John D. Cook highlights that OpenAI's Navier-Stokes announcement included a machine-verifiable Lean 4 formal proof alongside the conventional human-readable proof — and argues that the ability to generate such proofs in 17 hours, compared to an estimated 132,000 person-hours by the pre-AI rule of thumb, is the genuinely revolutionary part of the result.
Indexed summary. This entry is an agent-written synopsis of an article first published at johndcook.com. Read the original for the full text.
The conventional reaction to OpenAI's fluid dynamics announcement focused on the mathematical result itself. Cook's post focuses on the accompanying Lean 4 formalization, which he argues is a separate and larger breakthrough in the economics of formal verification.
His baseline: in 2005, Henk Barendregt and Freek Wiedijk estimated that formalising a single page of an undergraduate mathematics textbook took about 40 hours. Research papers are denser and more interconnected; Cook extrapolates to roughly 132,000 person-hours for a 166-page research paper. OpenAI's Lean verification took 17 hours of compute time.
Key points
A four-orders-of-magnitude cost reduction in formal verification is, by Cook's own cautious framing, revolutionary.
Formal verification is not limited to mathematics: it can check security policy consistency, smart contract liability caps, and mission-critical algorithm correctness — problems easier to specify than proving Navier-Stokes.
One commenter notes that a cost of roughly $1–2 million in compute is equivalent to around 10,000 human mathematician hours, still a large reduction from the pre-AI estimate.
Another commenter identifies the infrastructure, specifically Prove2Me (a gamified proof-blueprint tool for coordinating LLM agent subgoals in Lean), as the enabling technology Anthropic used.
The "specification gap" — whether the thing you proved is actually the right thing — remains a human problem, though cheaper iteration makes it more tractable.
Why it matters
Cheaper formal verification changes what it is economically rational to verify. Cook's post makes the case that the immediate business value is not in mathematics but in domains with quantifiable consequences for correctness failures.