Synthesis Superintelligence: from Semiconductors to Superconductors — Periodic Labs’ Liam Fedus and Ekin Dogus Cubuk
Summary
Periodic is betting that frontier scientific AI becomes most useful when its conjectures are forced through simulations, high-throughput experiments, and noisy physical measurements. Fedus summarizes the premise as “you can’t just think your way to a solution”: papers and pretrained knowledge can generate hypotheses, but reality supplies the decisive evidence. The resulting company resembles a modern Bell Labs, combining chemists, physicists, experimentalists, hardware engineers, and LLM researchers around one closed loop.
The core technical asset is not merely experimental data, but timestamped lineage showing how entire scientific campaigns progressed from failed attempts to working processes. That record includes computations, instrument data, code, discussions, intuitions, negative results, and eventual successes—data that largely disappears from published literature. Fedus’s distinction is decisive: Periodic wants to train “on the process of doing science,” not merely science’s final outputs, creating environments that models cannot solve by recalling memorized answers.
Periodic sees automated characterization as a major bottleneck to scaling materials discovery, rather than simply mixing more powders. X-ray diffraction patterns can contain multiple phases, distinct structures can fit the same signal, and neither synthesized materials nor failed experiments arrive labeled. AI earns its keep by combining thermodynamics, chemical plausibility, replicates, electrical and magnetic measurements, microscopy, and longitudinal evidence; Cubuk’s concise explanation is that the system can “look at everything” where human context is limited.
Lab automation is being optimized for data volume, quality, and reproducibility rather than humanoid spectacle or autonomy for its own sake. Periodic puts context-aware intelligence directly on instruments, redesigns hardware when off-the-shelf equipment creates bottlenecks or contamination risks, and automates routine work before difficult dexterity tasks. Fedus is explicit that “full autonomy is a non-goal”; one system already diagnosed an incorrectly loaded machine by discovering that a cyclic permutation and reversal made the complete dataset consistent.
The near-term commercial wedge is secure, on-site deployment into semiconductor companies, with a longer-term path from researcher copilots to outcome pricing. Forward-deployed teams integrate AI and simulation systems locally, train them on partner data, and let customers “own their own intelligence” rather than transmit proprietary information to a generic API. The aspirational endpoint is what Fedus calls a “matter compiler”: specify a material goal state, then have the system determine whether it can exist and how to synthesize it.
More simulation compute cannot remove the need for laboratories because perfect crystals, approximate theory, and inaccessible physical state omit the variables that govern actual materials. Cubuk invokes “More is different”: known microscopic laws do not make strongly correlated systems, microstructure, kinetics, or high-temperature superconductivity straightforward to calculate. Even 100 million density-functional-theory trials cannot test real microstructure, and scientific discovery is “almost by definition what you haven’t been trained on.”
For superconductors, Periodic’s thesis is that synthesis throughput may matter more than first obtaining a complete theory. Cubuk cites magnesium diboride: a Japanese group reportedly tried 30,000 materials, about 30 of which seemed to show interesting superconductivity, and surfaced a compound that had sat on shelves for decades despite BCS theory dating to 1957. A sufficiently capable automated lab could compress that career-scale search into a month, increasing “the surface area for luck”—but the guests reject the fantasy that a model will simply “zero shot the room temp superconductor.”
Deep dive
1. Reality, not intelligence alone, closes the discovery loop
Periodic’s founding line—“Intelligence is necessary but not sufficient”—means conjectures only become knowledge after reality accepts or rejects them. Fedus argues that no amount of rereading papers or thinking in isolation can enumerate every expected and unexpected experimental outcome.
The intended loop joins AI systems, physical-world simulations, and high-throughput experiments. Fedus expects “a different type of intelligence” to emerge when models must repeatedly propose, execute, observe, and revise rather than optimize against answers already recorded in textbooks.
Cubuk says the required team “has never existed, I think”: solid-state chemists and physicists, experimentalists, theorists, hardware engineers, LLM specialists, and computer scientists work together. The technologies they cite include recent high-throughput experimentation, robotic arms tried in roughly the last 3 years, and recent force-field work.
The historical model is Bell Labs, where theorists, chemists, and experimentalists shared an institution. Periodic’s bet is that the physical lab, concentrated mission, and unusual staffing mix require an irreducibly multidisciplinary organization.
2. Physical science requires reasoning under irreducible uncertainty
Fedus contrasts precise digital environments with furnaces and instruments: products do not emerge labeled, measurements can be stochastic, machines disagree, and recorded temperature may differ from actual temperature. The scientific agent must make decisions despite uncertainty and missing telemetry.
