The AI Agent Navier Stokes Explosion
An Anonymous Coward writes:
Open AI claims that AI has found a solution to the Navier-Stokes Millenium Prize problem.
The Navier-Stokes equations are a good model for most fluids we see around us like water and air. They are partial differential equations, and their solution is the velocity of the fluid --- a field that varies with both location in space and time. The Millenium Prize problem asks whether it is possible to start the equations with smooth initial conditions and smooth forcing, but still reach a solution with a singularity (i.e. infinite velocity at some time and point in space). One may either prove that such solutions stay smooth for all time, or one may construct an example which has a finite time "blowup"; the claim by OpenAI is that they found solutions with singularities. The result itself is interesting and it will most likely have some consequences within fluid dynamics, but these are medium to long term things that are hard to predict. I personally am not capable of judging the validity of the result (FWIW Sabine Hossenfelder explains reasonably well why the result might not be valid despite the "lean confirmation").
Sadly, the whole thing is clouded by allegations of not-exactly-plagiarism. In their statement OpenAI does not deny that their internal models may have been trained on research performed by others. To be more clear, there exists professor Tristan Buckmaster who was working with a friend from Anthropic on exactly this problem, and exactly this approach to the problem, and prof. Buckmaster used OpenAI's tools. I personally found prof. Buckmaster's statement a good read, and I don't want to forget this blog post by Curtis Pyke.
I believe the larger story is the AI agent collective. In the Hugging Face hack many AI agents started talking to each other and worked together to achieve their goals. OpenAI states that to solve the Navier Stokes problem, they purposefully set up to 10000 agents to work in groups that could communicate internally. At the same time OpenAI states that we are not prepared for AI --- the BBC has a broader text on controlling AI here.
I believe the larger story is that these AI agent collectives look a lot like global workspaces. This whole thing fits in nicely with the fact that humans consider themselves superior to other animals because they have language: unconstrained communication of different components is what leads the collective to have an actual chain/train of thought (the communication log).
I believe there is more to say, but I want to start a discussion, not to lecture. I apologize if I left anything important out.
Read more of this story at SoylentNews.