First-hand information for everyone
Main finding: The author argues that recent AI achievements in mathematics have intensified predictions of human intellectual obsolescence. But they do not prove that current systems can invent new concepts or sustain autonomous, recursive self‑improvement. (These are the two claims at the core of the singularity scenario.)
Concrete examples and methods: In 2023, hundreds of leading AI researchers signed a public statement equating extinction risk from AI with pandemics and nuclear war. Public alarms grew when a researcher’s social‑media thread drew more than 170 million views and prompted responses from industry figures.
In September 2026, OpenAI announced an AI‑generated solution to a Navier–Stokes problem. According to reports cited by the author, the work was said to have used as many as 10,000 AI “agents” working in tandem and many millions of dollars’ worth of computing. On October 6, 2026, the company released hundreds of AI‑generated manuscripts claiming advances across many mathematical fields.
The author’s own use of AI has been as a guided research partner. He describes a constant dialogue where AI helps explain papers, check calculations, test examples, and assist exposition—always under human direction.
Assessment and limitations: Large language models learn from vast collections of finished writing. They do not have access to the private imaginative processes that originally produced those works. Left without directed human judgment, they often recycle low‑quality material and can “hallucinate”—that is, invent facts, quotations, or references and present them confidently.
Human intelligence, the author stresses, also includes choosing worthwhile goals, judgment about persistence or change, imagination, and embodied experiences such as emotions. Standard tests and current benchmarks fail to capture these capacities.
On proofs and verification: A machine proof obtained mainly by exhaustive case checks or fast search can establish correctness while offering little conceptual insight. Historically, conceptual frameworks have mattered more than isolated correctness. Formal verification in a proof assistant (for example, Lean) can substantially aid checking. But being “Lean‑verified” is not by itself conclusive. One must inspect the formalized statement, assumptions, and completeness. Reported discrepancies exist between an announced Navier–Stokes argument and its Lean formalization.
About the announced Navier–Stokes construction: The reported result addresses a forced version of the equation with a specially designed external forcing term, a case allowed by the Clay problem’s statement. The author judges such constructed blow‑up solutions to be exceedingly complicated, likely highly unphysical, and probably unstable, though that instability remains to be established.
On autonomous self‑improvement: Some systems (like AlphaZero) improve via self‑play inside a fixed, human‑designed framework. Proposals such as the Darwin Gödel Machine and AlphaEvolve can propose and test algorithmic changes. But none of these systems demonstrate a self‑sustaining acceleration of general intelligence like the singularity scenario requires. The author draws an analogy to the biological challenge of creating self‑replicating life from nonliving chemistry. That problem has revealed deep, interacting complexities and suggests caution about assuming near‑term autonomous, self‑sustaining AI takeover.
Community impact and proposals: The mass release of AI‑generated mathematical claims has been deeply demoralizing for parts of the mathematical community. It threatens norms of open exchange and attribution. It also places heavy burdens on researchers to check, interpret, and claim credit for results.
To protect unpublished ideas and preserve scientific collaboration, the author proposes an academic AI consortium. The consortium would provide powerful research tools with enforceable guarantees that unpublished work will not be used for model training or shared without explicit consent.
Bottom line: AI is already a powerful tool that can accelerate parts of mathematical work when guided by human judgment. But, according to the author, current evidence does not show that AI can autonomously invent new mathematical concepts or trigger self‑sustaining recursive improvement. Significant technical, epistemic, and cultural challenges remain.