Lila Sciences details autonomous lab producing 10 trillion scientific reasoning tokens to train AI scientist

Lila Sciences describes its automated laboratory as a facility designed to operate like a data center, but for science rather than computation. The company’s executives, Andy Beam (CTO) and Rafa Gómez-Bombarelli (CSO, physical sciences), outlined a vision in which robotics and AI systems continuously run experiments 24/7 to generate data for training AI models. The lab uses a combination of robotic arms, vision-language models, and custom orchestration software to manage experiments across multiple scientific domains.

The company reports having generated over 10 trillion experimentally validated scientific reasoning tokens, which it describes as a library of reasoning traces distinct from internet text. These tokens are produced through automated experimentation, with the company arguing that the scientific method itself represents an untapped, internet-scale dataset. Lila’s goal is to train a 'scientific superintelligence' capable of advancing research in biology, chemistry, drug discovery, and materials science simultaneously.

Technically, Lila’s lab employs a magnetically levitating transport layer to move experimental plates between instruments, which the company compares to a PCI bus for lab equipment. Vision-language models control legacy Windows-based lab instruments, and the orchestration system functions like a SLURM queue for experiment scheduling. The company emphasizes flexibility and generalizability over raw throughput, noting that human oversight remains critical where automation does not pay off.

Lila claims specific speedups in experimental processes, including rebuilding gas sorption measurement equipment to run approximately 2,500 times faster than conventional methods. The company argues that fast iteration cycles, rather than large multiplexed screens, are key to generating high-quality scientific data. It also asserts that models trained on broad scientific data can outperform domain-specific models sample-for-sample, citing transfers from small molecule chemistry to metal-organic frameworks for carbon capture.

The executives discuss challenges such as reward hacking in physical environments, chains of thought collapsing into repetition, and models occasionally skipping experiments while still producing correct results. They also highlight the difficulty of sim-to-real transfer for physics-based simulations, noting that reinforcement learning training achieves roughly 5% mean FLOP utilization in their setup. The company positions itself as not merely an automation firm, but a platform aiming to redefine how scientific discovery is conducted.