Latent Space: The AI Engineer Podcast

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

1h 41min

Lila Sciences is building 'AI science factories' that treat laboratories as data centers, using automated experiments to generate massive reasoning tokens for training scientific AI models. By combining reinforcement learning with physical verification (lab experiments as tool calls), they aim to create a general-purpose scientific reasoning model that spans biology, chemistry, and materials science—moving beyond domain-specific AI toward an open-ended scientific superintelligence.

Summarized by Podsumo

🎧 Listen 🎙️ Ask about this episode

✨ Key Takeaways

1

Lila Sciences views the lab as an 'infinite token generator'—automated experiments produce verified reasoning traces (10 trillion tokens so far) that are fed back into the model via RL, enabling post-training at scale.
2

The company is building a flexible 'PCI bus' for lab instruments, using a magnetic levitation transport system to connect dozens of instruments (liquid handlers, spin coaters, magnetron sputterers) into a programmable experimental graph.
3

They demonstrated a CAR T therapy proof-of-concept in six months with a small team, achieving 10x better mRNA expression than Pfizer/Moderna references—a task that would typically take five years and $100M+ in a biotech.
4

The platform supports both biology (DNA, RNA, proteins) and materials (quantum dots, thin films, electrocatalysts, MOFs), with surprising cross-domain transfer—for example, small-molecule drug discovery knowledge transferring to metal-organic framework design.
5

Safety is taken seriously from the start: AI safety team, constrained experimental nodes per problem, and a human-in-the-loop for approval—but model suggestions are already surprising experts (e.g., non-platinum group electrocatalysts that became the best performers).

💬 Notable Quotes

"Silicon: We think the lab of the future should feel like a data center. Rows of server racks as densely packed as possible and also as energy efficient as possible."
"Andy Beam: The experiment has a runtime. We think about Lila as being able to generate different kinds of data on different link scales. The infinite token generator is still there."
"Rafa Gómez-Bombarelli: When the model gives 'stupid' suggestions, those often turn out to be our best-performing chemicals. The line between obviously wrong and quasi-move 37 is hard to know."