Dario Amodei posted a longer thread on the ART enzyme discovery this blog has now covered twice. Two things in it are worth separating: a more careful account of what actually happened, and an analogy to AI progress in math that doesn't hold up well against this blog's own reporting from the past two weeks.
The division-of-labor framing here is better than the launch post's
Amodei writes: "our life sciences team suggested a broad area of research, Claude read through the literature and a bunch of genome data and discovered something interesting, then Claude proposed experiments to verify the discovery and our team carried them out." That's a cleaner, more falsifiable breakdown than the announcement's "only high-level direction" phrasing, and it matches what the actual technical report shows: humans set the research brief and ran the wet-lab work, Claude agents did the genome mining, hypothesis generation, and experimental proposals in between. Credit where due — this is Amodei being more precise than his own company's marketing copy, not less.
The "independent discovery" claim checks out, with a timing wrinkle
Amodei states that "a Stanford team working independently described a novel RT system with an associated non-coding array that is in some ways similar to the one Claude found, though they are distinct systems that evolved independently from each other." This is broadly accurate and worth taking seriously: bacterial RT-and-noncoding-array defense systems are an active research area with real, recent Stanford-linked work — DRT3, a dual-reverse-transcriptase system pairing two RTs with a non-coding RNA template to build repetitive double-stranded DNA, appeared on bioRxiv months before Anthropic's announcement. Whether that specific paper is the one Amodei means, or a more recent one from the same active line of research, this points to something worth being precise about: ART isn't a discovery in an empty field. It's one new entry in a genuinely crowded, fast-moving area where multiple groups — some using AI, some not — are independently finding related RT-and-repeat systems in the same window of months. That's a real point in favor of the underlying biology being genuinely interesting; it complicates the framing that only an autonomous AI agent, reading raw DNA no human was looking at, could have found something in this space.
The math analogy runs into two weeks of this blog's own coverage
Amodei's argument for why AI-for-biology should be taken seriously rests on a specific historical claim about math: "In 2023 models struggled to do math at the level of an average high-school student... in early 2026 more significant open problems, and in late 2026 they are beginning to solve the top few open problems in all of mathematics." That last clause is the load-bearing one, and it's worth checking against what's actually been reported here recently rather than taken as settled background.
This blog covered OpenAI's Navier-Stokes claim — arguably the single most prominent recent instance of an AI lab claiming progress on a top-tier open problem — three weeks ago, and the coverage wasn't flattering: the mathematician who'd actually done related unpublished work, Tristan Buckmaster, said he'd contacted OpenAI first to flag it, that OpenAI's own math lead told him "very little human input" had gone into their result (a claim Buckmaster's own account contradicts), and that a direct question about whether the model had trained on his and a co-author's own draft work went unanswered. Eleven days before Amodei's thread, 25 Fields Medalists signed a declaration that opens by conceding LLMs "can solve major outstanding problems in many fields of mathematics" — then spends the rest of the statement objecting to exactly the pattern the Navier-Stokes episode showed: "rushed announcements with no writeup, unattributed prior work, and a broken transmission chain."
Amodei's sentence asserts the capability trajectory in exactly the terms that declaration was responding to, without the caveats the mathematicians closest to the actual results have been attaching to it in the same window of time. None of this means the underlying trend is fabricated — capability gains in math are real and this blog has covered several of them. But "beginning to solve the top few open problems in all of mathematics," offered as the evidentiary anchor for why biology should be expected to follow the same curve, is asserted here with more confidence than the specific recent case this blog checked in detail actually supports.