Proof-of-work Labs, and the Navier-Stokes Moment

20 Oct 2026

The swarm

In September 2026, OpenAI announced that a group of roughly 10,000 AI agents, running on a superior model they haven't released yet, produced a proof for the Navier-Stokes existence and smoothness problem (one of the Millennium Prize Problems). What matters the most to me is how the swarm got there.

None of those agents is individually more intelligent than the smartest mathematician alive right now, someone like Terence Tao. At least in most definitions of intelligence that we care about. But together they worked for about 88 hours and exchanged 2.7 million messages, each one passing on what an agent had tried and learned. Now imagine Terence having to swap 2.7 million messages with his smartest mates, with both sides gaining something from every conversation. No group of humans can do that of course, and the swarm did it in under four days.

I've been thinking about this for a bit. It's obviously super impressive, but what I find most interesting is the shift I've been seeing around it, both in how people think about AI and in how they think about building companies (this is the interesting bit to me).

The believers are not too surprised

If you've always been a hardcore believer in AI (or SI I should say), none of this surprised you too much, obviously it's super interesting to see it in real life though. But if you've always been of the opinion that there's nothing about human intelligence that can't be built out of physical stuff, then a swarm of AI agents solving a Millennium problem sounds like only the start of the exponential.

The argument that the hardcore (physicalist) believers follow goes something like this.

We are material beings, and our brains are matter following the laws of physics. People say there is something more to it, and maybe there is, but maybe there isn't.

If there isn't, then the act of "thinking" is something that matter can do just by itself. There are two options that are useful then to build an artificial intelligence as intelligent as humans:

• Option 1. Same pattern, any material. Option 1, The computation matters. What makes you intelligent is the pattern of information processing: inputs come in, get transformed via a specific mechanism, decisions come out. Neurons just happen to be what runs it, the same way a calculation gives the same answer on a Mac or a PC. If that's true, silicon can do it too. Everyone from the big labs is happy if this is true. They don't even need to change what their AI models run on, GPUs in their current form are probably just fine, you just need a lot of them. Option 1. Same pattern, any material.

• Option 2. Shuffle a cell and a chip and you can't tell them apart by what they do. Or copy the material and scale it. Option 2, The material matters. Something about the biology itself, like the chemistry or particular physical properties of neurons, is essential. This complicates things a bit more. But it's a question a bit more related to "human" intelligence and consciousness, which you might or might not care about. Let's say this Option 2 is true, a perfect simulation on silicon could act intelligent (which might be everything we care about) without being intelligent in the same way, much as a chess engine plays brilliant chess without thinking about the game the way a grandmaster does. Also, if Option 2 is true, then fuck it, the big labs might just switch from silicon to the same carbon (and etc) materials that humans are made from, and scale it up. Think of a mega brain, just made of the same stuff as ours. Sure, Nvidia would have to pivot, and we might de-accelerate for a bit until we find how to extract these materials at scale and put them together coherently (time to build a biology/chemistry/physics company), but it's still not stopping us from building an intelligence like ours, or something that looks so much like ours that we don't care. Option 2. Shuffle a cell and a chip and you can't tell them apart by what they do. Or copy the material and scale it.

Once you accept the logic above ^, you get human-level AI in principle. And from there, the step to smarter than humans is surprisingly small. Machines can copy themselves, so one smart mind becomes ten thousand. They can collaborate at a scale no team ever could. They never sleep (unless you run out of power or natural resources to keep them alive and running). And a single instance can be given more compute, whereas you can't give a person a bigger brain. The missing piece has always been learning, or the ability to improve themselves. That's what every single lab is trying to achieve, the key to truly unlocking the exponential.

Figure 1. The whole argument in five steps. A human, the same shape built from metal, the two being identical if you only look at behaviour, ten thousand copies, and then each generation learning from more data, from the world and from its own best version. The stream widths are illustrative, not measured.

Self-improvement is still the hardest part of this chain in my opinion, even more than the physical changes to where the compute runs that would need to happen if Option 2 above is true. True self-improvement (and there are various definitions of what this means, can be its own post) is a hard thing to achieve, and I don't think the Navier-Stokes result says much about it either way. These are 2026 models, very good and you can run them in swarms, but probably (I don't know, I don't work at OpenAI) not improving themselves yet in any strong definition. Which is exactly why the result is interesting. You don't need the superintelligence argument at all to take the next part seriously, it was about the compute all along, and this is just getting started.

