Eddy Keming Chen Puts Superintelligence Within a Year

Eddy Keming Chen, a philosopher at UC San Diego, said on Sept. 28 that superintelligence may arrive within a year ¹. His main exhibit is OpenAI's 166-page proof about fluid flow, which Lean software has checked and mathematicians are still untangling ⁸.

Chen, an associate professor with a joint appointment in philosophy and the Halıcıoğlu Data Science Institute, answered UC San Diego Today's one question ¹. "When we wrote our Nature Comment in Feb. 2026, I thought superintelligence might still be five years away," Chen said. "Recent developments have changed my assessment. I now believe we may reach this reality sooner than anticipated, perhaps even within a year" ¹. His example is OpenAI's claim that an internal system of AI agents produced a proposed solution to the Navier-Stokes Millennium Prize Problem, with a formalization in Lean, software that checks proofs step by step ¹⁵.

Three grades of evidence sit in that answer: a proof under review, its machine check and a revised forecast. Proof and check earn a small step up in capability, 0.4 points at reported confidence. His forecast stays off the scale, since the reading counts evidence and a forecast is a judgment about it. Chen wants the aims chosen now: "Faster scientific discovery and new medical treatments could bring enormous benefits for humanity. But there is disagreement about which goals to pursue and for whose benefit. Achieving superintelligence alone does not resolve those disagreements" ¹.

OpenAI's Lean-checked Navier-Stokes proof lifts the reading 0.4 points

The move

Capability steps up 0.4 at reported confidence: an internal OpenAI agent system produced a Lean-checked proof that Navier-Stokes solutions can break down. An internal model, a disputed path and a single problem keep the step in the lower third of its band.

The data5 rows · sources

Measure	Value
Reading before this day	1.9
This piece's move	+0.4 (Capability, reported)
Band for reported evidence	0.3 to 0.8
Reading after the day (with 4 other pieces that day)	1.7
Distance to 100	98.3

What did OpenAI's agents produce?

OpenAI's 166-page paper, posted Sept. 8, describes a fluid that "starts from rest and develops unbounded velocity in finite time while maintaining uniformly bounded kinetic energy" ⁵. Put plainly: a still fluid, stirred by a smooth force, spins up until its speed at one point runs to infinity. For air, George Karniadakis of Brown University told Nature that "the singularity appears when [the] vortex becomes around 70 nanometres wide," the Tufts Daily reported ¹⁴. At that width air is a crowd of molecules, and the equations have stopped describing it ¹⁴.

The run took under four days. OpenAI's announcement, quoted by Simon Willison, set the clock: "The agents arrived at their resolution on Saturday, September 5, about 88 hours after the first agents were launched" ⁶. About 10,000 agents sent 2.7 million messages and used about 130 billion output tokens, and GPT-6 Astra, a public model, spent 17 hours on the Lean version ⁶⁷. NPR put the compute at roughly $6 million to $10 million at OpenAI's public prices ⁸.

Human mathematicians built the method. Quanta Magazine's Konstantin Kakaes reported that the OpenAI team and two mathematicians, Tristan Buckmaster of New York University and Levent Alpöge of Anthropic, "relied heavily on work by Diego Córdoba of the Institute for Mathematical Sciences in Madrid and Luis Martínez-Zoroa of CUNEF University" ⁷. OpenAI started on Sept. 1 after a rumor it later tied to the pair, who had used Claude and OpenAI's Codex on the problem for most of a year ⁶.

