AI Scientists Are Here: Autonomous Labs & Synthesis Superintelligence — Periodic Labs
Summary
Periodic Labs’ central bet is that intelligence alone cannot create physical knowledge. Its operating maxim is, “Intelligence is necessary but not sufficient. New knowledge is created when ideas are found to be consistent with reality.” The company is therefore building AI, simulations, and high-throughput physical labs as one iterative system: conjecture, synthesize, characterize, update, repeat.
Physical science demands a different AI regime from math or coding because observations are noisy, incomplete, expensive, and slow. Furnaces degrade, temperature varies by position, vibrations disturb optical equipment, and finished samples do not arrive labeled; unlike digital environments, Periodic cannot create unlimited rollouts. The governing constraints are decision-making under uncertainty and extreme sample efficiency: “Science is very hard, but that’s what makes it special.”
Periodic’s potentially differentiated asset is not merely experimental data, but timestamped lineage of how science gets done. It captures failed attempts, computations, code, conversations, intuitions, instrument readings, and the sequence by which a campaign moves from repeated negatives to a positive result. This permits training on evidence and timestamped states where a pretrained model cannot simply rely on memorized answers, shifting the target from “the final output of science” to “the process of doing science.”
Automated characterization—not humanoid robotics—is the near-term bottleneck with the highest leverage. Mixing powders is relatively easy; identifying what was produced, reconciling multiple instruments, and selecting the next experiment constrains throughput. Periodic’s “every piece of equipment…140 IQ” framing means context-aware instruments that collect better data, while “full autonomy is a non-goal”: automation serves data quantity, quality, and diversity.
Simulation is a high-volume filter, not a replacement for reality. Density functional theory can estimate material stability, but practical approximations struggle with strong correlations, microstructure, and properties such as superconducting temperature; even 100 million perfect-crystal calculations cannot reveal every real sample’s defects and intermediate-scale structure. Periodic’s answer is a calibrated loop among simulations, AI, and experiments, with the lab remaining “the ultimate truth.”
Scaling requires simultaneous investment in laboratories, compute, instrumentation, and an unusually multidisciplinary organization. Off-the-shelf equipment eventually exposes bottlenecks and contamination risks, forcing custom hardware; additional labs can also access different expertise and experimental conditions. The organizational model deliberately combines solid-state chemists, physicists, experimentalists, theorists, AI researchers, infrastructure engineers, hardware builders, and product teams—an “irreducibly multidisciplinary problem.”
The immediate commercial wedge is secure, on-site deployment into semiconductor companies, with a longer-term path toward outcome pricing. Periodic first built the tools for itself as “customer zero”; forward-deployed engineers now integrate local AI and computational systems, train them on partner data, and let customers own specialized intelligence rather than send sensitive data to a generic API. Liam’s analogy is software’s progression from Copilot-style assistance to Codex-like autonomy: eventually, “help me get to this kind of goal state.”
The superconductivity thesis is that synthesis throughput may matter more than a complete theory. Ekin argues it is not hard to think of chemical spaces that could host superconductivity; reliably making and testing candidates is the harder problem. A Japanese group reportedly tried 30,000 materials, found roughly 30 interesting candidates, and surfaced MgB₂ despite its having sat on shelves for decades—so a closed-loop lab that compresses a career’s search into a month would “increase the surface area for luck.”
Deep dive
1. Periodic treats reality as the final training environment
Liam’s starting premise is that “you can’t just think your way to a solution.” The universe contains expected and unexpected outcomes that neither rereading papers nor reasoning in isolation can enumerate, so discovery requires conjectures repeatedly tested against physical reality.
That premise dictated the company architecture from day one: AI systems, simulations of the physical world, and high-throughput experiments had to be built together. Liam believes “a different type of intelligence emerges” when models must act, observe consequences, and revise their beliefs.
Ekin compares the intended institution to Bell Labs: solid-state chemists and physicists, experimentalists, theorists, hardware engineers, LLM specialists, and computer scientists working as one team. Recent robotic experimentation and force-field advances make the combination newly practical, but no existing team contained the complete playbook.
