Latent Space: The AI Engineer Podcast

Synthesis Superintelligence: from Semiconductors to Superconductors — Periodic Labs’ Liam Fedus and Ekin Dogus Cubuk

1h 24min

Periodic Labs is building a synthesis superintelligence by combining AI, high-throughput experimentation, and simulation to accelerate materials discovery. Founders Liam Fedus and Ekin Dogus Cubuk argue that intelligence alone is insufficient—new knowledge requires iterating between conjecture and physical reality. They discuss challenges like noisy experimental data, the limitations of density functional theory, and their vision for automated labs that can dramatically speed up discovery of novel materials, from room-temperature superconductors to better batteries.

Summarized by Podsumo

🎧 Listen 🎙️ Ask about this episode

✨ Key Takeaways

1

Periodic Labs integrates AI agents with robotic labs and simulation tools to close the materials discovery loop, treating experiments as reinforcement learning environments.
2

The team emphasizes the importance of negative results and full experimental lineage data, which are rarely published but critical for training robust AI models.
3

They highlight the fundamental difficulty of characterizing inorganic crystals: XRD patterns are lossy projections, and two different phases can produce similar spectra, requiring multimodal data and chemical priors to disambiguate.
4

Scaling materials discovery requires overcoming bottlenecks like automated characterization and instrument noise; the lab already builds custom hardware and instruments to improve data quality and throughput.
5

Periodic Labs aims to commercialize its tools for the semiconductor industry, starting with on-site deployment engineers who integrate AI and simulation into partners' workflows.

💬 Notable Quotes

"Intelligence is necessary but not sufficient. New knowledge is created when ideas are found to be consistent with reality."

— Liam Fedus
"Science is very hard, but that's what makes it special. Most things that require intelligence are actually more like science—there's auto uncertainty, auto noise, missing context, but you have to be the intelligent being and figure out what to do next."

— Ekin Dogus Cubuk
"We feel like even if GPT-8 gets really good, it will still have to run experiments to get results. Scientific discoveries almost by definition haven't been trained on."

— Liam Fedus