Physical rollouts are also scarce. Experiments take hours or days, cannot be replicated by adding arbitrary virtual environments, and do not provide enough parallel agents to reduce variance; sample efficiency therefore matters far more than in many coding or mathematical RL settings.
Cubuk frames every experiment as dimensionality reduction. The hosts invoke roughly (10^{23}) to (10^{27}) atoms in a room, yet thermodynamics describes much of a system using five or six variables; science succeeds by finding compressed descriptors without pretending they preserve the unknowable full state.
The hidden variables are concrete: furnaces degrade as evaporated material coats their heating elements, temperature varies by position, and optical instruments respond to vibration. Cubuk recalls a PhD lab whose night-time results changed because someone walking upstairs disturbed its laser setup.
3. Reinforcement learning is decomposed into tractable scientific judgments
Fedus divides materials discovery into three linked problems: decide which atomic configuration might be stable and possess the desired property; determine processing conditions that could synthesize it; then characterize what the lab actually produced.
End-to-end rewards such as “did you find a room-temperature superconductor?” are too delayed and noisy. Periodic instead builds environments around individual steps—for example, identifying phases from X-ray diffraction while penalizing spurious or chemically implausible assignments.
X-rays have wavelengths commensurate with atomic spacing, producing diffraction patterns that fingerprint crystal structure. Automating their interpretation prevents characterization from becoming the bottleneck once the lab begins running experiments at high throughput.
Periodic also timestamps its knowledge: given everything experimentally known on a date, what should the scientist have tried next, or what outcome followed? Unlike tasks built from public answers a pretrained model may remember, this proprietary evidence prevents “fake work” that reaches the answer without learning generalizable physics reasoning.
4. Known physical laws do not amount to a perfect materials simulator
swyx’s deliberately basic challenge—“don’t we have, like, most of the laws of physics worked out?”—elicits historical caution. Fedus recounts that Dirac could solve hydrogen and reportedly framed the remainder as “chemistry,” while noting that the quote’s accuracy is uncertain; interacting many-body systems proved nothing like a trivial composition of solved cases.
Conventional superconductivity is largely associated with electron-phonon coupling and an expected isotope effect. Cuprates above 77 Kelvin, including examples around 93 Kelvin, are called high-temperature because they are unconventional; Cubuk says their governing physics remains unknown.
Density functional theory can be highly accurate, but strong electron correlation produces important failures. Brandon invokes Phil Anderson’s “More is different”: simple underlying laws can yield emergent, universal collective behavior that is qualitatively difficult to derive or model.
swyx imagines separating desired properties from synthesis, like a TSMC for materials. Fedus calls the vision a “matter compiler,” while Cubuk cautions that unknowable atomic context may prevent perfect reverse engineering; the practical objective is to understand processes sufficiently to improve performance.
5. Phase identification turns partial failures into navigable gradients
Cubuk introduces phase transitions through melting ice and the Ising model: systems change abruptly from one organized regime to another. In synthesis, heating precursor crystals lets atoms react into a new crystal whose altered geometry should produce a different XRD pattern.
Actual campaigns rarely return a clean yes or no. Early attempts commonly contain precursor material, amorphous matter, the desired phase at low purity, and several unintended phases; simulations and AI must explain that mixture and decide which conditions to alter.
A previously unrecorded phase may require the agent to propose atomic configurations that explain the observations, then test their stability computationally. If the target is only 1% of a sample, accurate characterization supplies a gradient for “hill climbing” toward greater phase purity.
Brandon’s pushback—that phases might provide an unambiguous observable—does not quite hold. Two phases can fit the same pattern, so Periodic adds thermodynamic and chemical priors, replicates, and modalities such as electrical, magnetic, morphological, and electron-microscopy measurements to disambiguate them.
6. Every obtainable measurement is precious, but each is only a probe
Cubuk corrects the premise that Periodic must discard excess data: reality contains more information than any computer could store, but experimenters cannot observe every atom in the first place. “Any data we collect is very precious,” so the company retains it and adds extensive telemetry.
Brandon’s AI analogy treats the physical system as a foundation model containing the complete state, while an experiment is a tiny linear probe returning perhaps “8 bits.” Scientists and agents must reconstruct the latent physical situation from that severe coarse-graining.
Maxwell’s demon supplies the limiting thought experiment. An all-knowing sorter could apparently lower entropy, but Landauer’s argument says deleting the accumulated information costs energy; Periodic is not at the level where it can be Maxwell’s demon, though Cubuk calls swyx’s concern “visionary.”
