AI Across The Sciences

Author: Shahin Farschi (moderator), Joseph Krause, Carina Hong, Pete Florence Collection: UNLOCK-2026-Sessions

Summary

This wide-ranging panel examines how AI is transforming three non-biological sciences — materials discovery, mathematics, and robotics — and considers what each field's progress reveals about the broader trajectory of AI-for-science. The moderator is Shahin Farschi, partner at Lux Capital (25-year-old VC, 9 funds, 300+ companies). Panelists are Joseph Krause (Co-Founder & CEO, Radical AI), Carina Hong (Founder & CEO, Axiom Math), and Pete Florence (Co-Founder & CEO, Generalist).

Joseph Krause of Radical AI describes the materials science bottleneck: unlike biology (which had the protein data bank built up over 40 years for AlphaFold), materials science has no equivalent repository of experimental characterization data. Theoretical/computational data (from density functional theory, or DFT) can train force fields and interatomic potentials useful for simulation, but simulations are only a starting point for real materials discovery — not the discovery itself. Radical AI built a closed-loop system combining AI generation, robotic self-driving labs, and five characterization modalities (SEM, XRD, XRF, TGA, DSC) plus mechanical property testing (Vickers hardness, ductility, tensile stress-strain). A single PhD metallurgist at Radical AI ran 10 simultaneous discovery campaigns and produced 300 novel high-entropy alloys — compared to ~3,000 such alloys reported in the entire published literature over 50 years.

Carina Hong of Axiom Math (8-month-old company) is building AI for formal mathematics — both automated theorem proving and formal verification of code and chip designs. Her Axiom Prover achieved 100% on the Putnam exam (hardest undergraduate math test) and 99% on the Verina code verification benchmark without modification or tuning between the two tasks. Axiom Prover autonomously proved that the Ramanujan tau function misses 100% of prime numbers (an open number theory result). Hong argues that formal verification — proving code is provably correct — will transform chip design (currently verification teams are 3-4x the size of design teams) and software development in the era of AI-generated code.

Pete Florence co-founded Generalist to build general intelligence for physical world manipulation. He was motivated by his experience as a physics master's student at Cambridge (advisor: Richard Friend, inventor of optoelectronics) where he spent 20 minutes thinking and the rest of his day doing repetitive manual lab work. Generalist released Gen 0 in November 2025 (64% task success rate) and Gen 1 five months later (99%+ success rate) for dexterous manipulation. Gen 1 is trained from scratch on physical interaction data (not fine-tuned from VLAs), enabled by sufficient real-world robot data now existing. One hour of robot data suffices for 99% reliability on trained tasks; Florence predicts this will trend toward 2 demonstrations. The company works with auto manufacturers and logistics customers.