Skip to content
Stanford CS153 Frontier Systems | Teaching AI to Touch Atoms
This video isn’t embeddableWatch on YouTube →
Stanford OnlineStanford Online

Stanford CS153 Frontier Systems | Teaching AI to Touch Atoms

For more information about Stanford's online Artificial Intelligence programs, visit: https://stanford.io/ai Follow along with the course schedule and syllabus; visit: https://cs153.stanford.edu/ In this CS 153 Frontier Systems lecture, Periodic Labs cofounders Liam Ferriss and Dorje Chubak — veterans of OpenAI's post-training team and DeepMind's Genome project, respectively — walked students through their eleven-month-old mission to apply AI to the physical world, specifically using autonomous labs to accelerate the discovery of new materials including high-temperature superconductors. Operating out of a 40,000-square-foot Menlo Park facility where machine learning researchers work alongside physicists and chemists, their AI system (named Onnes, after the scientist who discovered superconductivity in 1908) runs a continuous loop of computational prediction, robotic synthesis, and experimental verification — closing the feedback cycle between digital intelligence and physical reality in a way no purely in-silico approach can. A key early lesson was abandoning their original plan of spending the first year purely computational: building smaller, semi-manual labs first allowed them to direct the research program and understand what to scale, faster. Dorje emphasized that the results of applying LLMs to actual atoms have exceeded even their own expectations, while Liam stressed that sample efficiency in reinforcement learning — not benchmark climbing — is the core technical frontier when physical experiments can't be arbitrarily scaled up the way digital rollouts can. They closed by pushing back on student anxiety about AGI displacing their careers, arguing that most scientific domains remain largely untouched by LLMs, that the bar to make meaningful AI-physical world progress is still surprisingly low, and that the history of civilization is essentially a materials story — making their work, in their view, among the highest-leverage bets anyone can make right now.

Dorje Chubakguest
Sep 29, 202642mWatch on YouTube ↗

CHAPTERS

  1. 0:10 – 2:58

    Periodic Labs: founders’ backgrounds and the biggest 11-month reality check

    The host introduces Liam Ferriss and Dorje Chubak and asks what assumptions changed most since founding Periodic Labs. They explain that the path to an autonomous, high-throughput lab required earlier hands-on experimentation than expected, and that the impact of “LLMs touching atoms” has exceeded their expectations.

    • •Company origin story and timeline (~11 months since founding)
    • •Big prior update: you can’t stay purely in silico for year one—early physical loops matter
    • •Building smaller semi-manual/semi-autonomous labs sped up learning and iteration
    • •Surprise: automated intelligence applied to matter yields even ‘crazier’ results than anticipated
  2. 2:58 – 5:46

    The end-to-end “AI ↔ robots ↔ materials” closed loop in Menlo Park

    They outline Periodic’s pipeline: an AI proposes materials, robots synthesize them, instruments characterize them, and results feed back into training. A key nuance is that AI is valuable not only for big scientific leaps but also for constant mundane error-correction that makes real labs work reliably.

    • •40,000 sq ft facility; interdisciplinary team (ML + experimental scientists)
    • •AI proposes candidate materials (motivated by superconductors)
    • •Robotic synthesis → measurement/verification → feedback into the model loop
    • •AI helps with operational minutiae: powder mixing, impurity control, detecting sample mix-ups
    • •Characterization challenges (e.g., interpreting X-ray diffraction) are core subproblems
  3. 5:46 – 8:22

    Why an “eval-first” mindset is harder in science (and where semiconductors fit)

    The discussion connects to earlier class content: start with a concrete eval, but science is inherently messy and partially observed. They argue you must first understand the “shape” of the scientific decision problem, then choose domains (like semiconductors) where bottlenecks are real and progress is measurable.

    • •Science lacks clean ground truth; decision-making under uncertainty dominates
    • •Blindly hill-climbing a fixed eval can fail without domain shaping
    • •Semiconductors: massive demand, shrinking nodes, materials bottlenecks across logic/memory
    • •Motivation: “computation is physical,” so improving materials accelerates compute progress
  4. 8:22 – 9:54

    Superconductors vs semiconductors: choosing the atom–electron interaction frontier

    Dorje frames a thesis: AGI may not generalize magically, so you should push AI toward a strategically important physical domain—electron/atom interactions. Superconductivity and semiconductor interface physics share underlying mechanisms, making the research program mutually reinforcing.