More sensors still reduce hidden variables, and the hosts confirm that Periodic effectively does “buy every sensor.” A future network of labs could also account for geographic effects and draw on different concentrations of physics and chemistry expertise rather than treating Menlo Park as the only viable location.
7. Simulation and experiment correct different kinds of error
Cubuk describes density functional theory through the Hohenberg–Kohn insight: ground-state properties can, in principle, be expressed as functionals of 3-dimensional charge density rather than an exponentially large wavefunction. Practical Kohn–Sham methods make otherwise intractable materials calculations possible.
Brandon summarizes the practical tradeoff as reducing an exponentially hard problem to an approximate calculation that is usually (N^3), with introduced error. DFT estimates formation enthalpy, helping determine whether a proposed material lies on the convex hull against alternative compounds.
Cubuk calls DFT the best stability-prediction method he knows, while stressing that current practical approaches lack the needed exchange-correlation functional and, for charge-density formulations, the kinetic-energy functional. Even perfect DFT would not fit (10^{23}) atoms in a computer.
Brandon asks how millions or 100 million simulated trials can be prevented from dominating the comparatively limited experimental stream. The answer is to scale the lab too: DFT usually models perfect crystals, cannot directly explore real microstructure, and cannot readily predict properties such as superconducting temperature.
Experiments calibrate simulations for each chemical system, while simulations estimate quantities that are difficult to measure cleanly. Cubuk’s conclusion is not simulation versus laboratory but “simulations, AI, and experiments” in a loop, each compensating for the others’ blind spots.
8. Inorganic materials resist biology’s compact representations
Cubuk provocatively suggests that biology and organic chemistry may often admit lower-complexity descriptions, with DNA and RNA approximating sequence-like representations; Liam notes that many people would describe the representation as more 2-dimensional. Brandon separately adds that biology benefits from a toolkit constrained by millions of years of evolution.
Three-dimensional inorganic structures can produce metals, insulators, superconductors, diamond, covalent bonds, ionic bonds, and metallic bonds. Even perfect-looking XRD is an average: atoms may shift in compensating patterns while appearing centered in the aggregate.
Cubuk contrasts biology’s sophisticated nanoscale machinery with solid electrolytes, where lithium dendrites can invade and crack an interface. Yet inorganic materials tolerate high temperatures, enable spacecraft, and underpin silicon compute—different capabilities rather than a simple hierarchy of difficulty.
9. The materials prize set runs from magnets to compute efficiency
Beyond room-temperature superconductors, Cubuk names better magnets with less dependence on scarce rare earths or transition metals such as cobalt, batteries using less cobalt, and more energy-efficient computation. These are materials problems whose commercial value follows directly from device performance and supply constraints.
He highlights the Landauer limit: energy per FLOP has historically fallen with Moore’s-Law-like exponential behavior, and approaching the thermodynamic floor would make computation far more efficient in what he calls “the real currency in the universe.” The discussion gives conflicting rough estimates of the remaining gap: Liam recalls 10 or 20 orders of magnitude, while Brandon recalls 3 or 4; no one claims a current checked figure.
Reversible computing could avoid information deletion and therefore some dissipation, though Cubuk’s answer is intentionally uncertain: “I don’t know if that works.” Brandon describes quantum computing as unitary and reversible, but Cubuk argues that even perfect quantum simulation might still model perfect crystals while missing microstructure and experimental reality.
10. “140 IQ” instruments optimize evidence, not robotic theater
Fedus says early lab scaling exposed human operating bottlenecks. Simple programs could scan fields of view and zoom, but their captures were too context-blind; AI-controlled instruments instead receive the experiment’s intent, target material, and prior evidence before deciding what data to collect.
Intelligence at the instrument improves future datasets, but deployment depends on physical timescales. Cloud calls, long reasoning traces, and simulations can be too slow or expensive for fine control—and a scientist waiting 2 minutes for analysis has a different workflow from one waiting 2 hours.
swyx tests whether incremental automation is a local optimum and a humanoid should simply run the existing lab. Fedus disagrees: automate repetitive, high-volume work first, leave unusually dexterous tasks to people when robotics would take months, and remember that “full autonomy is a non-goal.”
One AI system noticed longitudinal inconsistency from an incorrectly loaded machine, then found that cyclically permuting and reversing the records restored coherence. The diagnosis exposed a physical loading error and identified another automation opportunity, illustrating why data-quality control is foundational rather than clerical.
11. Negative results create the scientific training distribution
Periodic uses both open and closed models. Unique internal data can justify smaller, lower-latency systems, improve compute efficiency, and sometimes surpass frontier models at specialized tasks—not merely reduce cost—because outside models lack the relevant experimental distribution.
