Anthropic quietly built a wet lab for its AI drug push
Anthropic's answer to "AI is only a chatbot" is a wet lab in the Bay Area — and a federal flash-flood model that starts catching storms earlier than the forecast does.
Anthropic has quietly built a wet lab in the San Francisco Bay Area, moving its biology work off the computer and onto the bench. Two people familiar with the matter told Reuters the startup now runs physical experiments, and Anthropic's head of life sciences, Eric Kauderer-Abrams, confirmed the lab in an interview: "We believe that to do biology, the final test is still and will be for a while in real lab work." The company also wants Claude directing robotic units through experiments with limited human intervention — though a spokesperson said human oversight stays essential, and clarified the lab is not for drug discovery specifically.
The boundary Anthropic drew is the interesting part. Kauderer-Abrams told Reuters the company won't run clinical trials for now and is aiming at conditions the industry considers "undruggable" — neglected rare diseases, complex bispecific and trispecific antibodies — because "we're not competing with pharma and biotech companies that make their business in bringing drugs to market." That is a trust play as much as a scientific one: Anthropic sells models to Roche's Genentech, Bristol Myers Squibb and Novo Nordisk, and those customers have real reason to worry an AI vendor is reading their pipeline. Walling off customer data while building your own lab is the kind of claim that gets tested in contracts, not press releases.
The timing is the tension. In the past two weeks Anthropic researchers warned publicly that AI could lead to human extinction, the company reported finding cases where its systems could have been used toward biological weapons development (a finding that drew expert skepticism), and CEO Dario Amodei called for a development slowdown — while the company prepares a public listing that reports have pegged near a $2 trillion valuation. Kauderer-Abrams frames life sciences as the biggest upside of the mission; the same capability is what his own safety team flags as the biggest downside. Anthropic has spent the year staffing up for it: Novartis CEO Vas Narasimhan joined its board, it bought Coefficient Bio for about $400 million in stock, it shipped Claude Science, and in August it published a Model Hardware Standard so AI can drive lab equipment. Most drugs fail trials, and the company has not said which diseases it is chasing. We covered the measurement side of this push yesterday — Claude now leads 26% of the research at the lab that built it — but automation metrics are cheap; physical results are not.
A NASA-funded machine-learning system is giving US forecasters a real-time head start on flash floods, and it goes nationwide next month. TACLS — the Transient Artifact and Continuous Learning System — reads water vapor in the atmosphere from GNSS satellite signals and uses a long short-term memory model to flag where ordinary rain is turning into a life-threatening flood, well before the standard forecast catches up. It is already running at the National Weather Service offices in Los Angeles and San Diego; a version with map-based graphics is loading onto NWS systems now and will reach every forecast office in the second half of October.
The build is small and unglamorous: a UC San Diego graduate student spent about a year on the machine-learning side, and TACLS does not replace anyone. It sits next to the rain gauges and flash-flood guidance forecasters already use. Its false-positive handling is the part worth noting — an isolated sensor reading gets suppressed unless neighboring stations see the same signal, a sensible answer to the sensor-goes-funky problem that sinks a lot of operational ML. The system's reach is limited by where GNSS sensors happen to be dense, which skews it toward the earthquake-prone West. That is a real ceiling. It is also the rare AI deployment where the failure mode is a missed warning rather than a bad answer, and where being ten minutes early is the entire product.
If an AI model can flag a flood ten minutes early but only where sensors are dense, should federal money go to sensors or to models? Tell us in the comments.