    • •Hypothesis: general intelligence won’t automatically cover all scientific domains
    • •Strategic focus: electrons/phonons/magnetism regimes relevant to tech and superconductivity
    • •Not targeting unrelated domains (e.g., polymers) as a near-term generalization claim
    • •Learning in superconductivity improves broader synthesizability and thermodynamic intuition
  5. 9:54 – 14:14

    Superconductivity 101 by analogy: resistance, heat loss, and why discovery matters

    They give a quick conceptual explanation of superconductors using a “pipe/flow” analogy and contrast ordinary resistive losses with zero-resistance superconducting behavior. They connect improved superconductors to data-center efficiency, lossless transmission, fusion magnets, maglev, quantum computing, and more.

    • •Resistance as electron scattering → energy lost as heat
    • •Superconductors: zero resistivity observed after helium liquefaction (early 1900s)
    • •Modern relevance: efficiency and heat dissipation constraints in computing
    • •Applications: MRIs (today), plus future: fusion, quantum computing, maglev, grid transmission
    • •Materials understanding is broadly enabling across aerospace/auto/energy/semis
  6. 14:14 – 15:01

    Meet ‘Onnes’: naming the agent and the industrial-scale science philosophy

    They introduce the Onnes system, named after Heike Kamerlingh Onnes—discoverer of superconductivity and an early advocate of industrial-scale research labs. The name reflects Periodic’s belief that scientific progress can be systematically scaled with engineers, technicians, automation, and urgency.

    • •Agent name: Onnes (O-N-N-E-S), not “honest”
    • •Historical inspiration: liquefying helium and discovering superconductivity
    • •Operational inspiration: industrializing scientific research methods
    • •Core ethos: scale, seriousness, and urgency in experimental science
  7. 15:01 – 16:50

    Onnes vs ChatGPT-style systems: grounding intelligence in experimental reality

    Liam contrasts Periodic’s approach with general-purpose LLM development: the goal is deeper physical-world competence, not better coding. They emphasize that textbook knowledge alone won’t yield novel materials; models must be trained with new experimental data and iterative contact with reality.

    • •Target capability: predicting stability, synthesizability, and ‘what did we actually make?’
    • •Key claim: generalization is bounded by training data; you need new data from real labs
    • •Discovery requires synthesis + verification loops, not pure reasoning in isolation
    • •Shift in focus away from coding improvements toward physical-world construction
  8. 16:50 – 19:53

    Modeling and optimization for science: why active learning beats Bayesian optimization (in practice)

    In Q&A, Dorje explains that Bayesian optimization’s uncertainty estimates can be unreliable under distribution shift, while active learning shines in real-world, non-stationary regimes. Periodic uses active learning continuously to probe unknown regions, handle anomalies, and expand model competence incrementally.

    • •Physicists often gravitate to Bayesian optimization; practical limits come from uncertainty miscalibration
    • •Uncertainty estimation generalizes worse than point prediction under shift
    • •Active learning may fail on fixed academic splits but is crucial in real operations (e.g., autonomy)
    • •Periodic runs daily experiments that effectively implement active learning
    • •Unexpected anomalies become valuable training data and debugging targets
  9. 19:53 – 21:44

    AI scientist vs AI agent, and what counts as ‘new material discovery’

    They clarify that an ‘AI scientist’ is essentially an LLM-based agent orchestrating tool calls, including other neural nets. Dorje defines ‘discovery’ broadly: new crystal structures, new properties, or improved synthesis recipes—even if the material was known but hard to make well.

