Lila Sciences builds AI-run lab with 10T validated reasoning tokens

Lila Sciences operates a darkened warehouse filled with AI-guided robotics, instruments, and floating plates that zip along magnetically levitating tracks, running experiments continuously. The facility is the physical embodiment of a thesis: that science, treated as an infinite token generator, combined with reinforcement learning and scaled deployment, will yield what the founders call scientific superintelligence.

The lab has produced over 10 trillion experimentally validated scientific reasoning tokens—a category of data the company argues barely exists on the public internet. These are not text sequences but experimentally verified reasoning traces: records of AI reasoning paired with wet-lab results. Lila's CTO Andy Beam and CSO of physical sciences Rafa Gómez-Bombarelli argue that this form of grounded data is essential because, unlike internet text, it has been verified by nature itself. The lab operates across biology, chemistry, drug discovery, and materials science simultaneously, using the same AI and the same physical space. Lila insists it is not an automation company; it optimizes for flexibility and generalizability over raw throughput, meaning humans remain involved wherever automation does not pay.

Early results include a model-suggested catalyst for platinum-group-free electrocatalysts that a domain expert with 40 papers of background knowledge initially called "stupid" but later recognized as the best-performing design the lab has made. The lab also achieved in vivo CAR-T validation in nonhuman primates within six months—a milestone especially notable because AbbVie paid $2.1 billion（約3400億円） for Capstan on the strength of preclinical in vivo CAR-T data. Lila rebuilt a gas sorption measurement to run roughly 2,500x faster by prioritizing fast round-over-round iteration rather than large multiplexed screens. The company's strategy assumes that breadth—training on multiple domains together—creates transfer learning that beats domain-specific models sample for sample.

Key technical questions remain open. Models sometimes skip experiments entirely and remain correct, raising the question of how much to trust latent reasoning versus the experimental verifier. Reward hacking in a physical loop is also a hazard: chains of thought can collapse into repetition, and models have been observed becoming frustrated when asked to redo plate maps. Rafa inverts the bitter lesson for materials science: in AI, scaling is a roadmap, but in materials, scaling is a filter—only things that actually scale in the physical world end up mattering. The company's broader ambition is scientific superintelligence wired directly into the wet lab, driven by the hypothesis that if you have infinite experimentally verified data, you can build a general reasoner for any scientific problem.

Lila Sciences is betting on a hypothesis grounded in what the field calls the bitter lesson—the observation that scaling and brute-force learning often outperform hand-crafted domain knowledge. Applied to science, this thesis holds that a physical lab run by AI is an infinite generator of experimentally validated data, and that training on such data at scale will yield models that can reason across any scientific domain. The founders argue that breadth—training on chemistry, biology, materials, and drug discovery together—creates transfer that beats domain-specific models sample for sample. This flies against the conventional biotech playbook of single-asset focus.

The concrete evidence Lila offers includes a model's suggestion for platinum-group-free electrocatalysts that a 40-paper expert deemed "stupid" until they worked and became the best performers the lab has made. Similarly, the six-month path to in vivo CAR-T validation in nonhuman primates is notable because AbbVie's $2.1 billion（約3400億円） acquisition of Capstan hinged on the same milestone. Lila's architecture—instruments as graph nodes, a magnetically levitating transport layer between them, reinforcement learning with nature as the verifier—positions the lab as a data center, not a biotech facility. Rafa Gómez-Bombarelli inverts the bitter lesson for materials science: in AI, scaling is a roadmap; in materials, scaling is a filter, because only things that scale matter in the physical world. This framing suggests Lila believes the bottleneck is not discovery but manufacturability.

No comments yet. Be the first to share your thoughts!