Superhuman Hardware Design Is The Easy Part
Summary
- Frontier AI models can currently generate CAD, design custom PCBs, source COTS parts, & describe assembly, though they remain insufficient for ready-to-build hardware engineering.
- The interface between the digital & physical worlds is lossy, creating significant barriers to translating digital designs into physical atoms & measuring real-world results back into computers.
- Hardware development involves five distinct levels of difficulty: 101 (decomposing products & sourcing COTS parts), 201 (custom part manufacturing & CAM translation), 301 (complex electromechanical assembly & tolerance stacking), 401 (managing external vendor inputs), & 501 (mass-scale production).
- AI models excel at pure design but struggle with "design for manufacturing" because physical variability, non-uniform materials, sparse data, high measurement costs, inherent process latency, & qualitative user feedback are difficult to quantify & automate.
- Future superhuman hardware engineers will require better simulation tools to iterate at software speeds, a shift in demand toward raw materials & component suppliers, & AI labs must acquire rare "process knowledge" from physical experiments.
- The most incompressible step for AI remains gathering end-user feedback, as models cannot easily predict customer preferences without direct physical interaction.
- Hardware engineering will shift from design to solving challenges in the real world, including improving digital-physical interfaces, discovering customer needs faster, & building data pipelines that replicate master technician expertise.
The interface between the digital and physical world is lossy
- Keywords
- American Dynamism