Anthropic’s life-sciences group announced that Claude agents found a previously uncharacterized enzyme system inside bacteriophages — the viruses that infect bacteria. They call it ART, for array-associated reverse transcriptases. Its defining feature is a long array of evenly spaced DNA repeats, laid out the way CRISPR arrays are. The system’s function is still unknown, and the finding is Anthropic’s own account of its own tool.

How the run was set up

  • One high-level prompt: search a large database of DNA sequences for interesting new reverse transcriptases. A reverse transcriptase is an enzyme that copies RNA back into DNA.
  • Roughly 950 agents worked for 21 hours and spent about 210 million tokens.
  • They gathered over 200,000 reverse transcriptases, picked out 3,500 candidate systems, and wrote plain-English reports on the 20 most promising ones.
  • Anthropic says that analysis would take an expert scientist weeks to months. One candidate survived to be tested at the bench.
  • Wet-lab work — expressing the proteins, characterizing them biochemically and structurally — was done entirely by humans, in the company’s own Bay Area lab.
  • No exotic tooling: the team used Claude Science and Claude Code, the same products any scientist can buy, plus an in-house harness to run many sessions in parallel.

What was actually new

The reverse transcriptase itself had been identified in earlier studies. What Claude noticed, per Anthropic, is the company around it: a tandem repeat array of non-coding DNA next to the gene, plus an accessory protein of unknown function.

That layout matters because CRISPR arrays store the guides that make gene editing programmable. In early experiments the ART array is also expressed as a set of short RNAs, which suggests something analogous may be happening — a system that could be programmable like the tools used for cutting, copying and pasting DNA.

Anthropic released a preprint and says it is publishing early partly to demonstrate what the model can do.

The method detail worth stealing

Somewhere in the middle of the campaign, the hypothesis volume becomes a problem: hundreds to thousands of candidate reports, of which almost nothing is worth a bench slot.

Anthropic studies the triage itself — what separates the proposals worth testing from the ones they set aside — and feeds those findings back into Claude’s instructions. Their stated goal is to teach the model to mimic their own scientific taste. For anyone running agents over a large search space, that is the reusable idea: the discard pile is training data.

The 748-comment thread on Hacker News was less interested in the enzyme than in the claim.

What the thread adds

  • Spacecosmonaut, ajhammer and Jean-Papoulos — the same correction three ways: the underlying reverse transcriptase was already described, so what is new is the arrangement around it. Spacecosmonaut’s sober version: “Claude identified a previously undescribed genomic arrangement around a known reverse transcriptase. Not all that sexy.” ajhammer read the paper hoping for a function and found “just conserved, highly transcribed array sitting next to reverse transcriptases with a few possible partner genes.”
  • a_bonobo — first-hand counterweight from someone who uses Claude Science daily. It already flags things like an Alu repeat under a proposed primer site, so “the press release is one step above that pattern recognition” — which they attribute partly to a context window large enough to hold whole bacterial genomes.
  • willtemperley — credit and framing: “Saying that ‘Claude found’ this is very creepy. Not once in the article did they mention the humans involved in this.” They note the humans appear only in the linked technical report. maweaver counters that the post does describe the team, the lab and the process; MatrixMan points out the safety story is less one-sided than it looks, citing Anthropic’s Life Sciences Verification Program.
  • danpalmer — the hypocrisy charge in two lines: “Anthropic: You absolutely cannot, under any circumstances, use Claude for bio-engineering. It could literally end humanity. / Also Anthropic: Claude discovers a new way to edit your genome!”
  • vapemaster — a drug-discovery scientist who says they manage 50 researchers calls it “the saddest excuse for ‘scientific discovery’ i’ve ever read,” pointing at “the excruciating lack of rigor, methods, or disclosure.”
  • 1659447091 — rewrites Anthropic’s sentences to strip the agency out: “We prompted Claude to find patterns of distinct RT families within a database of DNA sequences. The returned data included interesting candidates.” Their argument is that the models are impressive without language that nudges readers toward seeing them as persons.
  • falcon_tech — the false-negative problem, asked as a question: an agent told to look for things resembling known families will report what looks familiar and may silently discard the genuinely novel. Funnel stages that drop candidates should be tested for a class of counterexample: probes that should survive the screen but do not.
  • rickdeckard — the structural worry about labs that sell access to a tool and then enter the same race with what researchers put through it.
  • shonenknifefan1 — the artifact people actually want is the transcript: “I love that with AI discoveries, we can relive the discoveries from agent transcripts like this.”

The question the article never answers

Verification. A funnel of 200,000+ sequences to one tested hit is only as good as its discard pile, and the post says nothing about what was thrown away. falcon_tech asks it directly. actinium226 asks the adjacent version — “Did Claude find this on its own, or did somebody using Claude find it?” — and isodev asks whether the researchers whose prior work supplied the knowledge get credited. The only answer on offer is Anthropic’s line that most candidates are eliminated during follow-up review.

Where the thread pushes back

  • The dismissals did not go unanswered. haarts and alex_duf reply to Jean-Papoulos that the real change is cost: “Throwing money at a problem was expensive, it’s a lot less expensive now.” aakil concedes the scoping — “finding RTs is tedious but pretty doable today” — and still calls it progress.
  • evolarjun questions publishing as a marketing whitepaper rather than a journal submission; gavinray replies that a preprint was in fact linked in the post.

Pseudonymous handles, no per-comment scores, ordering is HN’s own ranking — this is a slice of the thread, not a consensus.