Outside biologists narrow Anthropic's enzyme claim to architecture, not function

The pipeline Anthropic describes is an agentic literature-and-sequence search, not a model reasoning its way to a biological insight. By its own account roughly 950 Claude agents worked over DNA sequence databases for about 21 hours, gathering over 200,000 reverse transcriptases, flagging around 3,500 candidate systems and narrowing to 20, before one agent noticed an RT gene sitting next to an array of DNA repeats resembling CRISPR. These are Anthropic's own figures about its own system, so they are claimed rather than confirmed however plausible they look. The token-consumption figure is omitted entirely: the same page returned 210 million on one read and 200 million on another.

Anthropic then ran wet-lab work in house — protein expression in standard laboratory strains, biochemical and structural characterisation — and reports confirming that the array expresses distinct short RNAs. That is a real experimental result and it is the load-bearing part of the claim, but it was produced by Anthropic's own scientists and has not been independently replicated. Feng Zhang of MIT and the Broad Institute reviewed the findings and called them genuinely intriguing and worthy of further investigation. Review is not replication.

The corrective came from biologists with no stake in the result. Dimitri Perrin, who heads the computer science school at Queensland University of Technology, made the two distinctions that matter: the enzyme itself was already known, and what is new is the recognition that it may form part of a larger system; and the system is CRISPR-like in architecture with no evidence of being CRISPR-like in function. Kevin Blake, a microbiologist at Washington University, was blunter on applications, saying there is nothing to indicate a rival to CRISPR as a technology or a route to any therapeutic use.

There is a failure mode specific to this method that the coverage named: a search run at this scale over repeat-rich genomic data will surface patterns, and architectural adjacency between an RT gene and a repeat array is exactly the kind of pattern that looks meaningful before anyone knows whether it is. Anthropic's own wording is appropriately hedged — it says it does not yet know the function and that work to understand it is ongoing. There is no peer review and no journal submission, only a technical report on Anthropic's own servers.

• Claimed Anthropic says roughly 950 Claude agents searched DNA databases for about 21 hours, gathered over 200,000 reverse transcriptases and flagged around 3,500 candidate systems, narrowing to 20. Anthropic
• Claimed Anthropic's in-house wet-lab work confirms the repeat array expresses distinct short RNAs; the system's primary function is stated to be unknown and the work ongoing. Anthropic
• Reported Named outside biologists state the reverse transcriptase itself was already known, and that the system is CRISPR-like in architecture with no evidence of CRISPR-like function. Gizmodo