The AI industry is running out of internet-scale data, and Andy Beam of Lila Sciences argues that the next frontier isn’t scraping more text — it’s turning the scientific method into a reinforcement learning loop where the verifier is a physical lab. This reframes a core tension: the internet was a one-time fossil fuel of human-generated tokens, but science can act as an “infinite token generator” because every experiment produces fresh, grounded data. The challenge is that experiments have real runtimes, unlike simulated environments, making the engineering of orchestration and instrumentation the bottleneck rather than model architecture.

Lila’s concrete approach is to build a single generalist model trained on roughly 10 trillion experimentally-verified reasoning tokens spanning biology, chemistry, and materials. The counterintuitive finding is that breadth gives depth — a general model outperforms domain-specific ones because transfer learning across scientific disciplines yields better inductive biases. The lab itself is treated as a data center: instruments are connected on what Beam calls a “PCI bus” abstraction, and humans operate “below the API line,” meaning the system can drive instruments, read results, and iterate autonomously. This has already produced real outputs, including a CAR-T candidate designed in six months by two or three people and “Monster UTRs” hitting roughly 10x the mRNA expression of Moderna/Pfizer benchmarks.

For builders, the key takeaway is that the model is the product, not any single therapeutic or material — Lila is not a biotech company in the traditional sense. The operational insight is Rafa Gómez-Bombarelli‘s “bittersweet lesson”: only the things that you can scale matter, which applies to both data generation and infrastructure. The lab-as-data-center approach introduces real bottlenecks: instrument onboarding is slow, experiments have runtimes, and RL pathologies like collapsed chains of thought or models that “swear” still appear. Serious engineering work now lies in orchestrating heterogeneous instruments, reducing assay latency, and building the equivalent of model FLOPs utilization (MFU) for scientific discovery pipelines.