What Did We All Miss?

When OpenAI announced a solution to the Navier-Stokes Millennium Prize problem, a lot of the conversation I saw was about spending roughly twenty-two million dollars of compute on a one-million-dollar prize. Fair enough, it's a funny comparison. I was more interested in how the work got done.

Around ten thousand coordinating agents reportedly produced a proof and a Lean formalisation. OpenAI declined the prize money. A machine-checked proof at that level would be a substantial research result, even before anyone found a practical use for it.

The claimed result is a finite-time singularity: smooth solutions can break down. That doesn't give us a new method for building rockets or modelling turbulence. The part I want to understand is how the agents carried out the mathematics, and how much of that process transfers to other problems.

On the 6th of September, Jensen Huang declared that AGI has arrived and credited GPT-6 Astra, trained on more than 100,000 Nvidia Grace Blackwell GPUs. Two days later, the proof was released. Back in March he had told Lex Fridman “I think we've achieved AGI” while discussing whether AI could build and run a billion-dollar company. On the August earnings call, the claim was narrower: for many tasks, we could say we had achieved AGI.

Sam Altman's statements have been hard to reconcile too: that current systems meet OpenAI's definition, but that he still expects an internal system he would personally call AGI later this year.

I find this exhausting. Someone reads one of those headlines, asks the app to do their job, and watches it fail at something a competent colleague would handle. A definition chosen for an interview doesn't help them understand what the system can actually do.

And yes, Huang attributing AGI to a hundred thousand of his own GPUs is pretty funny.

AI is in product announcements, ads, LinkedIn posts, news segments, and tools I already used that have since acquired a sparkle icon. It's fucking everywhere. I still want to read about it, but friends outside tech often want the conversation to end before it starts.

Some of this resembles the tendency to form opposing camps: everything is a bubble, everything is slop, or we're about to make a god. But with friends who study business or work in other fields, I mostly notice fatigue. They've heard too many claims and don't want another thing to evaluate.

That's an observation about people I know. I don't share the feeling very often; usually I stop reading because I'm physically tired. So I don't think I get to lecture them about paying more attention.

Thiel's “0 → 1” framing is roughly how I think about this: some companies make something that wasn't possible before, while others sell access to something already built.

Reselling something useful is fine. What annoys me is a subscription wrapped around a prompt being marketed as a technological breakthrough. Enough of those announcements make it harder to take the next one seriously.

I'm more concerned about products that work well and do harm: surveillance tools, for instance, or military applications whose civilian use gets all the attention in the brochure. My objection there is to what people are building and who gets to use it.

Robotics is one area I wish I heard more about.

AMI Labs raised $1.03 billion in a seed round, with Yann LeCun as chair, to work on world models using the JEPA approach. Generalist AI raised $400 million in June and was reportedly discussing a $3 billion valuation in August. Its GEN-1.5 model is shown learning a physical task from a single demonstration, without gradient updates or fine-tuning.

Those companies clearly have access to capital. Yet I hear much less about their work than about the next language-model release. That may partly reflect what I follow. Robotics is harder for me to evaluate, and keeping up with even one part of this field takes time.

If robots become much easier to teach, I want to know how we're going to handle the changes to physical work. Should some jobs remain reserved for humans, including therapy, medicine, or particular trades? I haven't settled on an answer. The question involves income, responsibility, and what people want from the person doing the work, as well as whether a machine can perform it.

Imperva's 2026 report puts automated traffic above 53% of the web traffic it measures, with bad bots accounting for 40% of the total and AI-enabled bot attacks up 12.5 times year over year. Those are traffic measurements. They don't tell us what proportion of posts or conversations are generated.

There is separate evidence about how convincing generated conversation can be. In Jones and Bergen's Turing-test study, participants spoke to a human and a model for five minutes, then picked which was human. GPT-4.5, prompted with a humanlike persona, was selected 73% of the time. LLaMA-3.1 reached 56%; the ELIZA and GPT-4o baselines reached 23% and 21%.

I sometimes think I recognise generated writing. I wouldn't want to depend on that judgement, especially when I'm tired or reading something outside my field.

I still enjoy trying things without knowing whether they'll go anywhere. Make some ridiculous software to annoy your uni professor, or follow a dumb idea far enough to see what happens. It doesn't need to become a company.

I've noticed capable people in tech avoiding an attempt because they can't see a useful outcome in advance. I wonder whether the pressure to take a position on every technology contributes to that. Short-form media might play a part too, but I'm guessing.

I want room to be interested in the technology while objecting to how it's sold and used. I can dislike a surveillance product, or an AGI claim that tells me almost nothing, and still want to understand a new model or experiment.

The research interests me enough to keep reading. The marketing makes that harder, and friends who don't work in tech have much less reason to put up with it.

I'd like more of the conversation to involve trying specific things and less of it to be about whether all of AI is good or bad. Robotics is one of the subjects I'd like it to include.

~ A.