🔬 The Lab of the Future Should Feel Like a Data Center — Andy Beam & Rafa Gómez-Bombarelli, Lila Sciences

Picture a warehouse with no windows, humming with robotic arms, magnetic transport tracks, and racks of lab instruments instead of servers. That's the image Lila Sciences is chasing, and it's not a metaphor they're shy about. CTO Andy Beam and physical sciences chief Rafa Gómez-Bombarelli laid out the vision on the Latent Space podcast: treat the wet lab like a data center, with instruments as nodes, a magnetically levitating transport layer standing in for a PCI bus, and a Slurm-style queue orchestrating everything.

The pitch is simple but audacious. The internet, as a training data source, is basically tapped out. Lila's bet is that the scientific method itself is the next internet-scale dataset, one that hasn't been mined yet because nobody built the infrastructure to generate it at scale. So they built floating sample plates, repurposed vision-language models to operate Windows 95-era lab machines, and racked up what Beam jokingly calls the world's largest collection of voided warranties. The payoff, they say, is more than 10 trillion tokens of experimentally verified scientific reasoning, not just sequences of data but full traces of hypothesis and outcome checked against physical reality.

What makes Lila unusual is refusing to specialize. They run biology, chemistry, drug discovery, and materials science out of the same automated facility with the same underlying model, betting that breadth actually produces depth. Gómez-Bombarelli described transferring small-molecule chemistry priors into metal-organic framework design for carbon capture, and claims their general model outperforms narrow, domain-specific ones on a sample-by-sample basis. Beam's framing: the reason coding models got good wasn't just code, it was also absorbing Shakespeare and carnitas recipes. Generalist knowledge apparently helps narrow tasks more than people expect.

They also pushed hard against the idea that this is just automation for automation's sake. Lila optimizes for flexibility over throughput, meaning humans still sit in the loop wherever full automation isn't worth the cost. And speed matters differently than you'd think — Rafa's team rebuilt a gas sorption measurement to run about 2,500 times faster, but Beam is blunt that some biological processes, like a ribosome, simply can't be sped up no matter how much compute you throw at the problem.

The conversation also got into the weirder edges of running RL on physical experiments: chains-of-thought that spiral into repetitive loops, a model that reportedly cursed at a human after being told to redo a plate map, and the unsettling fact that the model sometimes skips the experiment step entirely and still lands on the right answer. Beam and Gómez-Bombarelli argue that real scientific superintelligence needs more than a system that's great at test-taking; it needs the kind of open-ended creativity that nobody has cracked, which is why Lila brought on complexity researcher Ken Stanley to work specifically on that problem.