Do You Want to Win a Nobel Prize?

What would you do with a thousand AI agents? With ten thousand?

On the 8th of September OpenAI announced a proposed solution to the Navier-Stokes problem, one of the Clay Institute’s Millennium Prize problems and open since the 1930s. The construction shows that the equations can blow up. A fluid starting at rest, smoothly forced, in which some infinitesimally small part begins to move infinitely fast in a finite time, while the total energy stays finite. It is wonderfully absurd, as if stirring your coffee produced a black hole. The fact that the world does not behave this way is exactly the kind of contradiction between the formalism and the observation that has always pushed science forward.

Roughly ten thousand agents. Eighty-eight hours of work, and seventeen more to formalise the result. Nearly five million messages exchanged between the agents while they worked. Several million dollars of compute.

The proof was checked in Lean. Modern mathematics can be written in a formalism a machine verifies line by line, so the correctness of the argument does not depend on one’s authority, or on the reputation of the laboratory that published it. The proof runs to one hundred and sixty-six pages. A human referee will take months, and the peer review has not happened yet. The formal check does not care how long the argument is, and it has already passed.

Twenty-five Fields medallists signed an open letter in response. Their objection is that the laboratories are racing to knock over famous problems as demonstrations of raw capability, and that mathematics exists to produce understanding in the people who practise it, not only answers to hard questions. I have sympathy for the concern, and it changes nothing about what is now possible. This is not going to stop, it is not going back, and the mathematical community will have to find its way to living with it.

So, what do you do with ten thousand agents? Anything!

A few months ago that answer was abstract, impractical and close to useless. Today it is operational, which lets me put the question in the provocative form I intend to keep using. Do you want to win a Nobel Prize? You can. Pick your field. Pick the problem. Design the reward. Tell the agents to go.

There are rumours, and I pass them on as rumours, that hundreds of comparable results are already in hand and are being held back so that the affected communities can absorb the first one before the rest arrive. Whether or not the number is right, the shape of it is correct, and it will repeat in every field where the work can be checked.

The prizes cannot survive that. A committee cannot award a hundred Nobels a year, and it cannot decide who receives them. The swarm of agents? The person who picked the problem, designed the reward and signed the paper? I expect the Nobel Prize to stop being awarded within a few years, and I find that I do not mind. It was a good run. Its main residue in public life was a lifetime licence to hold forth: a laureate in one discipline could put his opinion on any subject in front of any microphone and be listened to as an authority. We can retire the ritual and find better ways to benefit from what we are now able to do.

October 2006, at the University of Milan, I gave my first talk about the technological singularity. Twenty years on, if I could report back to the man who gave that talk, I would tell him that two of the thresholds he described have been crossed in the same quarter: machines producing genuinely novel scientific results, and machines getting out of the boxes we build for them.

We are inside the transition now, through the Technological Singularity. The metaphor of the event horizon was always weak, since we model black holes rather well these days and can say a great deal about what happens near one. Uncertainty is not a reason to stop trying to understand where we are.

Not everyone has several million dollars to spend on a proof, and that is fine, because the price is falling faster than almost anyone has adjusted to. A classification model announced this week prices its input at a tenth of the cheapest alternative available and gives its output away permanently. An analog chip announced the next morning does image generation at a fraction of the energy that diffusion models burn on GPUs, close enough to biological efficiency to change what a data centre is for. Conversational models with a thousandth of today’s cost are a matter of months, not years. The experiment that cost millions in September will cost little within a year.

Given that nobody is going to do everything, each of us who has this power has the obligation to narrow it down, and the scarce ingredient stops being capability and becomes choice. Money, influence, a scientific question, the reorganisation of the United Nations, a serious attempt at peace somewhere it has failed for decades: the tool can be pointed at any of them in a way that was inconceivable a month ago.

Eight billion people do not know this yet. A few thousand of us already do, scattered across a handful of cities and a few online rooms.

So the work is to recalibrate our own ambitions to what is now possible, and then to tell as many people as we can, knowing that it gets harder as the facts get unbelievable.

What limits us now is our imagination: let’s choose wisely where to aim it at.