Stanford CS153 Frontier Systems | Teaching AI to Touch Atoms

Shows how an AI-driven autonomous lab closes the loop between model predictions and physical experiments, and why sample efficiency in RL matters more than benchmark gains for science applications.

• Periodic Labs first planned a purely computational first year, then switched to building smaller, semi-manual labs early, which let the team steer the research program and learn what to scale faster.
• Liam Ferriss's case for sample efficiency: physical experiments cannot be scaled up the way digital rollouts can, so learning from fewer real experiments matters more than climbing benchmarks.
• The team works from a 40,000-square-foot Menlo Park facility where ML researchers sit alongside physicists and chemists; its system, Onnes, is named for the scientist who discovered superconductivity in 1908.
• The founders told students most scientific domains remain largely untouched by LLMs and that the bar for meaningful AI progress in the physical world is still surprisingly low.

• benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.

• videoStanford CME295 Transformers & LLMs | Autumn 2026 | Lecture 1 - Transformers
• videoGuillermo Rauch on Vercel's Shift Toward Agentic Infrastructure
• videoTuhin Srivastava (Baseten) on Production Inference Economics
• videoAnthropic and Chai Discovery on AI-Accelerated Drug Development

• podcastAI researchers debate how close we are to recursive self-improvementDwarkesh Patel
• articleFoundries vs Navigators: Lowering the Cost of ScienceAdrian Sanborn
• articleLila Scienceswww.latent.space

Checking sign-in…

Loading comments…