AI Chatbot Company Says Its AI Spotted an Unknown Set of Genes in Viruses, and Its Lab Confirmed They're Active
Anthropic, the company that makes the Claude chatbot, says hundreds of copies of Claude combing through DNA databases noticed an unusual group of genes in viruses that infect bacteria, something no one had described before. Human scientists in the company's new lab confirmed that part of the system is active. What the system actually does is still unknown.
Anthropic, the company that makes the Claude chatbot, today announced that it has opened a biology laboratory, and shared its first result: a set of genes it says Claude was the first to notice.
Here is what happened. Anthropic set roughly 950 copies of Claude working for about 21 hours, searching huge public databases of DNA. They were hunting for a particular kind of enzyme, a protein that copies genetic instructions from one chemical form, RNA, into another, DNA. The copies gathered more than 200,000 of these, picked out about 3,500 unfamiliar candidates, and narrowed those to 20 they wrote up as reports for human scientists to read.
One report described something new. In viruses that infect bacteria, one of these copying enzymes sits right next to a partner gene nobody understands and a long row of short, evenly repeated stretches of DNA. The enzyme itself had been seen before; Anthropic says Claude "appears to be the first to notice" the partner gene and the row of repeats that go with it. That arrangement is interesting because, the company says, only a handful of other natural systems combine these features, and all of them turned out to be able to cut, copy or paste DNA. The best known is CRISPR, which scientists turned into a widely used gene-editing tool.
Human scientists in Anthropic's lab then put the genes into laboratory bacteria and found the row of repeats is actively read out as a set of short RNA pieces, a first sign the system does something. What it does, the company says plainly, "we don't yet know."
There are reasons for caution. The work is a preprint, shared publicly but not yet reviewed by outside scientists, and it comes from the company selling the AI. The Next Web, a technology news site that read the paper, reports that when the search was run ten more times, none of the runs found the same pattern again. Feng Zhang, a CRISPR pioneer at MIT and the Broad Institute who reviewed the preprint for Anthropic, called the finding "genuinely intriguing" and said it "merits further investigation."