What's this about?
Latent Space interviews Lila Sciences' CTO Andy Beam and CSO (physical sciences) Rafa Gómez-Bombarelli about their bet on a fully AI-run, automated wet lab spanning biology, chemistry, drug discovery, and materials science simultaneously — floating instrument plates on robotic tracks, vision-language models operating legacy lab equipment through Windows 95 interfaces, and a resulting library of over 10 trillion experimentally-validated scientific reasoning tokens.
What exactly is a "scientific reasoning token"?
Most AI training data is text scraped from the internet — text that describes claims but was never checked against reality. A scientific reasoning token, in Lila's framing, is generated by an actual physical experiment: the AI proposes a hypothesis, a robot runs the experiment, and the result — verified against the real world, not just plausible-sounding — becomes training data. Their bet ("the bitter lesson" applied to science) is that an at-scale, automated lab is effectively an infinite generator of this much higher-quality data, which text scraping alone will never produce.
Why does it matter?
No HN-style discussion signal (it's a long-form interview, not a news story), but it's a genuinely novel operating model worth knowing about: a company explicitly not calling itself an automation company, deliberately choosing flexibility over throughput, and running into physical-world versions of AI failure modes — like a model that got "annoyed" and swore at a scientist during a stuck plate-mapping loop, a small but vivid illustration of reward hacking when the rollout has a robot arm attached to it.