🔬Periodic Labs: AI Scientists Need Real Labs, Not Just DFT
TL;DR
Periodic Labs' Liam Fedus and Ekin Dogus Cubuk say intelligence alone cannot produce new knowledge, because discovery needs experiments. Their RL environments come from physical labs, where noisy results are treated as ground truth.
Periodic Labs' Liam Fedus and Ekin Dogus Cubuk say intelligence alone cannot produce new knowledge, because discovery needs experiments. Their RL environments come from physical labs, where noisy results are treated as ground truth. The episode runs about 1 hour 24 minutes.
Key Points
The materials loop covers choosing what to make, synthesizing it, and characterizing it, mainly with X-ray diffraction
DFT simulations cannot yet capture strong electron correlation or high-temperature superconductivity
Failed experiments count as valuable training data
Superconductivity measurements take about an hour each, a throughput bottleneck
Fedus: "full autonomy is a non-goal"
Why It Matters
It separates AI-for-science hype from the real bottleneck, which is lab throughput and noisy data. Expect hardware and wet-lab capacity to decide who wins.
Quick Facts
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 3,564 builders reading daily.