Chen's forecast fell from five years in February to one year in September

Timeline

Chen's AGI claim and Marcus's reply came in February. OpenAI's internal model began training Aug. 28, its agents reached the result in about 88 hours, and Chen's revised forecast appeared three weeks after the paper.Sources [1] [2] [4] [5] [6] [9] [12] [15]

The data9 rows · sources

Date	Event	Source
Feb. 2, 2026	Chen, Belkin, Bergen and Danks argue in Nature that AGI is here	[2]
Feb. 17, 2026	Marcus, Quattrociocchi and Capraro reply in Nature	[4]
Aug. 28, 2026	OpenAI's internal model begins training	[15]
Sept. 1, 2026	OpenAI launches its Millennium Prize effort after a rumor	[6]
Sept. 5, 2026	Agents reach the Navier-Stokes result after about 88 hours	[6]
Sept. 8, 2026	OpenAI posts the 166-page paper and its Lean files	[5]
Sept. 11, 2026	Clay Institute calls the problem apparently settled, review deliberately unhurried	[9]
Sept. 11, 2026	Twenty-five Fields medallists call AI and mathematics goals severely misaligned	[12]
Sept. 28, 2026	UC San Diego Today publishes Chen's revised forecast: perhaps within a year	[1]

What did Lean check, and what must mathematicians still check?

A Lean check proves one thing: that the argument establishes the statement written in Lean. OpenAI's repository says its files establish Clay's alternatives (C) and (D), breakdown in all of space and on a periodic box, and credits Google DeepMind's Formal Conjectures project for the Lean statement of the problem it adapted ¹¹. Lean checked the proof against a question that began as Google DeepMind's text.

Mathematicians still have to confirm that the Lean question matches the problem Clay posed in 2000. Kakaes spelled it out: "The crucial bit of verification that must still be done by humans is to guarantee that the statement being shown to be true in Lean is logically equivalent to what mathematicians set out to prove" ⁷. Charles Fefferman's official statement lets alternative (C) use a smooth external force, and OpenAI's result uses one ⁵¹⁴. Tapio Schneider, in a guest post on Terence Tao's blog, marked the rest: "The unforced version of the problem, whether smooth solutions exist for all time without external forcing, remains open" ¹³.

Javier Gómez-Serrano of Brown University told NPR the Lean code compiled and "the community seems to have the consensus that it is correct" ⁸. On Sept. 11 the Clay Mathematics Institute wrote that "the Navier-Stokes problem has apparently been settled," adding that its process "is deliberately unhurried" ⁹. Its rules require publication in a qualifying outlet, two years of waiting and general acceptance before Clay will consider a solution ¹⁰. Formalizing and checking the proof in Lean took 17 hours. A checker certifies a proof in hours; a field certifies a problem in years.

GPT-6 Astra formalized the proof in 17 hours; Clay's rules ask for two years

Compared

Hours on each clock. The agents took about 88 hours and GPT-6 Astra 17 more to formalize the proof in Lean. Clay's rules then ask for two years after publication, about 17,520 hours, before Clay will consider a solution.Sources [6] [10]

The data3 rows · sources

Unit: hours
Item	Value	Source
Agents reach the result	88	[6]
Lean formalization and check by GPT-6 Astra	17	[6]
Clay's minimum wait after publication two years	17520	[10]

Is AGI already here?

UC San Diego Today framed superintelligence as a onetime science-fiction premise that researchers now debate in earnest, with little agreement on whether or when it arrives ¹. Chen's forecast rests on an earlier claim. "Superintelligence is greater than general intelligence," he said, and general intelligence, he holds, has arrived ¹. He argued it in a Nature Comment published Feb. 2 with three UC San Diego colleagues, Mikhail Belkin, Leon Bergen and David Danks, whose fields run from data and computer science to linguistics, cognitive science and policy ²³. Danks now teaches at the University of Virginia ¹¹⁹. "We are at a Copernican moment where we are questioning the very notion of whether general intelligence is uniquely human," Chen said ¹. Belkin used the same frame in February: "Copernicus displaced humans from the center of the universe, Darwin displaced humans from a privileged place in nature; now we are contending with the prospect that there are more kinds of minds than we had previously entertained" ³.

