Anthropic claims Claude identified a previously uncharacterized phage enzyme system in 21 hours
Anthropic's life sciences team reported that approximately 950 Claude agents searched public DNA data for 21 hours, utilizing about 210 million tokens, to identify a previously uncharacterized phage enzyme system, named Array-Related Reverse Transcriptase (ART). Dario Amodei stated that the precise functions, biotechnological applications, and significance are still unclear. Feng Zhang reviewed the preprint and described the discovery as genuinely intriguing.
The funnel reduced over 200,000 reverse transcriptase clusters to about 3,500 candidate partner families, resulting in approximately 20 reports for human review. The ART trio includes: a reverse transcriptase known to exist in giant phages, a partner gene with unknown functions, and a uniformly spaced DNA repeat array. The agents visually identified tandem repeats next to the original sequence and documented them as a CRISPR-like array. Early lab experiments indicated that this array expresses a set of short RNAs. Ten repeated searches did not capture the array again.
Anthropic concurrently established a life sciences lab, limited to BSL-1 and BSL-2, using Claude to scan uncharacterized protein families and then validate them in wet experiments. The underlying literature on reverse transcriptases already exists, and the company noted that this set of agents recognized the repeat array and accessory proteins. The discovery was described as a workload that a PhD student would be proud of, not a tool for editing genomes.
The 210 million tokens represent a supervised literature mining effort, not an automated publication. Humans posed questions, conducted experiments, and wrote preprints; the agents contributed by highlighting repeat structures in noisy sequences. The false negative rate indicates that the same prompt cannot serve as a reproducible discovery engine. The CRISPR analogy sells attention, not already aligned cutting activity.
In market mechanisms, what is sold is the model's ability to generate hypotheses, while what is bought is the endorsement of laboratories and preprints. Demand comes from model companies wanting to demonstrate scientific utility, while supply consists of public genomes and wet lab capacity. The beneficiaries are those who can incorporate discoveries into narratives of safety and capability, like Anthropic, while the pressured parties must explain the "unclear significance" of scientific edits. Funding flows from inference bills into lab reagents, then exchanged for research reputation through preprints.
Source: Public information
ABAB AI Insight
Anthropic extends its cutting-edge models from conversational products to its own wet lab, with ART being the first external operation. The 950 agents working in parallel treat the context window as a cheap research assistant. The real scarcity lies in subsequent biochemistry: the array expressing short RNAs only indicates it has been transcribed, not that it can memorize viruses or cut genomes. Zhang Feng's comments provide professional permission, while Amodei's reservations prevent it from being labeled as a second-generation CRISPR.
The capital pathway serves the narrative of research to model valuation. The life sciences team demonstrates that Claude can find patterns in unstructured sequences, but also proves that patterns do not equal tools. The ten repeated runs failing to capture the "discovery" turn it into a sampling success. The White House requested new models to pass U.S. security reviews in the same week, with scientific news and export controls operating within the same company.
This is similar to the announcement of seeing Google DeepMind's protein folding preprint before waiting for pharmaceutical translation. The industry is at a stage where agents scan public databases in exchange for preprints. Those who can turn uncertain enzyme systems into reproducible experiments will be the ones who convert tokens into intellectual property.
Structural judgments belong to the eve of technological replacement. The labor hours of literature mining are replaced by agent clusters, while experimental validation remains human. The mechanism is: the cost of discovery drops to a single inference bill, while the cost of confirmation remains in petri dishes; media headlines consume the former, while patents consume the latter.
ABAB News · Cognitive Laws
Agents can highlight repeat sequences but cannot sign off on petri dishes.
A discovery in twenty-one hours may not be found in ten repeated runs.
Enzymes of unclear significance are first used for model valuation, then for laboratory use.