Superhuman Hardware Design Is the Easy Part

Frontier AI is improving at CAD, PCB design, and selecting components, making superhuman hardware engineers seem close. But generating an ideal design is only one part of building a working product. The central obstacle is the lossy interface between digital plans and physical objects: designs must be translated into parts, assembled, tested, and repeatedly adjusted. This work spans distinct challenges, from sourcing compatible off-the-shelf components and manufacturing custom parts to coordinating complex assemblies, managing external suppliers, and producing at scale. Each step brings practical constraints - availability, cost, buildability, tolerance errors, and delivery schedules - that may not be visible in a model’s design files.

The physical world is also difficult for AI to observe and learn from. Materials and processes vary and change; measurements are sparse, costly, and sometimes too slow to collect; and important feedback, such as comfort or ease of assembly, is qualitative. These limitations make it hard to identify failures, diagnose their causes, and feed reliable information back into the next design cycle. The argument is that hardware progress will therefore be uneven: models may master machine-readable design tasks well before they can reliably manage manufacturing and real-world iteration. Even one-shot electromechanical designs would not eliminate the need to solve procurement, production, testing, and feedback challenges separately.