In one of my last posts, I tried to calculate an estimate of how many years of funding for a small research group was equivalent to what OpenAI spent to proof the Navier-Stokes problem (Answer: the estimated $20 million would correspond to around 125 years of funding).
Yesterday, a similar scientific research story came up, this time from Anthropic 's side.
Matthew von Hippel, a former theoretical physicist, challenged AI companies to solve an open problem from his old field with the kind of resources an academic actually has: computing particles scattering amplitude at nine loops. Without digging deep on what this technically means, imagine that it deals with a lot of complicated mathematical manipulation of tensors (multi-dimensional matrices) combined with functions with dozens (hundreds) of terms.
A month later, two physicists at Anthropic had the answer.
The instructions to Claude?
- Model statement: "The problem is to compute the Six-particle (hexagon) amplitude in planar N=4 SYM at nine loops."
- periodical instructions like "keep working on this".
The cost billed at API rates would be around $1–2k. The heavy computation itself was about $100.
It’s not a Millennium Prize problem. But it is a real frontier calculation, the kind that advances human knowldge.
Where it draws the line difference against the Navier-Stokes proof:
- the LLM did not invent a new method: it used the ones Lance Dixon and his collaborators developed over the years
- delivered the result in the format they had already set up
- results were validated by tools already built by Dixon’s team, and similar results were achieved by another group from Beijing University
Knowledge came from people and papers that were already published. The AI did the “heavy lifting” and closed the “small gap” that was missing.
But to me, this raises another question: can this mean a shift in how we approach scientific funding and PhD scholarships, similar to what we are seeing in the industry?
In industry, it is reported that fewer juniors are being hired, and senior engineers are becoming what is now called a “10x engineer” when helped by AI.
There is a parallel with Academia: professors bring the knowledge, experience and funding, and young (inexperienced) PhD students do a lot of the hard, repetitive (and sometimes novelty) work, taking key learnings for their research journey.
If a senior researcher can now do in one week, for the price of a laptop, what used to take a PhD student years… could this mean a shift in the focus of scientific funding?
How can we assure that funds are still routed for the training of next generations and not re-routed for the use of LLMs?
