No PriorsAI for Atoms: How Periodic Labs is Revolutionizing Materials Engineering with Co-Founder Liam Fedus
CHAPTERS
Why AI Needs the Physical World: Periodic Labs’ “Foundation Model for Atoms”
Elad Gil introduces Liam Fedus and frames Periodic Labs’ core thesis: AI won’t meaningfully accelerate science without being connected to real experiments and physical reality. The conversation sets up materials engineering as the next major frontier beyond language and code.
- •Liam Fedus’ role in building ChatGPT and leading post-training at OpenAI
- •Periodic Labs as an “AI foundation lab for atoms” focused on materials/chemistry
- •Motivation: scientific progress requires experiments, not just thinking or text generation
- •Framing the shift from digital-only AI to systems that affect the physical world
From Physics to AI: Dark Matter Research and Why Physicists Flood Into ML
Liam recounts his physics background, including dark matter research, and discusses why physics training maps well to modern AI work. They connect this to a broader migration of high-energy physicists into machine learning after major milestones like the Higgs discovery.
- •Undergrad physics and early research on directionally sensitive dark matter detection
- •Physics as a disciplined way to reason about complex systems
- •High leverage and rapid progress in AI attracted physicists
- •Post-Higgs “what’s next?” dynamic in high-energy physics pushing talent toward AI
Google Brain in 2016–2017: The Cambrian Era of Scaling, Sparsity, and Transformers
Liam describes joining Google Brain during a formative period when key ideas like distributed training and early scaling approaches were emerging. He highlights the research culture of small teams pushing boundaries with limited hardware compared to today’s industrial-scale AI labs.
- •Google Brain’s early, small-team frontier environment
- •Era of foundational ideas: distributed training, mixture-of-experts, transformers
- •Liam’s focus on architecture and sparsity for efficient scaling/serving
- •Contrast between early experimentation and today’s capital-driven scaling labs
OpenAI and the Road to ChatGPT: Turning GPT-4 Into a Product
Liam explains the practical challenge of productizing GPT-4 and how internal debates shaped the initial product direction. He notes that the “general chatbot” approach won out over narrower ideas, catalyzing mainstream awareness of AI capabilities.
- •Objective: productionization and product strategy around GPT-4
- •Early product brainstorming (writing bot, coding bot, meeting notes)
- •John Schulman’s push to keep it general → chatbot as core product form
- •ChatGPT as a public inflection point for AI adoption and awareness
From ChatGPT to Atoms: Why Materials Needed Better Reasoning and Tool Use
Liam outlines why 2022-era models were insufficient for serious physical-world science and why subsequent improvements changed the feasibility. He emphasizes reasoning, test-time inference, error correction, and tool use as prerequisites for closed-loop experimentation.
- •Physical-world science requires interacting with reality, not just text generation
- •2022 models were “too weak” for robust experimental orchestration
- •Improvements that matter: reasoning, test-time inference, reliable tool use
- •Rise of agents (especially coding agents) as enabling infrastructure for labs
The Data Problem in Science: Simulation vs. Experiment and the Need for Ground Truth
They unpack how data scarcity and noisiness in materials science differs from internet-scale language data. Liam describes the limits of literature-derived values and why experiments are required to ground models in reality.
- •Two primary sources: physics simulations and experimental measurements
- •Foundation models inherit strong priors from internet/papers—but that’s not enough
- •Literature values can vary wildly (orders of magnitude), creating unreliable targets
- •Experimental grounding provides a path toward “truth” rather than just modeling noise
Closed-Loop Discovery: Turning Experiments Into an Active Learning Engine
Liam argues the breakthrough is not just collecting a dataset but building an interactive loop that decides what to test next. The system uses experimental results to find inconsistencies, reconcile sources, and guide subsequent experiments for faster discovery.
- •Closed-loop system beats static datasets for scientific progress
- •Use experimental data to detect aberrations and patterns
- •Cross-checking experiment vs. simulation vs. literature to improve reliability
- •Iterative experiment selection as the core accelerator for materials discovery
Generalization Boundaries: When Quantum-Driven Models Transfer—and When They Don’t
The conversation explores how far learned representations can generalize across scientific domains. Liam notes meaningful transfer within quantum-mechanical regimes, but limited transfer to domains governed by different abstractions like fluid dynamics.
- •Best progress occurs where there’s abundant domain-specific data
- •Transfer works better in areas governed by similar quantum-mechanical principles
- •Generalization breaks across fundamentally different physics regimes (e.g., fluids)
- •First-principles structure provides partial reusability but not universal coverage
System Architecture: LLMs as Orchestration Layer + Specialized Atomic Models
Liam describes Periodic’s architecture: language models provide a natural interface and planning layer, while specialized neural nets handle atomic symmetry and fast, low-latency predictions. The overall system routes between literature, experimental data, and domain tools.
- •LLMs used as the “orchestration” and interaction layer for scientists
- •Specialized models for atomic systems with symmetry awareness
- •Tool-like use of specialized nets (as tools/reward functions) inside a larger agentic system
- •Multi-modal ingestion: literature + internal experimental data + model predictions
Commercialization Strategy: Intelligence Layer for Advanced Manufacturing
Elad presses on go-to-market: why materials isn’t as universally “plug-and-play” as language. Liam positions Periodic initially as a software/intelligence layer for companies bottlenecked by materials and process engineering, with potential to evolve toward a discovery-style value capture model.
- •Start with “customer zero” internally: transforming how Periodic’s science is done
- •Target customers: enterprises constrained by materials/process engineering
- •Product framing: control plane/system-of-record for experiments and solutions
- •Business model: begin as software; discovery/royalty model is possible longer-term
If Periodic Works: “Generating Matter,” Faster Physical Progress, and a New Industrial Revolution
Using sci-fi as a prompt (The Diamond Age), they discuss what a successful “atomic rearrangement” platform could mean for industries and daily life. Liam predicts meaningful speedups despite physical constraints, creating a step-change in pace for semiconductors, energy, aerospace, and manufacturing.
- •Vision shift: from generating text/software to generating physical matter
- •Impact areas: semiconductors, aerospace, energy, advanced manufacturing
- •Even with physics limits, order-of-magnitude speedups are plausible
- •Goal: give humanity agency over synthesis and materials design at scale
Scaling the Lab: Multidisciplinary Teams, Capital/Compute, Hiring, and Robotics as the Next Interface
Liam highlights multidisciplinary collaboration as essential, then details practical constraints: compute is expensive and often dominates cost, while physical automation has lead times and reliability challenges. They close on hiring needs and why better robotics would dramatically accelerate closed-loop experimentation and broader physical-world AI.
- •Multidisciplinary core: chemists/physicists + AI researchers + engineers
- •Analogy to AI scaling: moving from small-team science to industrial-scale throughput
- •Capital intensity: GPUs/compute as a primary cost; physical systems add lead-time complexity
- •Hiring focus split into “bits and atoms” roles (ML training/infrastructure + control/system engineering + product)
- •Robotics: not strictly required, but a major accelerator for scalable, flexible lab automation
