Terence Tao on Proof Assistants, AI Mining Four Centuries of Field Notes, and Cats That Blink Back
How this was made Verified AI
Every Intellegix briefing is generated from that day's broadcast and run through automated checks before it publishes — with a human paged on any flag. Here is the trail for this edition.
Fields Medal winner Terence Tao published a blog post addressing what mathematicians should know about the Lean theorem prover — specifically, questions of reliability and AI's role in formal mathematics — drawing 129 upvotes and 25 comments. Lean is a proof assistant: a programming language in which mathematical proofs can be expressed and formally verified by computer. It has been gaining traction as AI models are increasingly used to help generate proof steps, and several high-profile results have been formalized within it.
Tao's post, as read by commenters, addresses two related questions. First, how reliable is Lean itself — can a proof Lean accepts be trusted as correct? The answer is nuanced: Lean's kernel is small enough to be audited and trusted, but the broader library ecosystem on which most proofs depend can contain its own errors. Second, what does AI assistance actually guarantee in this context? A proof step that an AI suggests and Lean formally accepts is verifiably valid — but Lean can only verify formal validity, not whether the formal statement correctly captures the mathematical question the researcher actually intended to ask. That gap between formal correctness and mathematical meaning — the specification problem — means AI assistance does not obviously resolve the deepest layer of uncertainty, and may introduce high confidence in a subtly wrong answer if the formalization itself is off. Tao's optimism about AI in mathematics, the discussion noted, is explicitly conditioned on recognizing this problem remains unsolved.
The more immediately striking AI application surfaced in a post about mining four centuries of historical archives, which drew 151 upvotes and 77 comments. A researcher ran AI analysis across digitized naturalist field notes, expedition records, and specimen catalogs spanning roughly 400 years, and surfaced findings human researchers had missed: a forgotten meteorite impact record, documentation of rhinoceros populations in regions where they are now considered never to have lived, and several other rediscoveries. The economic case for digitized archives has always been theoretically clear and practically underdeveloped because reading four centuries of handwritten field notes is an enormous undertaking. Language models capable of handling historical orthography, cross-referencing across documents, and flagging anomalous entries change that calculus — extending researchers' reach across document collections that would otherwise require lifetimes to process manually.
Jane Street's engineering blog contributed a post on autoregressive diffusion for generating synthetic market data, sitting at the intersection of AI and quantitative finance. The firm explored whether diffusion models trained on historical market data could generate synthetic scenarios preserving real markets' statistical properties — fat tails, time-varying autocorrelation structures, discrete price levels, and feedback loops between prices and participants. Practical applications would include backtesting trading strategies against synthetic scenarios rather than being limited to historical data, stress-testing against tail events, and reducing the overfitting risk that comes from repeatedly testing against the same historical record. The post reached appropriately cautious conclusions.
From the 'paper awaiting its viral moment' category: a 2020 Nature study from the University of Sussex on cat slow blinks resurfaced on HN with 90 upvotes and 36 comments. The research found that cats use eye-narrowing as a positive affiliative signal toward humans and that humans can reciprocate with measurable effect. The comment thread was, by general consensus, charmingly empirical.