Superhuman Hardware Design Is The Easy Part
The author argues that frontier AI models are getting good at generating CAD and designing parts, but physical hardware is hard to automate because the link between digital designs and the real world is lossy in both directions. Writing a design into atoms and reading measurements back out both introduce errors that models will have to work through.
- Hardware work breaks into levels (101 to 501), from sourcing COTS parts to building a million units, and each level needs different skills that have to be solved separately.
- Design for manufacturing is the hard part: models may produce ideal parts, but making them buildable and cheap requires constant real-world interaction.
- Real-world data is non-uniform, sparse, expensive to collect, and slow, so measuring what is actually happening on a factory floor or supply chain is difficult.
The core argument is that AI's progress in hardware design runs into a physical bottleneck: the boundary between digital bits and physical atoms is lossy in both directions. The author, who spent eight months building a first FPV drone (buying parts in January, setting one on fire in July, crashing it into a neighbor's roof in August), notes that frontier models can now generate CAD, design PCBs, source off-the-shelf parts, and describe assembly. Benchmarks like CAD Arena, EE Bench, and his own LuxoBench show steady gains. But he argues the models are still far from superhuman in practice, and that the hard part is not pure design but getting designs into the world and getting measurements back out.
He breaks the "write" problem into five levels of difficulty. At the simplest, you pick off-the-shelf components, check they're compatible, and actually buy them, which can stall even elite teams (he cites a 2006 Falcon 1 hot fire delayed by a single capacitor stuck 6,000 miles away). Custom parts add a translation step from CAD to CAM to machine instructions, where a master machinist's judgment about stock, fixturing, and tooling introduces variability. Full assemblies bring tolerance stacking, where every part meets its drawing but the whole fails to fit, plus coordination across teams and vendors. At the top, scaling from one unit to a million multiplies variability, financing needs, training, and vendor limits. His key distinction is between "pure design," which AI may handle well by finding the ideal part for a job, and "design for manufacturing," which requires constant contact with real suppliers and processes and should therefore be harder for models to master.
The "read" side gets the same treatment through an example from a friend who buys 45 million chickens a week. Even with identical specs, suppliers, and machines, his output varied by up to 20% week to week, and fixing that took heavy measurement, supplier pressure, and sophisticated data work. In one case he caught a supplier injecting water into chickens, so the weights passed intake checks while the product was less meat than paid for. The lesson the author draws is that sensing the physical world is as lossy as writing to it, which means superhuman hardware engineers will need extensive real-world instrumentation and feedback, not just better models, and that the transition will likely arrive unevenly across domains rather than all at once.