Chen sets the bar for superintelligence high: AI that would "exceed the abilities of leading human experts across almost all areas of thought," at a level past what humans could reach "individually or collectively" ¹. "Insofar as individual humans have general intelligence, a frontier AI has that too," Chen said, but "our argument for artificial general intelligence (AGI) does not establish the arrival of superintelligence" ¹.

Gary Marcus, the NYU professor emeritus, answered in Nature on Feb. 17 with Walter Quattrociocchi of Sapienza University of Rome and Valerio Capraro of the University of Milan-Bicocca ⁴. "Such claims confuse high performance on benchmarks (often gameable) with real world flexibility, which historically has been at the center of what it means to be an artificial general intelligence," Marcus wrote ⁴. Their longer essay anticipated this month's exhibit: "Reports that language models have produced correct proofs for isolated open problems in mathematics, including specific Erdős problems, do not alter this assessment" ⁴.

What does the objection change? The Lean check stands, since a compiled proof is a fact about one theorem. Marcus targets the leap from one theorem to general intelligence, and there he and Chen sit closer than their headlines suggest. Chen's own sentence names the missing test: "establishing superintelligence across almost all areas would require much broader evaluation" ¹. Both men want a measurement across many environments, and that measurement has yet to happen.

Chen says AGI has arrived; Marcus calls that a redefinition

Both sides

Chen and Marcus split on whether AGI is here. Both point to the same missing test for superintelligence, evaluation across almost all areas, which has yet to be run, so the debate leaves the ledger step unchanged.Sources [1] [4]

The data2 rows · sources

Side	Who	Claim	Source
For	Eddy Keming Chen	Frontier AI already has general intelligence, and evidence such as the Navier-Stokes proof puts superintelligence perhaps within a year.	[1]
Against	Gary Marcus	Such claims confuse benchmark performance with real-world flexibility and redefine what AGI historically meant; statistical approximation falls short of general intelligence.	[4]

Why the step stays small

Three facts hold it down: an internal model, a disputed path and a narrow base.

Start with the model. OpenAI says the system behind the proof began training on Aug. 28, and it stays in-house ¹⁵⁷. The capability component asks about public systems.

Then the path, on which a Lean file is silent. OpenAI's own announcement conceded: "While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models" ⁶. A test counts as uncontaminated when the answer's path stays clear of the test-taker's inputs, and OpenAI itself leaves that open.

Last, the base. OpenAI says the same model solved more than 100 long-standing problems in a month of training, and The Decoder noted that which problems it solved, and how, remain open questions ¹⁵. James Maynard, a Fields medallist at the University of Oxford, named the cost to the field: "So far it's been very difficult to really extract any human understanding from this new AI proof" ⁸. On Sept. 11 he and 24 other Fields Medal winners, Terence Tao among them, called the goals of AI companies and of mathematicians "severely misaligned" ¹². Manlio De Domenico of the Complexity Thoughts newsletter put it plainly: "It has not yet become an officially recognized solution under Clay's process" ²¹.

Where does that leave the reading? Capability moves up 0.4, reported. Reported, because the claim rests on OpenAI's paper and announcement, carried by the press of record and backed by an outside compile of the Lean files; confirmed needs a public system and a clean provenance. This ledger moved capability 0.6 for Anthropic's Fermat formalization on Sept. 4 and 0.6 when FrontierMath Tier 4 fell on Sept. 10. Navier-Stokes has more depth than either: new mathematics on an open problem of the first rank, the test the Fermat piece named. It has less provenance. Fermat also came from an internal model, but it restated a public human proof and Kevin Buzzard put his name to the check; FrontierMath fell to a public model. This result has a disputed path and a Lean statement awaiting a named check. Chen's forecast adds 0.0.

