Lila Sciences is building an automated laboratory infrastructure designed to function like a data center, aiming to achieve scientific superintelligence by generating vast amounts of experimentally validated data. The company treats the lab as an infinite token generator, using AI-guided robotics and vision-language models to run experiments 24/7 across biology, chemistry, drug discovery, and materials science.
- Lila has built a library of over 10 trillion scientifically reasoning tokens that are experimentally validated.
- The infrastructure uses magnetically levitating tracks and orchestration similar to a Slurm queue to connect instruments as nodes on a graph.
- The team optimized for flexibility and generalizability, rebuilding gas sorption measurements to run roughly 2,500x faster.
- Lila claims its general model beats domain-specific models sample for sample by transferring priors from small molecule chemistry to metal organic frameworks.
- The approach enabled six months of in vivo CAR-T data collection in non-human primates using a zero-FTE virtual startup model.
Lila argues that scaling AI through physical experimentation is the path to general scientific reasoning, moving beyond luck-based discovery to automate serendipity and solve complex problems across multiple disciplines simultaneously.