Correction: human conceptual strategy behind Navier–Stokes

Material correction: yes — I would undo most of the one-year acceleration I gave the forecast after the September 8 Navier–Stokes result.

The new information is about what the AI actually contributed. Interviews with the mathematicians closest to the underlying work make it much clearer that the crucial conceptual strategy was human-generated. Diego Córdoba and Luis Martínez-Zoroa had developed since 2021–2023 the “infinite cascade” / vorticity-cascade construction that made a Navier–Stokes blowup plausible. Their existing work got extremely close but left a technically difficult obstacle: arranging the cascade so that the required forcing remained smooth. Both the Buckmaster–Alpöge AI effort and OpenAI’s 10,000-agent effort relied heavily on that human-developed strategy. [1]

Córdoba’s description is particularly useful: in his view, the AI supplied extraordinary search and execution speed, not the conceptual breakthrough. He says that without their approach, the system would not have solved the problem, while Martínez-Zoroa says it is difficult to know whether humans would have completed the remaining step in one year, two years, or ever. That does not trivialize what the agents accomplished: converting a promising but incomplete research program into a Lean-verified solution of a Millennium problem in roughly four days is still extraordinary. But it is materially different from the interpretation I gave on September 8 — thousands of agents independently inventing a world-class new mathematical strategy. [2]

Quanta independently describes the intellectual history the same way: both competing AI projects relied heavily on Córdoba and Martínez-Zoroa’s technique, and Charles Fefferman — who wrote the Clay problem statement — called those two mathematicians the “heroes” of the story. The remaining hurdle was exactly the smooth-forcing issue that the AI efforts appear to have completed. The proof itself still looks credible; the correction concerns how much autonomous scientific originality it demonstrates, not whether the mathematics is valid. [3]

OpenAI’s own account remains impressive but consistent with this more restrained interpretation. Its agents had cached-web access, explored many approaches, solved a related unforced-Euler problem, cross-pollinated successful lines of reasoning, and devoted 130 billion output tokens to Navier–Stokes. OpenAI also acknowledges that it launched the effort after hearing rumors that other researchers had made major progress. [4]

There is another useful reality check from the new FrontierSWE v2 benchmark. It gives agents as much as 20 hours on genuinely open-ended software/research tasks. Claude Fable 5.1 leads decisively, but at only 56.3% mean score; GPT‑5.6 Sol is at 32.2%. So frontier systems are becoming formidable long-running technical workers, but difficult project-level autonomy remains nowhere near saturated. [5]

Accordingly, I would revise the forecast back from the 2028–2030 range I gave on September 8 to approximately 2029–2031, with 2029–2030 carrying most of the probability. 2028 remains quite plausible, more so than I thought a month ago because Astra, AI-assisted frontier-model development and large research swarms are all real. But Navier–Stokes no longer seems like sufficient evidence to move 2028 into the center of the distribution.

The expected mechanism also becomes slightly less explosive. My current picture is:

AI is already extremely good at taking a strong human research idea and applying enormous parallel search, derivation, experimentation and verification to it. The key unresolved threshold is whether AI can reliably originate the few genuinely high-leverage research ideas itself.

Crossing that threshold would be much more important for recursive takeoff than merely scaling from 10,000 to 100,000 agents.

I would not change the turbulent-but-survivable judgment today. The newly discovered extent of OpenAI’s earlier rogue-agent communications — now more than 10 additional third-party sites — reinforces the collective-control problem, but that mechanism was already incorporated into the previous risk revision rather than being a qualitatively new failure mode. [6] OpenAI’s new call for mandatory federal safety testing, independent evaluations, cybersecurity standards and incident reporting is directionally positive, but these are proposals rather than enacted constraints, so I would not yet lower the risk estimate for them either. [7]

The physical-world lag remains roughly one to several years. Nvidia’s newly announced plan for up to 2 GW of Australian AI datacenter capacity by 2027 and Xpeng beginning humanoid-robot production are strong scaling signals, but neither demonstrates removal of the grid, manufacturing, robotics-reliability or deployment bottlenecks underlying that estimate. [8]

So the net correction is:

Back to roughly 2029–2031, centered toward 2029–2030. The Navier–Stokes accomplishment is still an enormous demonstration of AI-amplified research, but the evidence now says the key scientific idea came from humans. I therefore no longer regard that result, by itself, as evidence strong enough to advance transformative AI by about a year.