    • •AI agent: LLM reasoning + tool orchestration (tools can include other neural nets)
    • •Discovery dimension 1: novel crystal structures beyond known prototypes
    • •Discovery dimension 2: improved properties (e.g., higher superconducting Tc)
    • •Discovery dimension 3: better synthesis recipes/efficiency and optimized outcomes
    • •Long-term aim: expanded human control over arranging atoms in the world
  10. 21:44 – 25:54

    Why conviction formed quickly: unmet hype, poor benchmarks, and massive materials upside

    The host shares why Periodic felt credible compared to prior ‘AI for the physical world’ efforts: measurable progress was often missing. A benchmark paper suggested frontier models were weak at condensed-matter reasoning, and introductions led to rapid alignment on the opportunity—materials as a civilizational bottleneck.

    • •Frustration with gap between marketing and real-world AI-for-physics results
    • •Benchmarking frontier models on condensed-matter reasoning showed poor performance
    • •Fast founder–investor alignment via peer review/intros; rapid decision to start
    • •Materials as key constraint across eras (Stone/Bronze/Silicon) and future tech
    • •Thesis: someone had to build the real lab+AI loop to unlock value
  11. 25:54 – 27:23

    Student anxiety about AGI: what remains undone and how to pick impactful projects

    Dorje addresses student concerns about careers under rapid LLM progress, arguing only a few domains are truly transformed so far. He encourages students to choose problems they care about that aren’t yet “revolutionized,” offering analog circuit design as an example, and frames the moment as a rare high-impact era.

    • •Acknowledges widespread anxiety about education/careers under AGI narratives
    • •Reality check: breakthroughs are uneven (coding/translation/self-driving/force fields)
    • •Advice: use LLMs as leverage to improve what’s still unsolved
    • •Example project direction: ML for analog circuit design
    • •Framing: ‘golden era’ moments are short—capitalize with ambitious work
  12. 27:23 – 29:03

    Simulation vs reality: DFT strengths/limits, catalysis difficulty, and ML force fields

    Dorje gives a technical overview: DFT is strong for ground-state properties but weaker for band gaps/excited states, while catalysis is hard due to messy, unknown active structures and defects. They highlight that ML has significantly improved force fields, and Periodic builds new architectures (including GNNs) but recognizes remaining limits.

    • •DFT: reliable for ground-state formation enthalpy; less reliable for band gaps/excited states
    • •Catalysis: difficult because true atomic structure/defects/steps often unknown and decisive
    • •ML revolution in force fields; large gains from modern neural approaches
    • •Periodic combines LLM expertise with GNN/physics-informed modeling
    • •Ongoing limitation: no single method ‘solves’ all chemical physics regimes
  13. 29:03 – 31:35

    The chicken-and-egg of discovery: no ‘complete’ dataset, so iterate with sample-efficient RL

    They discuss why science cannot have a final comprehensive dataset—once fully known, it becomes ‘textbook’ rather than discovery. Progress requires iterative exploration (active learning), leveraging large public datasets where useful, and aggressively improving sample efficiency—especially for model-based reinforcement learning in costly real-world loops.

    • •Discovery loop is universally chicken-and-egg: you don’t understand what you haven’t found yet
    • •Large datasets help (e.g., massive DFT corpora), but can’t cover future unknowns
    • •Science has no natural endpoint; automating it expands what becomes possible indefinitely
    • •Physical experiments are expensive → sample efficiency becomes a central ML bottleneck
    • •Focus areas: model-based RL, better use of limited experimental data, scalable iteration
  14. 31:35 – 42:08

    Open research problems and practical barriers: synthesis, characterization automation, and partial observability

    They identify high-leverage academic directions: intentional synthesis planning, automating characterization analysis with LLM toolchains, and reasoning under noisy/incomplete context. They close by discussing barriers to physical-world AI (infrastructure cost and build complexity), why these barriers should fall, and how to choose domains with faster experimental loops.

    • •Key open problem: intentional synthesis (reducing brute-force trial-and-error)
    • •Automating characterization: instrument outputs need heavy analysis; LLM+tools can accelerate it
    • •Decision-making under uncertainty: detecting inconsistent evidence in literature/experiments
    • •Motivation & progress tracking: focus on improving many small subcomponents, not only end goal
    • •Barriers: building labs is hard but should get easier with better robots/AI; domain choice matters (vs particle accelerators)

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.