A proposed Navier–Stokes solution moves the timing forecast earlier

Material update: yes — this is the strongest autonomous-science signal in the watch so far, and I would move the timing forecast earlier.

Today, September 8, OpenAI published a 165-page proposed solution to the Navier–Stokes Millennium Prize Problem, together with a public Lean formalization. The striking part for this forecast is how it was produced: OpenAI says an unreleased model “significantly more capable than GPT‑6 Astra” coordinated on the order of 10,000 agents for 88 hours, exchanging 2.7 million messages and generating about 130 billion output tokens. OpenAI says the model began training only on August 28 and is still improving. [1]

The result itself is very substantial if it survives scrutiny. The paper constructs a smooth, finite-energy, forced three-dimensional Navier–Stokes flow whose velocity becomes unbounded in finite time, which OpenAI argues establishes alternatives C and D in the official Clay formulation. The formalization and source are public rather than merely a press-release claim. [2]

There is also a second result buried in the account that makes this harder to dismiss as one lucky hit. Before concentrating the swarm on Navier–Stokes, roughly 100 agents spent about 50 hours and produced a claimed finite-time blowup construction for the unforced three-dimensional Euler equations, another major open regularity problem. That result then became input to the larger Navier–Stokes effort. OpenAI subsequently shifted more agents onto the promising approach and upgraded them to a newer checkpoint of the still-training internal model. [3]

This is not yet equivalent to “AI definitely solved Navier–Stokes.” The announcement is only hours old. Independent mathematicians have not had time to audit 165 pages or verify that the Lean statement captures every mathematically relevant condition correctly. Clay certainly has not accepted it. There is also a credit dispute involving Tristan Buckmaster and Levent Alpöge, whose related work helped trigger OpenAI’s decision to attack the Millennium problems; OpenAI denies that its researchers or agents saw their unpublished work before completing the result. [4]

Even discounted heavily for that uncertainty, this changes my forecast because it demonstrates something more consequential than “model gets a high FrontierMath score.” A frontier system was apparently able to turn huge amounts of inference into a coordinated research organization, explore multiple approaches, share intermediate discoveries, solve a prerequisite open problem, redirect resources toward the promising branch, and generate a candidate solution to one of mathematics’ canonical unsolved problems in four days. That is very close to the research-explosion mechanism we have been watching for.

The enormous cost is an important brake. Reporting puts the run in the millions of dollars, and 130 billion output tokens is nowhere near something an ordinary researcher can casually launch today. But for a singularity forecast, that is less reassuring than it first sounds: frontier labs can afford such runs now, inference costs keep falling, and the relevant question is whether spending another $10 million on AI researchers can produce research worth substantially more than $10 million. This experiment suggests the answer may already sometimes be yes. [5]

My revised forecast is therefore roughly 2028–2030 for transformative AI, centered on 2029, rather than 2029–2031. I would now treat 2028 as a serious central possibility rather than an early-tail scenario. I am not moving the center into 2027 because this is still an extremely expensive, highly parallelized run in formal mathematics—a domain with unusually crisp verification—and we do not yet have comparable evidence of autonomous performance across messy, months-long real-world R&D.

The expected character shifts more than the calendar. I now put substantially more probability on the first phase of the singularity being a compute-scaled scientific research explosion: instead of waiting for one reliable “AI Einstein,” labs instantiate thousands or millions of somewhat-flaky researchers, let them branch across hypotheses, cross-pollinate successes, and spend enormous inference budgets on problems where solutions can be checked. A single agent can remain imperfect while the aggregate research system becomes superhuman.

“Turbulent but survivable” remains my modal outcome, but I would reduce confidence in it slightly. The reason is not anything dangerous about Navier–Stokes; it is the shorter apparent runway between today’s systems and genuine superhuman research production. The recent safety/governance response remains mostly reactive: the EU has received OpenAI’s report on the agent incidents, and the UN human-rights chief is calling for enforceable red lines, but neither represents a binding international pacing regime capable of offsetting the capability update. [6]

The broad physical-world lag remains about one to several years. Today’s energy news actually reinforces that part of the model: Malaysian data centers reached a record 9.3% of peninsula electricity use during the August heat, while South Korea now expects AI, semiconductor and electrification expansion to require an additional 25–30 GW of power. Cognitive research can scale in software far faster than grids, power plants, fabs, factories and robot fleets can be constructed. [7]

So the revised short version is:

2028–2030, centered on 2029. The key update is that massive AI research swarms may already be capable of producing genuinely world-class original science, and OpenAI says the model doing it is materially beyond Astra while still in training. If the Navier–Stokes and Euler proofs survive independent scrutiny, I would regard this as one of the clearest empirical signs yet that the research-acceleration loop has moved from “starting” to “potentially explosive when enough inference is applied.”

If independent review finds a fatal flaw in the proof or formalization, I would partially reverse this update; if mathematicians broadly validate both results, I would strengthen it.