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

What It Covers

Andy Beam (CTO) and Rafa Gómez-Bombarelli (CSO) of Lila Sciences explain how their AI science factory treats physical laboratories as data centers — generating experimentally verified reasoning tokens across biology, chemistry, and materials science to train a general scientific reasoning model, bypassing the exhausted internet-scale pretraining data problem through continuous lab-in-the-loop reinforcement learning.

Key Questions Answered

• •Science as RL Verifier: Post-pretraining AI progress depends on finding new data sources beyond the exhausted internet corpus. Lila's thesis treats physical experiments as the ultimate reinforcement learning verifier — nature itself provides ground truth rewards. This mirrors how math and coding use verifiable rewards, but extends to wet labs and materials synthesis, creating a feedback loop where experimental outcomes directly steer model training toward better scientific reasoning.
• •Lab-as-Data-Center Architecture: Lila connects instruments via planar motor systems where 96-well plates magnetically levitate along tracks — functioning like a PCI bus for biology and materials. Each instrument is a node; physical transport layers are edges. The target end-state is a lights-out, 24/7 facility resembling a data center: multi-story, millions of square feet, maximizing tokens-per-square-foot rather than human ergonomics, with AMRs replacing bench-height instrument assumptions.
• •Flexibility Over Throughput: Unlike traditional lab automation optimized for raw throughput on fixed protocols, Lila prioritizes generalizability — the model can design novel experimental protocols, issue API calls to instruments, and receive feedback even for experiments never previously conceived. Humans remain below the API line for steps where manual execution is faster than automation, such as removing caps, making every action a software-addressable call regardless of executor.
• •Cross-Domain Transfer Produces Real Results: A model trained on small-molecule drug discovery data successfully reasoned over metal-organic frameworks for CO₂ capture — a materials science domain with no explicit training. Similarly, a general scientific reasoning model trained across life sciences, chemistry, and materials consistently outperforms domain-specific models on a sample-for-sample basis, suggesting broad scientific training reduces per-domain data requirements, sometimes to near zero for adjacent problems.
• •10T Experimentally Verified Reasoning Tokens: Lila has assembled roughly 10 trillion reasoning tokens — not nucleotide-level sequences, but quasi-English reasoning traces including tool calls and experimental feedback, generated through RL environments and experimentally verified. Pre-training corpora typically run 15–30 trillion tokens, placing Lila's dataset in the regime where emergent capabilities appear. Starting from open-weight models like Nematron avoids the ~$1B compute cost of pretraining from scratch.
• •Zero-FTE Virtual Startup Model: Lila's commercial structure enables two-to-three-person teams to compress five years of biotech R&D into six months at roughly 10% of traditional investment. A proof case: an internal team of two to three people developed in-vivo CAR-T mRNA constructs with 10x expression improvement over Moderna/Pfizer references, reaching non-human primate B-cell depletion data surpassing the Capstan dataset — the asset that sold to AbbVie for $2.1B — within six months via platform access fees plus milestone-based upside sharing.
• •Proxy Measurement Acceleration: Iteration speed, not parallelization breadth, is Lila's primary scaling lever. For gas sorption in metal-organic frameworks, standard BET pressure measurement takes roughly one day per sample. Lila built a parallel proxy measurement instrument achieving equivalent readouts across 96-well plates in approximately one hour — a roughly 2,500x throughput improvement. Identifying faster proxy measurements that correlate with target properties is a repeatable strategy for compressing experimental cycle times across materials domains.