No PriorsAI for Atoms: How Periodic Labs is Revolutionizing Materials Engineering with Co-Founder Liam Fedus
CHAPTERS
- 0:05 – 1:05
Why “AI for atoms” matters: foundation models meet the physical world
Elad sets the stage: Liam Fedus helped create ChatGPT and is now building Periodic Labs—an “AI foundation lab for atoms.” The conversation frames the core challenge: translating AI progress from text and code into materials, chemistry, and real-world experimentation.
- •Liam’s role in building ChatGPT and leading post-training at OpenAI
- •Periodic Labs’ mission: apply AI to materials/chemistry and the physical world
- •Why this is a different kind of foundation-model problem than language
- 1:05 – 2:58
From physics to AI: why so many physicists end up in ML
Liam traces his origin as a physics student and researcher, and Elad asks why physics backgrounds are so common in frontier AI. They discuss the “principled” training physics provides and how shifts in high-energy physics pushed talent toward ML’s higher leverage.
- •Undergrad physics and dark matter research as Liam’s starting point
- •Physics training: rigor, principled thinking, careful scientific method
- •Post-Higgs high-energy physics bottlenecks nudging researchers toward AI
- 2:58 – 3:46
Early Google Brain era: transformers, MoE, and scaling with small teams
Liam describes joining Google Brain during a formative period (2016–2017) when core ideas like distributed training and transformer-era innovations were emerging. He emphasizes how much frontier progress once came from small groups and limited compute.
- •Google Brain in 2016–2017 as a “Cambrian era” for ML research
- •Key innovations: distributed training strategies, sparsity, mixture-of-experts
- •Contrast between early diversity/entropy and today’s industrialized scaling
- 3:46 – 5:26
At OpenAI: turning GPT-4 into products and the birth of ChatGPT
Liam explains the push to productionize GPT-4 and the internal debate over what the first killer product should be. John Schulman’s preference for a general chatbot helped crystallize the direction that became ChatGPT.
- •The challenge: productizing GPT-4 beyond research demos
- •Early product ideas (writing, coding, meeting bots) vs. a general chatbot
- •ChatGPT as the “starting gun” for broad public awareness of LLMs
- 5:26 – 6:46
Why materials now: connecting AI to reality requires experiments
Liam argues that major acceleration in science and technology requires AI systems to interface with the physical world via experimentation. He notes that 2022-era models were not capable enough, but improvements in reasoning and tool use made the connection increasingly feasible.
- •Science progress depends on experiments, not just “thinking in a room”
- •2022 models were too weak; later reasoning/tool-use advances changed that
- •Agents and test-time inference as prerequisites for real-world closed loops
- 6:46 – 7:44
Data in the physical world: simulation vs. experiment and the need for grounding
The discussion turns to the data bottleneck: unlike the internet for language models, physical-world data is sparse, noisy, and inconsistent. Liam highlights why experimental grounding and closed-loop iteration are essential to move beyond unreliable literature aggregates.
- •Two sources: physics simulation data and real experimental data
- •Literature-derived properties can vary by orders of magnitude
- •Closed-loop experimentation: detect anomalies, cross-check, design next experiments
- 7:44 – 10:39
Sample efficiency and priors: leveraging internet-scale knowledge without starting from scratch
Liam explains that Periodic benefits from strong priors inherited from general-purpose models trained on massive corpora. That allows higher sample efficiency when moving into specific chemical or materials discovery domains, while focusing internal ML effort on where frontier capability is lacking.
- •Periodic leverages existing LLM capabilities (e.g., coding) rather than reinventing them
- •Combination of open and closed models; zero effort spent improving coding models
- •General text/paper knowledge provides priors, but is insufficient without experiments
- 10:39 – 11:39
What generalizes and what doesn’t: domains, first principles, and limits of transfer
They explore how generalization works across physical domains: quantum-mechanical regimes may share transferable structure, but that transfer doesn’t automatically help with areas like fluid dynamics. The takeaway is a layered view of generalization tied to governing physics.
- •Best progress where there’s abundant domain data
- •Quantum-governed systems can share generalizable structure
- •Transfer breaks across different physical abstractions (e.g., quantum vs. fluids)
- 11:39 – 12:47
System architecture: LLMs as orchestration layer plus specialized atomic models
Liam outlines a hybrid architecture: language models provide a natural interface and orchestration layer, while specialized low-latency, symmetry-aware neural nets handle atomic-scale predictions. The system treats these specialized models as tools and components inside a broader experimental workflow.