A null result may mean the intended crystal was absent, an impurity phase destroyed the property, or a different previously unrecorded structure appeared. Periodic deliberately runs negative controls; Cubuk notes that classifiers cannot learn effectively when literature reports successful syntheses but usually omits failed ones.
Failure remains contextual rather than absolute: an unsuccessful synthesis could reflect “a skill issue,” the wrong method, or current technology. The valuable unit is therefore often an entire campaign—a sequence of failures, processing changes, and eventual success—rather than one experiment labeled negative.
Fedus says Periodic records conversations, intuition, computations, code, lab execution, and outcomes to recover the campaign’s full lineage. Compressing those patterns into weights still matters: otherwise frontier labs could have stopped training at GPT-4 and relied solely on inference-time reasoning.
12. Characterization and custom hardware determine laboratory scale
Asked to map the abstraction stack, Cubuk names atomistic structures, continuum models, thermodynamics describing the infinite-time endpoint, and kinetics describing what happens within finite time. A theoretical advance could sit above multiple campaigns, while individual instruments and operations sit below them.
His practical bottleneck remains automated characterization: “it’s not that hard to mix powders,” but random throughput is useless without rapid analysis and an informed next experiment. Human campaign guidance and characterization have consequently shifted toward mixed or predominantly AI-driven workflows.
Search itself is familiar rather than mystical. Evolutionary methods have long helped predict low-energy structures, while semiconductor process engineers use design of experiments, Bayesian optimization, or related zeroth-order techniques; the novelty is connecting those choices to rich physical feedback.
Parallelization has limits. Combinatorial sputtering can create hundreds or thousands of compositions across a gradient, but diffusion blurs the result; pooled superconductivity screening resembles pooled COVID tests, yet a single superconductivity measurement can still take about 1 hour.
13. Periodic is scaling a new lab architecture and a commercial flywheel
Fedus says available resources will expand both compute and physical labs, including new lab types beyond the initial Menlo Park setup. Each build teaches which instruments justify higher capital expenditure and where the next version needs greater ambition or precision.
Scale increasingly requires custom hardware. An off-the-shelf system may be fast everywhere except its weighing step, while a robotic arm parked above previously mixed samples introduces contamination; redesigning those details lowers the experimental noise floor and raises usable throughput.
The organization spans AI research and infrastructure, computation, experimental science, hardware, and product. Cubuk emphasizes experienced people remaining hands-on—the Bell Labs ideal of senior theorists and lab leaders working directly with experiments rather than migrating entirely into administration.
The current revenue wedge is semiconductors: forward-deployed engineers work inside highly secure customer environments, run inference locally, and train systems on partner data. Fedus expects a progression analogous to software—from Copilot-like assistance toward pricing a requested physical outcome as autonomy improves.
14. Commercial success is intended to finance deeper science
Cubuk wants materials research to become visibly profitable, attracting talent the way ChatGPT increased interest in computer science and machine learning. Bell Labs is invoked not only as a research culture but as proof that foundational physical research can coexist with enormous commercial impact.
Fedus argues that technology and capital reinforce each other. Chatbots changed slowly from 2010 to 2015, then dramatically from 2021 to 2026 after product-market fit transformed access to hiring, compute, and data; Periodic wants an equivalent virtuous cycle for physical-world intelligence.
Commercialization does not preclude public infrastructure. The team cites contributions to Pymatgen, Materials Project tooling, Custodian, TorchSim, JAX-MD, SGLang, and Megatron, alongside academic gift grants whose first funded paper was reportedly nearing publication.
15. Superconductor discovery expands the surface area for luck
Brandon’s final steelman grants perfect automation but asks how it finds superconductors without a theory of strongly correlated systems. Cubuk replies that plausible chemical spaces are plentiful; the harder problem is synthesizing candidates, making “synthesis superintelligence” the operative target.
Alex Müller reportedly tried nickelates before succeeding with cuprates; later, Harold Hwang produced superconducting nickelates as thin films. Nickel, copper, cobalt, oxygen, and related transition-metal systems provide connected search directions even without a complete predictive theory.
Cubuk’s strongest precedent is magnesium diboride, the highest-temperature ambient-pressure conventional superconductor discussed. A Japanese group tried roughly 30,000 materials and found about 30 interesting candidates, although MgB2 had sat on shelves for decades and BCS theory dated to 1957.
Compress those 30,000 trials from a career into a month and the laboratory does not eliminate luck—it increases “the surface area for luck.” Models will continue to propose, simulate, characterize, and learn, but both guests insist that no system will simply think its way to discovery: “no one’s gonna zero shot the room temp superconductor.”