Warp Sciences' Kevin Weil and Sol Reisberg: chemists work by hand first, the work is recorded, and the lab is automated step by step

The physical sciences are short of experimental data, so generate it at scale in one narrow area

Weil says OpenAI's science work started with mathematics because it can be done entirely in silico and pushed with more compute; chemistry, biology and materials move by experiment, and models there are short of data. His conclusion is to generate data at scale, but in one area that matters scientifically and commercially, not as a light dusting across all of science.

“the models are data poor in the physical sciences”Kevin Weil

Lead optimisation is directed evolution in weekly rounds, and contract labs do the bench work

Reisberg describes a small-molecule biotech as starting from a molecule with a hint of activity, making about 50 variants a week, testing them for solubility, toxicity and potency, picking the slightly better one and making 50 around it the next week, climbing toward a drug. He says the physical work is now almost all outsourced to contract research organisations (CROs), and a biotech often never handles the drug until the clinic.

The pain points are intellectual property, supplier concentration and the cost gap

Reisberg says two large providers do most of this work and do it well, fast and cheaply; the problems are depending on one or two suppliers and what he calls multiple documented cases of intellectual property leaking. US contract labs can do it, he says, but labour costs more and their prices typically run four to five times higher, so a biotech has to choose between cost and protecting its IP.

First the hands, then the nervous system, then the brain

Weil says step one is delivering the molecules and tests customers order, by hand at first; chemists will wear egocentric and stereo cameras, pressure-sensing gloves and wrist cameras, so the work produces the data automation needs. The work is then broken into the smallest repeatable chunks, placed on a two-by-two of how often it is done and how hard it is to automate, and automated from the easy corner, from 0% to 10% and 20% and on to 100%. The brain is a synthesis-route model trained on every success and failure.

“First be the hands, then build the nervous system, and then over time become the brain.”Kevin Weil

Synthesis process data can be shared; compound test results must be protected

Reisberg separates making molecules from testing them. Making, he says, is commoditised: given a molecule on paper he has about a 70% chance of making it on the first try and near certainty within ten; how the hands move and which reactions work or fail is the factory's own process data, which customers will not mind sharing. Which molecules are potent or toxic is the customer's core IP; he proposes protecting it with target exclusivity (one customer per area) and pooling data only across non-competing customers, citing Inductive Bio's customer consortia on molecular properties as precedent.

“making molecules is commoditized”Sol Reisberg

A fixed lab environment is the advantage, and the company makes no drugs of its own

Weil says lab chemistry is dextrous and variable (some solids are sticky, some burn in air), but unlike a home robot the hood's lighting and temperature are fully controlled. The methods will differ by task: reverse-engineering instruments so AI can drive them, teaching robot arms some actions, or building a dedicated machine for repetitive work such as weighing solids. He stresses that Warp builds infrastructure for everyone hunting drugs and does no drug discovery of its own.