AI just moved into the wet lab

Anthropic spent the spring building something a model lab does not usually build: a wet lab. On September 23 it published the first result — Claude, running as 949 autonomous agent sessions over 21.5 hours, flagged a pattern in bacteriophage DNA that nobody had annotated. The company calls it ART. The honest version is narrower than the CRISPR comparisons travelling with it.

What Claude actually found

ART — array-associated reverse transcriptases — is a reverse transcriptase, a partner gene beside it, and an array of evenly spaced DNA repeats running 3 to 21 copies. It sits mainly in phages and carries none of the cas genes a CRISPR system would have nearby. The agents screened roughly 200,000 reverse transcriptases unsupervised; human scientists then found the array is transcribed as distinct short RNAs, hitting 8% of all phage RNA fifteen minutes after infection.

Now the subtractions. The enzyme was not new — earlier work described it in a jumbo phage, and what Claude spotted was the array and partner gene around it. Nobody has shown the enzyme is active, that those RNAs are its substrates, or that any of it edits DNA. The function is unknown, and this is a preprint. Resemblance to a CRISPR array is a reason to investigate, not a result.

The finding is about context, not biology

The tempting read is that a general model beat the specialist genome models. It did not: Anthropic ran Evo 2 and gLM2 on the same locus and both identified the repeats with high confidence. What separated attempts was how much raw DNA reached the model — recognition climbed from 29% of attempts to as high as 96% once a few hundred nucleotides were in context. Interpretability work then found internal signals that fire on repeated DNA, and on repeats that are not DNA at all. Generic pattern machinery, aimed at a genome.

The money, and the shadow

The same day, Enveda raised $311M to push nature-derived AI drug candidates into trials. Biology is where the capability argument is being cashed now — and where the risk frameworks point, since unsupervised search across sequence space is the capability those evaluations were written for. Related: Bioinformatics; Healthcare; Astra and critical cyber.

The point

This result is a promising lead, not a discovery. The agents screened 200,000 candidates, far more than any researcher could by hand, and found one anomaly worth a week of lab testing. The slow step is now the lab work itself, which is why Anthropic built its own lab. Judge the next announcement on whether the enzyme is shown to actually work. Related: OpenAI's Millennium Prize fight.