# Superhuman Hardware Design Is the Easy Part AI will soon design hardware better than humans. Design is digital. Building, testing and learning from the physical world are not, and the frontier moves there next. — Christian Keil, a16z^[[a16z.news/p/superhuman-hardware-design-is-the](https://www.a16z.news/p/superhuman-hardware-design-is-the), 20 Oct 2026] > [!abstract] Core idea > The bits-to-atoms interface is **lossy in both directions**: > - **Write**: turning a design into a built object. See [[Five Levels of Building Hardware]]. > - **Read**: measuring the physical world back into a computer. See [[Why the Physical World Is Hard to Read]]. ## Where models stand (Oct 2026) - Models already generate CAD, design PCBs, source COTS parts and write assembly steps. - Benchmarks improve steadily: CAD Arena, EE Bench, LuxoBench (Keil's own). - LuxoBench test: three frontier models helped build a desk lamp. Verdict: "not quite ready for primetime." - Keil will not bet against the Bitter Lesson. Design falls first. ## The split - **Pure design** (the ideal part) happens inside a computer. Models master it fast. - **DFM** (a part that meets spec *and* is cheap and easy to build or buy) needs real-world contact. It is harder. - Each later step in design → build → test → iterate needs more physical contact than the one before. - Result: AI capability in hardware stays **spiky**. The more machine-readable a test result, the faster models improve on it. ## Three predictions 1. **Simulation becomes the main loop.** Better sims give software-speed iteration and fewer trips to wind tunnels and thermal-vacuum chambers. Builds become "confirmatory, not experimental." Names: Navier, Inviscid, Lanyon. See [[Simulation-Based Optimization]], [[Surrogate Models]]. 2. **Demand shifts left, to raw materials and parts.** Millions of new "professional-grade" designers all hit supply chains. - Parts: Senra (harnesses), Podium (control panels), Hadrian (machining), Cofactr (components), Diode (PCBs) - Software: Sift (telemetry), Instrumental (inspection), Revel (test) - Materials: Mariana Minerals (copper) 3. **AI labs must buy process knowledge.** The machinist who hears a bad cut. TSMC's decades of polishing-pad tuning. This data barely exists. ## The data funnel Eric Jang's estimate of robotics training data in 2026: | Data type | Volume | |---|---| | Egocentric video | Millions of hours | | UMI gripper | Hundreds of thousands of hours | | Teleoperation | Tens of thousands of hours | | High-quality, hardware-specific demos | Hours | - Keil expects this funnel to persist after the crossover. - Fix: generate the data yourself. **Periodic Labs** runs autonomous end-to-end experiments on superconductors, magnets and semiconductor materials. Keil calls it the "platonic ideal" of AI-for-hardware data. - Compare: [[RL Data Provenance]], [[RL Environment Stack]]. ## Open problems 1. A better digital–physical interface 2. Faster ways to learn what customers want. A model can design a million Apple Watches. It cannot know which one people will buy. 3. Data pipelines that capture master technicians' process knowledge > [!question] From the comments > - Nan Deng: CAD Arena gains may reflect better *drawings*, not better hardware. Score the same design as a file and as a measured part. > - EnergyPipeline: compare cement jobs in oil wells. The value is in knowing what actually happened downhole. ---- tags: #boat #deeptech links: [[Industrial AI MOC]] [[Five Levels of Building Hardware]] [[Why the Physical World Is Hard to Read]] [[Precision Manufacturing]] [[Digital Twins]] [[AI x Robotics Flow]]