While large language models have thoroughly revolutionized digital domains—mastering code, mathematics, and complex reasoning—they have yet to make a comparable dent in the physical world. The reason is simple: software has an instant answer key, but the physical world requires physical experiments. Now, two former leads from OpenAI and Google DeepMind are building the infrastructure to bridge that exact gap, creating what they call a “synthesis superintelligence” designed to engineer matter itself.

The Limits of “Thinkism”

Liam Fedus (former Post-Training Lead at OpenAI and a creator of ChatGPT) and Dogus Cubuk (former Research Scientist at Google DeepMind) recently founded Periodic Labs to address a fundamental bottleneck in AI development.

Currently, AI models rely heavily on what the founders call “thinkism”—the idea that if an AI thinks long enough on existing literature, it will intuitively discover breakthroughs like high-temperature superconductors. Fedus and Cubuk argue this approach is inherently flawed. Science pushes the boundaries of the unknown; it cannot rely solely on the final, published artifacts of past experiments. To create net-new knowledge, an AI must run actual physical experiments, fail, learn from the telemetry, and run an error-correction loop—a process previously exclusive to human scientists.

Orchestrating the Physical Lab

Periodic Labs is building an end-to-end infrastructure where an AI doesn’t just read about science; it actively orchestrates it.

• Intelligent Synthesis: The AI suggests combinations of precursor powders and determines the precise processing conditions (like temperature profiles and pressures) needed to synthesize entirely new materials.
• Bridging Disciplines: Historically, experts in computer simulation and experts in physical experimentation operate in silos. Periodic’s AI acts as a fluent bridge, running PhD-level simulations while simultaneously performing PhD-level experimental data analysis (such as reading X-ray diffraction patterns to identify newly created materials).
• The “Chain of Thought” for Science: Unlike public literature, which only publishes successful outcomes, Periodic is capturing the entire experimental lineage—the hypotheses, the computational predictions, the robotic executions, and the physical results. This proprietary data snapshot creates a uniquely powerful reinforcement learning environment for the AI.

The Target: High-Temperature Superconductors

While the lab infrastructure is highly generalized, Periodic’s first major moonshot is the synthesis of high-temperature superconductors. Superconductors are the holy grail of materials science, holding the key to sci-fi futures like lossless energy transmission, advanced quantum computing, and commercial fusion.

Furthermore, the founders note a massive economic shift on the horizon. Much like the pharmaceutical industry evolved to place billions of dollars of value on drug discovery IP, the materials science sector is approaching a similar inflection point. By closing the loop from discovery to physical synthesis, Periodic aims to drastically shorten the notorious “20-year bottleneck” it usually takes for a lab material to become an industrialized technology.

As the digital frontier becomes increasingly crowded with frontier models mimicking one another, the physical frontier remains wide open. By directly connecting artificial reasoning to robotic laboratories, Periodic Labs isn’t just training an AI on the state of science—they are training it on the process of science.