AI for Atoms: How Periodic Labs is Revolutionizing Materials Engineering with Co-Founder Liam Fedus

What It Covers

Liam Fedus, co-creator of ChatGPT and former OpenAI VP of post-training, explains how Periodic Labs builds closed-loop AI systems for materials science, combining specialized neural networks, automated experimentation, and large language model orchestration to accelerate physical world discovery across semiconductors, aerospace, and energy sectors.

Key Questions Answered

• •AI Architecture for Materials: Periodic Labs uses large language models as an orchestration layer that directs specialized symmetry-aware neural networks built specifically for atomic systems. This two-tier structure — general LLM on top, domain-specific models as tools — allows lower latency inference while preserving natural language interfaces for scientists querying experimental data.
• •Closed-Loop Experimental Data: Published literature alone is insufficient for materials AI because reported property values for the same material can span multiple orders of magnitude. Periodic Labs addresses this by running automated experiments that feed back into the model continuously, creating an active discovery loop rather than a static training dataset.
• •Sample Efficiency via Strong Priors: Periodic Labs avoids retraining from scratch by leveraging tens of trillions of tokens from open-source models as a foundational prior. When entering specific chemical spaces, the system reaches useful accuracy with far fewer experiments than a randomly initialized model would require, compressing the data bootstrapping problem significantly.
• •Domain-Specific AGI Timelines: Fedus argues that AI self-improvement is already occurring in software engineering — where unit tests provide cheap, instant verification — but the same loop for physical sciences requires hours of GPU runs and calibrated lab equipment. Builders should expect AI autonomy to arrive domain-by-domain, not as a single general threshold.
• •Capital Structure Mirrors Frontier Labs: Periodic Labs' primary cost is compute, not physical lab infrastructure, which is counterintuitive given the hardware involved. Companies building AI for physical sciences should model their capital requirements closer to frontier LLM labs than to traditional biotech, while also accounting for long lead times on well-calibrated physical systems.