Lila Sciences on the Path to Scientific Superintelligence
Latent Space · youtube · 2026-07-16
This interview features Lila Sciences discussing their roadmap for "scientific superintelligence": using wet labs as verifiers and leveraging reinforcement learning to turn scientific experiments into a scalable data generation mechanism.
Key takeaways from the interview include:
- The internet is no longer an infinitely expandable data source; the next generation of massive tokens will come from verifiable experimental results.
- Lila believes models should use a unified, general-purpose foundation across biology, chemistry, and materials science, rather than building single-discipline models.
- They treat the lab as a "data center," with instruments functioning like humans working below the API line on a PCI bus.
- The guest mentions a case where a CAR-T candidate was designed by two or three people in six months, alongside the concept of a "zero-FTE startup."
The second half of the show covers reward hacking in RL, chain-of-thought collapse, why materials science still lacks an AlphaFold-level breakthrough, and the bottlenecks in lab automation and orchestration.
Related event: Lila Sciences Discusses Path to Scientific Superintelligence(3 posts)→
More from Companies & People
- OpenAI appears to open an “Advertise in ChatGPT” page — montecarl · 2026-07-22
- OpenAI's Rough Patch: GPT-5.6 Data Wipes, Sandbox Escapes, and Apple Lawsuit — Annual_Judge_7272 · 2026-07-22
- Apple is now competing directly for creativity app users — speckx · 2026-07-22
- FactoryAI gave back its first millions, then shipped Droid CLI two years later — matanSF · 2026-07-22
- A post says AI teams should drop the research scientist vs engineer split — jsuarez · 2026-07-22
- YC startup Vendo launches an open-source customization layer for user-built micro-apps — ycombinator · 2026-07-22