Warp Sciences Trains Lab Robots on Chemists' Recorded Demonstrations
Kevin Weil, formerly science lead at OpenAI, has launched Warp Sciences, an AI startup that aims to automate laboratory work for drug discovery. TechCrunch reported the launch on 22 October 2026.
The company trains robots to make and test promising compounds that researchers have already identified. The target is physical wet-lab work: the pipetting, synthesis and assay steps (liquid handling, chemical preparation and measurement) that still run largely on human hands at the bench. Warp Sciences is not positioning its system as a model that reasons about chemistry in the abstract. It is positioning it as a set of robotic systems that carry out the bench procedures themselves.
The training method is the more unusual part of the announcement. According to the same report, human chemists at Warp Sciences wear cameras and sensor gloves while they perform lab tasks. That captured motion and sensor data is used to teach the robots those tasks. In effect, skilled practitioners demonstrate a procedure, and the robots learn it from the recording rather than from hand-written control code.
The company has raised $215 million at an $865 million valuation. Lightspeed led the round, and IVP, Premji Invest and Lux Capital also participated.
The figures place the round among the larger early-stage raises in AI-for-science, though the funding alone does not indicate what the system can do on a bench. Drug discovery has long been limited by how many physical experiments a lab can run. Compound design can be accelerated computationally, but each candidate still has to be made and measured. A system that closes that loop faster, and runs experiments around the clock without fatigue, addresses a real bottleneck in the pipeline.
In my view, the demonstration-based approach deserves closer scrutiny than the headline funding. Teaching by recording experts is a sound way to capture tacit technique, the small judgments about timing, viscosity and glassware that rarely make it into a protocol document. It also means the quality of the robots' behaviour depends on the breadth and consistency of the demonstrations. A procedure captured from one chemist on one day may not transfer cleanly to another instrument, another reagent lot or a changed protocol. Whether the company can generalise from a limited set of recorded tasks to the variety of work a drug discovery lab actually runs will determine the platform's value.
The broader context here is a shift in where AI systems are expected to add value. For much of the current wave, the bottleneck was generating text, code or images. In the lab, the bottleneck is physical action, which brings in hardware reliability, calibration, safety interlocks and the cost of failed runs. A language model handles none of those on its own. Warp Sciences is betting that the data problem can be solved by instrumenting people rather than waiting for large robot-manipulation datasets to accumulate.
If the approach works, the immediate gain is a shorter cycle between a researcher's proposed compound and a measured result. That compression, repeated across many candidates, is what the company is selling. Whether it delivers will show up in bench results and in how quickly discovery programmes move, not in the valuation.