It was about the compute all along

Here's the detail I keep coming back to. Before going for Navier-Stokes, OpenAI pointed about 100 agents at a related problem, the Euler equations (Navier-Stokes without viscosity), and they solved it in roughly 50 hours. Then they moved their agents over and scaled up to around 10,000. Their model didn't change, it wasn't smarter. They took the same approach and put a lot more inference behind it.

You can think of a hard open problem as a search. Each agent following a line of attack has some tiny chance p of finding a way through. With N agents working in parallel, the chance that at least one of them gets there is P = 1 - (1 - p)^N, which for tiny p is roughly N · p. In other words, doubling your agents roughly doubles your odds. Real agents share what they find, so they do better than independent attempts, but the shape holds.

To be fair, it's hard to fully separate "more agents" from "better model" here. OpenAI also switched to a further-trained model halfway through the run. And maths is a bit of a special case, because Lean can verify a proof in hours. If you're working on a new battery chemistry or a drug, you still need a physical experiment to tell you whether you were right. This will be solved a bit later, with self-contained physical chemistry labs, but either way, the direction is pretty clear. Once the model is good enough, a breakthrough starts looking like a function of how much inference you can point at the problem. Full stop.

Building a company in 2026

Even if you're not a hardcore believer in AI/SI, this should change how you think about starting something. The old question was "who can you hire?". Talent was the scarce thing, so companies competed on people. It was all about the aura of the people in your startup. And it arguably still is (because you still need smart people for now), but much less. In a growing number of fields, the scarce thing is becoming compute. How much inference can you give each of your employees? And more importantly, how much can you aim directly at the one hard problem your company exists to solve?

Pick a field. Find the breakthrough that would change it, whether that's a material, a drug target or a proof. If models are good enough to find it in the limit of compute thrown at it, the company that wins is the one that throws enough agents at it first. The breakthrough is the company.

It reminds me a bit of the early days of modern science, when so much was still waiting to be discovered that one well-educated person with a bit of patience, strong first principles thinking, and determination/curiosity, could change an entire field. I think we might be heading back there. The difference is that the limit is no longer how many brilliant people are alive. It's how many GPUs you can get.

Neo Labs Proof-of-work labs

I think neo labs, as we know them today, are going to die. They raise billions to train their own models, and those models will never keep up with the frontier.

What I think comes next is something I'm calling proof-of-work labs. The name comes from the blockchain space, where you burn compute until you find the answer and then collect the reward. A proof-of-work lab does the same thing with intelligence. It doesn't train anything. It rents the best model available, picks a very niche problem whose solution is worth a lot of money (a material, a drug target, a proof), and points as many agents at it as it can afford until it breaks. It's the true conceptualisation of the idea that the only important thing is what you point self-improving AI at. The model is a commodity (sort of, of course), the compute is the cost, and the breakthrough is the product.

You could argue a neo lab is just a proof-of-work lab whose niche problem happens to be the model itself. Instead of a drug or a battery, they bet on a new architecture, usually out of some hypothesis or crazy idea about how intelligence should work, probably based on a re-interpretation of the Bitter Lesson, a new simple algorithm that can scale in a different way that big labs haven't found out yet. The problem is that this is exactly the one problem the big labs will never leave alone. Making their AI better is their whole business, and they'll be pointing their own (very powerful, not available yet to the public) AI at it, with more compute than anyone else. My bet is that the big labs end up with plenty to do between consumer products, API access and, mainly, improving their models. Everyone else becomes a proof-of-work lab (or a company meant for human pleasure/fun/free-time, which I'll explain in a separate post).

Again, the obvious question is why a proof-of-work lab would ever beat the big labs, who own both the best models and the most compute (OpenAI solved this one themselves, after all). I think the answer is focus. A big lab can't point 10,000 agents at every niche problem in every field, and it won't bother with most of them. A proof-of-work lab picks one, brings the domain data, the experiments and the verifiers that field needs, and rents the intelligence.

Navier-Stokes was the first proof of work, quite literally. The next ones will come from whoever is willing to burn the most inference on the right question. Asking the right questions is all that matters (and compute, of course).