ~125% IRR 📈

And the best part is: Hinton's AlexNet won the 2012 ImageNet competition, and ImageNet was the dataset Fei-Fei built. Twelve years later, he backed her company. Full circle. 🔄

5) The late money: a great IRR, but a poor multiple of money 🤏

The $1b Series C in Feb-26 at $5.4b returns ~1.5x in 7 months. That's a great IRR, but in terms of venture returns, this is a poor multiple of money. Although I am sure no one will complain about seeing DPI this quickly!

Congrats to all involved!

someone's going viral after copying my exact style.
honestly that's just proof the design works.

I bet he saw this post and asked claude to replicate it. it's flattering :)

I'm going to share exactly how to do it anyway,
so anyone can build 3d interactive STEM lessons.

pls try it with different subjects.
make it useful for anyone who wants to learn.
one thing I noticed: his page took a while to load
because it was stuffed with too much.

my goal is simple:
it needs to load really fast.
it should be lightweight.
it should be useful for learning.
it should be visually stunning.

AI is cool, but these 3d interactive lessons can still produce inaccurate results right now,
so they really need input from experts.

an open-source repo is coming!
I'm hoping people will add new learning material or fix inaccurate visuals just by prompting.

it needs to be done properly.
let me know what you want to see??
spoiler: it's never about crafting the perfect prompt or /skills.
you won't need any of that.
where we're going, you just ask for stuff.

had to take a 24h break.

came back to X and found out my post got millions of views.

got emails I wasn't expecting,
including a really warm one from the @OpenRouter CEO (Hi @alexatallah )

got replies, likes, and even follows from people I really look up to, like @bcherny and @addyosmani .
(if you read this please please don't nerf opus 5.5 I beg you)

some people are asking what I did to the algo.

honestly? no idea.

I'm at 4,000+ followers now and still going up.
was around 500 a few days ago.
only had 200+ last month.

I'm a nobody.
I just post stuff I think is fun,
stuff I like building.

I’ll keep doing that,
keep posting original stuff.

thank you, genuinely. it means a lot.

stay tuned. peace.

... here’s what’s cooking right now in 35 seconds

AI moves so fast that I thought this book was going to be a year too late. It turns out it's right on time.

While I was writing, I kept having to change chapters from "someone should do this" to "someone has done this," and then go talk to the people who had done it. A good example is the chapter on virtual cells. I'd written that we needed large initiatives to build one. By the time I finished, I was writing about the ones that exist — the Chan Zuckerberg Initiative's is now the largest and best-funded push to model a human cell.

The Eureka Machine is out today.

It's about building a full-stack scientific superintelligence: a living map of human knowledge, a model of physical reality, high-fidelity simulations, autonomous labs, and an agent swarm of AI scientists on top. There's no better science to start with than the science of AI itself, because that's where the feedback loop is tightest. Then you expand the aperture to physics, chemistry, and especially pre-clinical biology.

This will be the most meaningful application of AI, and one of the most important things humanity does. And it may be the last organic invention we need to make. After that, it starts inventing everything else for us to solve our most pressing problems.

While writing, I couldn't let the idea go. Eight of us started @Recursive_SI to go build it.

AI is going to massively accelerate science, and unlike most of what gets argued about right now, that isn't zero sum.

I hope you all enjoy the book. I look forward to discussing it with you all.