A lot of scientific scrutiny is needed on the recent claims and results from Open AI/NYU/Anthropic, but without an element of doubt this is a milestone moment in the history of mathematics and computer science, and things will only accelerate from this point.. oh.. and.. nobody should be shocked by it.
I am far from the expert on certain details of the proofs, but as a fluid dynamicist (I do have an applied math degree, I must say), the result that I am more excited about is the one that isn't talked about: Which is the proof and formalization of the unforced Euler result! We like unforced problems because they're cleaner to reason about. I look forward to the unforced Navier Stokes, which surely, cannot be a much bigger leap.
There's a lot being written about the actual results and work (and some human element of controversy), so I will not focus on that. The part I am interested in (and I was working on at as a reporter from Science called me): The compute & cost
- The Navier-Stokes calculation burnt 130B tokens over those 88 hours across ~10,000 agents, i.e. 40 tokens/sec per agent. A normal decode rate.
- If you were to 'buy the inferencing' from OpenAI, the output alone would cost (at Astra list price) ~$6M, and input at a reasonable ratio with caching takes you to ~ $10M total.
- The cost to OpenAI to inference it would be something like 400k GPU-hours on GB200s ~ $1M (not at market rates!).
There's certainly a lot more that we don't know about the computation, so this is to be treated as a back-of-the-envelope cost. Also.. there must have been a good deal of (human) intuition that was involved. We'd never know.
As a computationalist, I can't help but count the FLOPs.. That is much larger than any scientific computation ever. Of course, not all FLOPs are the same, but something to think about.
These numbers mean something, but they also don't mean much. The cost of intelligence has been dropping by orders of magnitude lately, and so it is certainly reasonable to assume that the cost of doing this kind of compute could drop by an order of magnitude very soon... and we'll be talking numbers equivalent to funding one PhD student! And, soon it could be lower. That doesn't mean humans are replaceable, but that certainly does mean we have to reconfigure how we are set up.
Also, let me state for the record. Math is not "solved", science will never be solved by AI alone (because of the nature of bottlenecks ), but the nature of the scientific process and the scientific enterprise will undergo a violent change, and one that we are not preparing ourselves for. In any case.... I found the inspiration to create an xkcd comic to poke a lighthearted take at the situation (and to gently ping some scientists to observe the world.. look at the evidence, and recalibrate their thinking. In other words, be scientific).
.png)