Anthropic's Crispr-like find from 950 Claude agents still waits on the wet lab

Partway through the run, one of the agents got excited in writing: "I can see by eye a tandem repeat array ... that's a Crispr-like ... repeat array?!" ^[10]. The same agent allowed that the system could be a retron, a caveat the technical report records ^[11]. Crispr and retrons are both bacterial immune systems, and retrons have some gene-editing use without Crispr's range, Wired reported ^[12]. Anthropic's announcement led with the Crispr comparison anyway ^[1], which Wired called not too surprising ^[13].

Here is what was pitched: a possible new gene-editing system. Here is what was done. Researchers at Anthropic prompted Claude to search genomic databases for "interesting new examples" of reverse transcriptases, proteins that copy RNA into DNA ^[5]. Seth Shipman of the Gladstone Institutes, whose lab has built gene-editing systems out of retrons, does not think ART is a new Crispr system ^[14][16]. "The novel thing is how they found it, not what it is," he said ^[15]. Mining databases for new reverse transcriptases by hand can take months, and Shipman called doing it in a day impressive ^[17]. Jason Gill of Texas A&M agreed on the pattern-finding. "These models are good at finding patterns, better than a person staring at it with their eyeballs can," he said ^[19].

Shipman also drew a line on credit. "I think we have to be careful about saying that Claude autonomously discovered something, because there are scientists involved in the study," he said ^[18]. People chose the protein class and wrote the prompt ^[5].

The compute side is easy to size. About 950 agents running for 21.5 hours comes to roughly 20,400 agent-hours ^[1]. Wired's account does not include what that cost.

The bench side is harder. Fyodor Urnov of Berkeley's Innovative Genomics Institute, which collaborates with Anthropic but was not involved in this work, said: "I sincerely compliment Anthropic for telling the world about their discovery" ^[23]. Stanford's Le Cong was cooler. "The experiments are still in the queue. The PR is already live," he said ^[20]. Cong compared the search to scanning Santa Monica Beach for a diamond: "AI found this thing that looks very shiny," he said, and then the lab has to find out whether it is glass ^[21]. Anthropic has set up its own wet lab for drug discovery and says the finding is not the end of its work ^[22]. Whether ART can edit genes at all, and whether it would be a useful editor, is unknown ^[4].

For the research lead deciding whether to fund agent-driven search, I'd sort each AI-for-science claim on two axes. The first is whether the search beats the manual work it replaces. The second is whether a bench has confirmed what the output does. Fast and confirmed: budget against outcomes. Fast and unconfirmed: budget the agents as triage, priced against months of manual database mining, and fund validation as its own line. Slow and confirmed describes the process most labs already run. Slow and unconfirmed is a pass.

ART sits in the second cell. Shipman and Gill vouch for the speed, and the bench record is one experiment in a report that has not been peer-reviewed ^[17][3]. My recommendation is to keep the Crispr comparison out of any business case until the lab reports what ART does. The tradeoff in that cell is that cheaper triage sends more shiny candidates to the same bench, and each one needs the kind of experiments Cong says are still queued for ART ^[20].