2. Scientific reasoning begins by discarding almost all of reality
Ekin’s distinction from coding and math is dimensionality reduction. A program’s relevant functions and APIs may fit in context; a physical system begins with roughly 10^23 or more atoms, far beyond exact observation or storage, forcing scientists to identify a few useful variables.
Thermodynamics is the canonical success: steam engines appeared impossibly complicated until temperature, pressure, energy, and a handful of related variables explained much of their behavior. Atomistic pictures—local tetrahedra, octahedra, energies, fluctuations, and reaction barriers—provide additional compressed descriptions.
Experiments are similarly lossy. Scientists record everything accessible while knowing they will miss hidden state: a furnace’s heating element accumulates evaporated material, temperature varies across its chamber, and an upstairs person walking at night once altered a PhD lab’s laser results through vibration.
3. Noise and slow rollouts rule out one giant end-to-end reward
A materials campaign must choose a potentially stable configuration with useful properties, determine how to synthesize it, then characterize what actually emerged. Each stage is non-trivial; “things don’t come off the furnace labeled.”
Liam rejects an RL loop that launches an experiment, waits days, and rewards whether it found a room-temperature superconductor. Physical runtimes, scarce agents, instrument delays, and noisy outcomes make variance reduction infeasible at that horizon: “You can’t leave your GPU sitting for a week.”
Periodic instead builds tractable environments around individual stages. For X-ray diffraction, an agent can be rewarded for identifying phases present in the raw pattern and penalized for spurious or chemically implausible phases, relieving a characterization bottleneck that otherwise grows with experimental throughput.
These stage-level agents can then be stitched into the end-to-end discovery loop. The broader lesson is that scientific RL needs intermediate rewards grounded in actual measurements, rather than a single heroic payoff at the end of a slow physical campaign.
4. Timestamped evidence prevents models from faking scientific work
Periodic can freeze the state of knowledge at a particular date: this was the experimental evidence available, these were the possible decisions, this was the scientist’s next choice, and this was the subsequent outcome. Those snapshots become training environments for evidence-based scientific judgment.
Liam’s concern with already-memorized data is that a pretrained model can reach the correct result without performing the physics, calculation, or simulation; reinforcement learning then up-weights a shortcut that will not generalize to genuinely novel systems.
Internally generated data has provenance across the entire process, making such leakage less plausible. The system can learn which action was reasonable from the evidence available then, not merely reproduce the endpoint now that history has revealed it.
The same finite experimental basket can support multiple environments: one instrument alone, combinations of modalities, subsets of a campaign, or evidence bounded by time. That does not create new physical observations, but it expands the number of scientifically meaningful learning problems constructed from them.
5. Knowing basic laws does not make many-body matter solvable
swyx’s deliberately basic challenge—“Don’t we have most of the laws of physics worked out?”—elicits the episode’s central scientific correction. Ekin invokes the story that Dirac regarded quantum mechanics as finished and dismissed what remained as “chemistry”; solving hydrogen did not make nitrogen, oxygen, or solids a simple composition exercise.
High-temperature superconductivity is one conspicuous gap. Conventional electron-phonon superconductors show effects such as the expected isotope dependence, but cuprates above 77 Kelvin, including examples around 93 Kelvin, do not obey that known mechanism; their governing physics remains unresolved.
Even density functional theory, powerful and accurate in many regimes, can miss effects once strong electron correlation matters. Brandon’s summary is Phil Anderson’s “More is different”: simple microscopic laws can produce qualitatively new collective behavior when many components interact.
Liam connects this to deep learning’s unexplained power laws, emergence, and universality. The discussion does not promise perfect reverse engineering of every atom; the practical target is sufficient understanding to improve material performance—a prospective “matter compiler,” not omniscience.
6. Phase identification turns messy outcomes into an actionable search signal
Ekin explains phase transitions through melting ice and the Ising model: a system changes little until a threshold produces a qualitatively different state. In synthesis, precursors heated together may form new crystals, and the altered atomic geometry appears as a changed X-ray diffraction fingerprint.
The real output is rarely binary. Early attempts commonly contain leftover precursors, amorphous material, unintended compounds, and only a small fraction of the target phase; the agent must determine what formed and recommend how conditions should change.
Some observed phase may not exist in any paper or database. The AI must propose atomic configurations consistent with the measurement, test their stability computationally, and help hill-climb a desired phase from perhaps 1% purity toward a cleaner sample.
