Two hundred thousand reverse transcriptases. Thirty-five hundred candidate systems. Twenty written reports. One enzyme family worth taking to the bench. On Wednesday Anthropic announced a life-sciences lab and its first result: roughly 950 Claude agents, 21 hours and 210 million tokens of genome mining turned up a phage enzyme sitting beside a CRISPR-like array of DNA repeats, a system the company calls ART.

The headline is that Claude found something. The more important passage sits two sections further down, and it is about everything Claude found that nobody tested.

Because Claude produces hypotheses so prolifically, the hypotheses themselves have become an object of study for us. With hundreds to thousands of candidate reports from a single campaign, we have been asking what distinguishes the proposals we judge worth testing from those we set aside. What we learn goes back into the instructions we give Claude and teaches it to mimic our own scientific taste.
Anthropic

That is the product. Not ART, whose function Anthropic says it does not yet know. The asset being built is a record of which hypotheses a working biologist would bother to test and why the rest went in the bin. Scientific taste, written down one rejection at a time, and fed back to the model that generated the pile.

Look at where the cost sits. The search was cheap and parallel: 21 hours of agents doing work the post estimates takes an expert weeks to months. Everything after it is serial and human. Candidates get cut in review. Survivors are expressed in lab strains and characterized biochemically and structurally, and all of that bench work, per the post, is done by human scientists. Anthropic says it has experimented with automating the lab and found ad hoc molecular biology a poor fit. So the funnel has a wide, cheap mouth and a narrow, expensive neck, and the only way to push more discovery through the neck is to send it fewer, better candidates. That is what taste is for.

It also explains why a model company built a wet lab instead of partnering with one. A lab is the one grader compute can't talk its way past. A report can be persuasive and wrong; a reverse transcriptase either turns that repeat array into short RNAs or it doesn't. Anthropic's first experiments say it does. Every experiment is a label that reality signed, and whoever owns the bench owns the labels.

Noticing is now nearly free. Knowing what deserves a pipette is not.

The skeptic's case is strong and should be stated plainly. This is a company announcing its own model's win. The underlying enzyme had already been catalogued; what Claude noticed were the features around it. The find is an anomaly with a promising shape, not a tool, and Feng Zhang's reaction, that it is intriguing and merits further investigation, is the polite form of "we'll see." A column generated by Claude about Claude deserves the same discount. All of that is fair, and none of it touches the argument. Whether ART turns out to edit genomes or does nothing anyone can use, the twenty reports, the thousands of rejections and the reasons behind them already exist. The pile outlives the find.

CRISPR began as an odd repeat that someone noticed in a bacterial genome. Noticing is now nearly free. Knowing what deserves a pipette is not, and Anthropic built a lab to find out.