What Lean and Clay have confirmed, and what mathematicians still have to check

The record

The left column earns the step: compiled Lean files, a statement adapted from a third party's, and Clay's word. The right column holds it small: the fit, the unforced case, provenance and the unseen other problems.Sources [6] [7] [8] [9] [10] [11] [13] [15]

The data9 rows · sources

Column	Item	Source
Checked	The Lean code compiled, Javier Gómez-Serrano told NPR	[8]
Checked	Lean files establish Clay's alternatives (C) and (D)	[11]
Checked	Lean statement adapted from Google DeepMind's Formal Conjectures project	[11]
Checked	Clay Institute says the problem has apparently been settled	[9]
Waiting on people and time	Whether the Lean statement matches the problem posed in 2000	[7]
Waiting on people and time	The unforced version of the problem remains open	[13]
Waiting on people and time	Clay's process: publication, two years, general acceptance	[10]
Waiting on people and time	Whether de-identified Codex data helped improve the internal model	[6]
Waiting on people and time	Which of the 100-plus other claimed problems were solved, and how	[15]

Who keeps the ability to change course?

Chen asks whether the ability to evaluate and guide AI can keep pace with its capabilities. Leading AI companies, he noted, say their researchers increasingly use AI agents to run experiments and train future AIs ¹. "If we delegate more of the development of future AI to systems whose reasoning we do not fully understand, how do we retain the ability to assess the results and, when appropriate, change course?" Chen asked ¹. The Decoder reported on Sept. 27 that Anthropic puts the human share of research-direction decisions in single digits, and that OpenAI runs GPT-5.6 Sol across its development cycle; a study it covered, from a team building its own model, found people made the final call on goals in 93.4 percent of cases ¹⁶.

OpenAI answered the mathematicians with an independent advisory group and one stated limit. "Importantly, the group will not be responsible for advising us on how to pace our internal progress on mathematics," the company wrote ⁸. OpenAI, which supplied Chen's exhibit, kept the pace for itself.

His remedy is access, with universities pairing technical research with philosophy and the social and data sciences, and giving researchers room to scrutinize developers' claims ¹. "We especially need independent and rigorous evaluations of what the most capable AI systems can do and whether their safeguards still work," Chen said. "That requires access to the systems and the freedom to publish critical findings" ¹. He also sets aside a human ceiling: "There is no reason to assume that human intelligence represents a fundamental ceiling for artificial systems" ¹.

UC San Diego takes up the argument in October. Ted Chiang, whose "Story of Your Life" became the Academy Award-nominated film "Arrival," gives a free lecture at the Halıcıoğlu Data Science Institute on Oct. 16, in a series honoring the late author and San Diego State professor Vernor Vinge ¹¹⁷. Its title disputes a premise of Chen's Nature Comment: "No Machine Has Passed the Turing Test" ¹⁷. On Oct. 17, Chiang and Kim Stanley Robinson, a UC San Diego alumnus, hold a free conversation and a 10th-anniversary screening of "Arrival" at the Epstein Family Amphitheater, presented by ArtPower ¹¹⁸. Chen closed with the last sentence of Alan Turing's 1950 paper: "We can only see a short distance ahead, but we can see plenty there that needs to be done" ¹²⁰.

By the numbers

166 pages in OpenAI's Navier-Stokes paper, posted Sept. 8 ⁵
Roughly 10,000 agents for about 88 hours, then 17 hours of Lean work by GPT-6 Astra ⁶⁷
About 130 billion output tokens on Navier-Stokes, out of 300 billion across every problem attempted ⁶
$6 million to $10 million: NPR's estimate of the Navier-Stokes compute at public prices ⁸
Two years at minimum between publication and Clay's consideration of a solution ¹⁰

What to watch

Clay's review is the slow, decisive marker: peer-reviewed publication and general acceptance would lift this step toward the confirmed band. Sooner, named mathematicians checking the Lean statement against Fefferman's text would settle the fit question. OpenAI publishing its 100-plus problems, misses included, would test breadth, which Chen and Marcus both ask for. The open question that decides the next move is Chen's own: whether outside researchers get access to these systems, and the freedom to publish what they find, before the next result arrives.