Brandon’s pushback—that a phase might provide an unambiguous observable—does not fully hold. Replicates help, but different structures can match the same pattern; thermodynamics, synthesis conditions, and chemical plausibility remain priors rather than certainty.
7. Multimodality is valuable because every instrument is a narrow probe
X-ray diffraction may leave two phases indistinguishable, while electrical, magnetic, morphological, or electron-microscopy measurements separate them. Ekin argues that consistently reconciling ten modalities is taxing for humans but “not super intelligent for an AI”; it primarily requires context and computation.
Liam’s formulation of superintelligence here is operational: inspect every instrument, every replicate, and longitudinal evidence across experiments without over-indexing on one reading. Better decisions emerge from combining weak, partially conflicting signals into a stronger latent model.
Periodic therefore does not delete collected data. Ekin clarifies that nature contains more information than any computer could store, but humans cannot observe most of it in the first place; accessible measurements remain precious, and heavy telemetry reduces the hidden-variable count.
Brandon likens an experiment to an eight-bit linear probe of a vast foundation model: reality contains the full state, while the instrument emits a tiny coarse-grained projection. The scientific task is reconstructing what happened from that narrow output.
8. Simulation compresses quantum mechanics but cannot erase the lab
Ekin describes density functional theory as the dominant materials simulation method. The Hohenberg–Kohn result replaces an exponentially large wave-function description, for a ground state, with a three-dimensional charge density whose functionals determine the system’s properties.
He corrects an important simplification: DFT need not be approximate by definition, but practical work lacks the exact exchange-correlation functional and, for charge-density approaches, the kinetic-energy functional. Formation enthalpy then estimates whether a material’s energy is competitive with alternative arrangements—the convex-hull test for stability.
Simulation wins where experiments provide difficult or inaccurate measurements, while experiments win on actual reality. Periodic calibrates computations against data for each chemical system and uses the combined loop to filter many theoretical candidates into a much smaller set worth making.
Brandon’s scaling concern remains: compute might enable 100 million DFT trials while labs produce far fewer samples. Ekin’s answer is that DFT usually models perfect crystals, not consequential microstructure, and cannot easily predict properties such as superconducting temperature; volume does not repair a missing physical representation.
9. Inorganic materials resist biology’s compact representations
Ekin provocatively frames biology and organic chemistry as having lower representational complexity: DNA and related systems admit something close to a one-dimensional sequence, whereas inorganic crystals lack an equivalent to a SMILES string and combine covalent, ionic, and metallic bonding in three dimensions.
That freedom supports wildly different outcomes—metals, insulators, superconductors, diamond, high-temperature structures, and semiconductor devices—but creates inference ambiguity. Even perfect-looking XRD is an average: atoms may shift in opposing directions through disproportionation while the aggregate pattern places them in the middle.
Brandon’s counterweight is that biology’s apparent tractability borrows a toolkit refined by millions of years of evolution, while cells and organisms become immensely complex at larger scales. The exchange ends without declaring a universal winner: biology gains complexity across organization levels; materials gain it through bonding, defects, interfaces, and microstructure.
10. Energy-efficient compute may be as consequential as a dream battery
Asked for science-fiction targets, Ekin prioritizes superconductors, improved magnets, and reduced reliance on difficult-to-source rare earths or transition metals such as cobalt. Batteries matter, but he elevates computation near the Landauer limit as a foundational materials objective.
Energy per FLOP has improved exponentially in a Moore’s Law-like progression. The discussion recalled an older estimate of roughly three to four orders of magnitude from the theoretical limit; Ekin explicitly cautioned that he had not checked the figures in about ten years.
Getting closer would reduce both computation energy and the heat that must be dissipated—important because energy is, in Ekin’s phrase, “the real currency in the universe.” Reversible computing might evade dissipation associated with deleting information, but he gives an honest non-answer: “I don’t know if that works.”
Quantum computing receives similar skepticism. It could break security and may enable new possibilities, but a perfect quantum calculation of a perfect crystal still omits real microstructure; excitement about quantum logic can obscure the unresolved question, “What would you simulate?”
