Navier-Stokes result prompts debate on Astra and recursive self-improvement
The author asks whether the Millennium Prize solution came from brute-force compute or a creative "Move 37"-style insight, and speculates on what follows once Astra is wired to robot arms.
Some takes on the Navier-Stokes result: - Gun to my head, if you asked me a month ago when LLMs would be able to solve a Millennium Prize Problem, I would have probably said "mid-2027". - My key question about this result is whether it's more a "brute force" result (tons of compute wielded by an extremely powerful model solves problem) or a "Move 37" kind of result (genuinely creative and novel reasoning). I'm not a mathematician, so can't really tell! I think the answer to this question should be very informative vis-a-vis timelines for RSI. - Last year, OpenAI used a complex multi-agent setup to win gold on the IMO. This year, Meta Muse Spark won gold on the IMO. This should give you some food for thought regarding the kind of progress we should see in the next 12 months, given today's Navier-Stokes result. - All the videos of Astra connected to robot arms on your timeline are a precursor for what's next. Things will get *really* insane when intelligence at least as powerful as this internal OpenAI model is connected to the real world and is able to work with physical objects and run experiments. Once intelligence like this exists and is a bit cheaper to run (see previous point), there will be *a lot* of capital chasing this problem.