- •LLMs used for orchestration, planning, and interfacing with literature/data
- •Specialized atomic models incorporate symmetry and run with lower latency
- •Tool-like composition: specialized nets act as tools/reward functions within a system
- 12:47 – 16:10
Commercialization strategy: “intelligence layer” for industrial science and manufacturing
Elad probes how Periodic will bring this to market given the smaller “footprint” of materials vs. language. Liam describes starting with scientists as “customer zero,” then expanding to enterprises bottlenecked by materials and process engineering—positioning Periodic as a control plane for experimentation and decision-making.
- •Start internally: transform how Periodic’s own scientists work
- •Target customers: industries constrained by materials/process engineering
- •Business framing: software-first intelligence/control layer for experiments
- 16:10 – 17:48
If Periodic succeeds: accelerating ‘matter generation’ across semiconductors, energy, aerospace
Elad asks for a 10-year vision; Liam describes moving from systems that generate text/software to systems that help generate and rearrange matter. He predicts substantial speedups (even if physics imposes limits) by improving iteration cycles and extracting insight from complex experimental data.
- •Shift from digital outputs to atomic rearrangement and synthesis workflows
- •High-impact sectors: semiconductors, aerospace, energy
- •Even 10–100× iteration improvements could make life feel dramatically different
- 17:48 – 19:40
Scaling in the physical sciences: multidisciplinary teams, automation throughput, and bottlenecks
Liam emphasizes that progress requires tight collaboration among physicists, chemists, AI researchers, and engineers. He draws an analogy to AI scaling laws and argues physical sciences will similarly industrialize—where automation increases experimental throughput but also creates new “intelligence bottlenecks” in interpreting data.
- •Irreducibly multidisciplinary collaboration as the core organizational advantage
- •Analogy to ML’s shift from small teams to scaling-law-driven industrialization
- •Automation increases throughput; scientists can become bottlenecked on interpretation
- 19:40 – 21:42
Capital, compute, and hiring: building ‘bits and atoms’ capabilities
They discuss the cost structure: GPUs and compute dominate, while physical infrastructure has long lead times and calibration complexity. Liam also shares how Periodic hires across “bits” (ML training/infrastructure) and “atoms” (control systems, lab engineering) plus product engineering to bridge them.
- •Compute/GPU expense as the primary capital driver
- •Physical infrastructure: lower cost sometimes, but harder logistics and reliability demands
- •Hiring taxonomy: ‘bits’ (ML, infra) and ‘atoms’ (control/system engineering) plus product
- 21:42 – 25:39
AGI, spiky intelligence, and machine self-improvement timelines
Liam argues intelligence isn’t a single scalar: systems can be world-class in narrow areas yet brittle under small perturbations. He predicts machine-driven improvement will arrive first in domains with cheap verification loops (software engineering, then AI research), and later in physical sciences via closed-loop experimentation.
- •“Spiky” capability profiles and non-intuitive generalization failures
- •Recursive self-improvement looks like domain-specific architecture search
- •Fastest self-improvement where evaluation is cheap (unit tests); slower where GPUs/experiments are costly
- 25:39 – 27:59
Robotics and lab automation: not required, but a major accelerator for closed-loop science
Elad asks whether general robotics is necessary for Periodic’s closed loop; Liam says it’s not strictly required but would dramatically speed data generation and lab scaling. They contrast today’s careful, bespoke lab automation with the potential of reliable, general-purpose robots operating in unstructured labs.
- •Goal: high-quantity, high-quality, diverse experimental data
- •Current approach: hybrid of humans + reliable autonomous subsystems + off-the-shelf robotics
- •Dexterous general robots would accelerate new lab spin-up and throughput
- 27:59 – 29:25
Beyond Periodic: why robotics is the next major interface layer (and closing reflections)
Liam names robotics as the most exciting adjacent frontier because it expands AI’s agency into the physical world and addresses widespread labor shortages. They close by reinforcing the theme: interface layers that connect models to reality will define the next decade.
- •Robotics as the transformative bridge from models to physical action
- •Physical-world workforces dwarf software engineering—large untapped impact
- •Decade outlook: broader agency/control in the real world drives the next wave