11. Intelligent instruments matter more than a fully humanoid lab
“Every piece of equipment…140 IQ” arose from a mundane bottleneck. Simple programmed microscopy captured fixed fields of view and zoom levels, but the images were not useful enough; an AI-controlled instrument instead sees experimental intent, target material, and prior evidence while choosing what data to collect.
Latency matters even when a full experiment lasts hours or days. Fine control can occur on shorter timescales, and an analysis may require many reasoning passes, searches, simulations, and tool calls; two minutes versus two hours also changes whether a human scientist can work interactively.
swyx challenges the incremental automation path as a possible local optimum and suggests dropping in a humanoid. Brandon’s answer is pragmatic: leave rare, dexterous tasks to people if automation would take months, while automating routine operations that consume substantial technician time.
The objective is explicit: “Full autonomy is a non-goal.” Automation exists to produce more, better, and more diverse data—not to remove humans for its own sake.
12. Standardization lowers the experimental noise floor and exposes mistakes
Liam recounts an AI discovering a cyclic error in instrument data. Given the declared run, the pattern was impossible; reversing a cyclic permutation restored consistency, allowing the team to trace the fault to incorrectly loaded physical equipment.
That incident turned data-quality monitoring into a hardware opportunity: redesign the workflow so the same mistake cannot recur. Robots may make fewer errors, but their larger value is repeatability—creating a controlled process whose deviations become legible rather than disappearing into laboratory folklore.
Ekin argues that published measurements across people, labs, geographies, and eras can carry a noise floor “actually usually worse than DFT accuracy.” Standardized workflows could narrow those variations and make physical data substantially more useful for model training.
The team also over-instruments rather than discarding sensor streams. Maxwell’s demon appears as the distant limiting analogy: only an almost all-knowing observer would collect enough information for deletion costs to become the central problem; today, the scarce resource is still observable ground truth.
13. Failures and campaign lineage are the proprietary learning substrate
Periodic uses both open and closed models, choosing among latency, cost, and capability. The stronger claim is that exclusive physical data can let a specialized system exceed frontier models at their highest reasoning settings while using compute more efficiently.
Null results supply what materials literature usually omits. A failed synthesis may provide evidence that the target was absent under specified conditions, reveal a new unintended structure, or serve as a deliberate negative control against toxic or property-killing impurity phases.
Ekin preserves the uncertainty: failure never proves a crystal is permanently unsynthesizable; “it could just be a skill issue,” an inadequate method, or missing technology. But contextual negatives are still indispensable to classification because a dataset containing only published successes cannot teach the boundary.
The richer unit is an entire campaign: the failed sequence, changing conditions, calculations, conversations, code, and eventual success. Liam’s formulation is the data strategy in one line: train on “the process of doing science,” not merely its polished final output.
14. Better weights will not eliminate experimentation
Brandon asks whether models such as “Fable 6” or GPT-8 could absorb campaign-level reasoning and perform it without a specialized loop. Liam concedes that uncertainty also exists in machine learning, but argues that physical-world interaction creates distinct challenges worth compressing into dedicated weights.
His evidence is the frontier labs’ own behavior: if inference-time reasoning alone were sufficient, they could have stopped training at GPT-4. Their continued investment in training implies value in compressing capabilities into weights rather than reconstructing everything through long reasoning at inference.
Ekin adds the harder boundary: even if “Fable 7” becomes dramatically better, scientific discovery concerns what the model has not seen. It must still “tinker with the universe” to obtain new evidence; “no one’s gonna zero shot the room temp superconductor.”
There is nevertheless a larger recursive loop. Better materials could improve chips, chips improve AI, and AI improves experiments—the “big unified wheel”—but every turn remains grounded by a physical test.
15. Characterization is the bottleneck around which human work reorganizes
Ekin maps several abstractions rather than offering one hierarchy: atomistic structure versus continuum models, and thermodynamics’ infinite-time endpoint versus kinetics’ account of how quickly reactions occur. A theoretical advance can sit above these layers and redirect multiple campaigns at once.
Periodic’s founding bottleneck hypothesis was automated characterization. Randomly mixing powders is not especially valuable if the system cannot identify the result, compare it with simulations, and intelligently choose tomorrow’s experiment.
