Latent SpaceJune 17, 20261h 16m

🔬 The Limits of AI in Science - Why We Need Self-Driving Labs — Joseph Krause, Radical AI

TL;DR

Materials AI breaks where biology-style representations stop: Krause says strings like SMILES work for molecules because bonds and elements capture much of the problem, but alloys also depend on microstructure, processing route, supply chain, cost, and manufacturability, which cannot be compressed into a simple text representation.
Radical's moat is experimental throughput, not model cleverness: The company has produced about 1,200 alloys in 5 to 6 months, with 300 novel to literature and roughly 10 that look especially promising, while aiming to scale from 8 to 20 alloys per day today to 100 per day by June or July.
The company is building a real self-driving lab, not just lab automation: Krause draws a Waymo analogy, saying an automated lab is like lane-keeping assistance while a self-driving lab runs full research campaigns, decides what to test next, and routes samples through synthesis, characterization, and property testing with minimal human intervention.
Humans still matter because scientific intuition is hard to encode: Metallurgists at Radical annotate SEM images with observations like dendritic formation and still handle some synthesis steps, because the company is trying to capture tacit expertise such as how to read micrographs or judge when a molten alloy is fully melted.
Qualification and manufacturing remain the long pole: Even if discovery gets faster, aerospace alloys still face around 10 years of qualification through FAA or military standards, and Krause is explicit that Radical has not solved manufacturing-scale data capture yet.
Open source fits their thesis because experiments are the moat: Radical publishes models like Matrix and open tools because Krause believes science models will broadly commoditize, while proprietary advantage will come from the self-driving lab, the experimental loop, and the data generated inside it.

The Breakdown

Radical AI has already made 1,200 alloys in a few months, including 300 that were never reported in literature, and Joseph Krause argues that this only matters because the real bottleneck in materials is not model intelligence but experimental ground truth. His core case is blunt: there is no AlphaFold-style one-shot for materials, so self-driving labs are the only serious path to getting new alloys from composition to real-world products.