Human placement consequently changes over time. Campaign selection began fully human-led and is increasingly mixed; characterization moved sharply toward AI, freeing scientists from repeated refinements to work at higher levels rather than eliminating their contribution.
Search methods are not unique to LLMs. Evolutionary algorithms already support low-energy structure prediction, while semiconductor process engineers use design of experiments, Bayesian optimization, and related zeroth-order techniques; the novelty lies in connecting search to richer, standardized feedback.
16. Physical scaling mixes parallelism, clever pooling, and custom hardware
Materials do offer multiplexing tricks. Combinatorial sputtering creates composition gradients that test hundreds or thousands of crystals spatially in one run, though Ekin says diffusion has limited its success by allowing components to move.
Superconductivity measurements remain a bottleneck at roughly one hour per candidate. One proposed shortcut pools several candidates into one sample, tests whether any superconductivity exists, then narrows down which source did—analogous to pooled COVID testing.
New funding would scale both compute and physical labs while enabling additional kinds of laboratories beyond the Menlo Park setup. The team expects each build-out to improve the next through better instrument choices, higher ambition, and knowledge of which expensive machines generate the most return.
Scale ultimately forces vertical integration. Off-the-shelf systems revealed a weighing machine as a throughput constraint and a robotic arm whose resting position risked contaminating completed mixtures; bespoke hardware can remove such bottlenecks and push the noise floor lower.
17. The company is productizing its internal loop through semiconductors
The org chart follows the loop: AI research and infrastructure, computation, experiments, hardware engineering, and product. Ekin’s hiring principle is equally deliberate—experienced researchers remain hands-on, echoing Bell Labs figures who continued calculations and experiments rather than graduating into pure administration.
The examples make the culture concrete: Murat Aykol works across computational and experimental intuition; Joe Checkelsky brings superconductivity, physics, and synthesis; Dima Bahdanau goes from attention research into training traces and solid-state chemistry; Reiichiro Nakano physically joins scientists to automate machines.
Periodic built its tools for itself as “customer zero” and is now taking them into semiconductor companies. Because partner data is private, forward-deployed engineers operate on-site in secure environments, integrate AI and simulation systems, perform local inference, and train models on customer-specific evidence.
Liam expects a progression from research copilots toward autonomous systems and eventually outcome pricing: “Help me get to this kind of goal state.” The partner owns intelligence specialized to its data, rather than sending proprietary knowledge to an untrained external API.
18. Commercial success is intended to finance a scientific flywheel
Ekin wants solid-state physics and materials science to become visibly profitable, attracting students and capital as ChatGPT increased interest in computer science and LLM work. Bell Labs is the model not only for research culture but for science tied to commercially valuable infrastructure.
Liam’s comparison is capital formation: chatbot progress from 2010 to 2015 was hard to describe, while 2021 to 2026 was “night and day.” Product-market fit changed the resources available for hiring, compute, and data; “technology and capital are incredibly intertwined.”
Periodic is also contributing to open projects including pymatgen, Custodian, TorchSim, JAX-MD, SGLang, and Megatron, while funding academic groups through a grant program. The commercial loop and open ecosystem are presented as complementary ways to accelerate synthesis intelligence.
19. Superconductivity may yield to search before it yields to theory
Ekin’s categorical call is that promising chemical spaces for superconductivity are not hard to imagine; synthesizing the candidates is hard. That is why the company describes its objective as “synthesis superintelligence,” not merely a model that generates theoretical structures.
His historical specimen is Alex Müller’s search among transition-metal oxides: nickelates did not initially work, cuprates did, and the latter produced a Nobel Prize. Later thin-film work showed nickelates were not a foolish direction—nickel and copper are neighbors on the periodic table, and synthesis capability changed what could be tested.
Magnesium diboride sharpens the argument. A Japanese group reportedly tried 30,000 materials, found about 30 with interesting superconducting signals, and identified MgB₂ even though it had existed on shelves for decades and BCS theory dated to 1957.
The material was neither predicted cleanly from theory nor blocked by fundamental synthesizability; somebody finally tested it in the right way. A closed-loop lab that executes a career’s 30,000 attempts in a month would not abolish luck—it would “really increase the surface area for luck.”