Designing the Physical World with AI

The conversation opens with a framing of the core tension: AI can generate designs and run simulations at near-zero marginal cost, but the physical world — construction, manufacturing — remains slow and fragmented. Aaron Price-Wright sets up the discussion by contrasting 'moving atoms vs. bits' and introduces his two guests as working on physical AI at opposite scales: macro (large infrastructure projects) and micro (circuit board design and manufacturing).

Alex Modin of Unlimited Industries describes construction as an industry where a single large project (power plant, hospital) can take over a year just for the design phase, involving hundreds of engineers across mechanical, electrical, civil, and structural disciplines. His company's goal is to automate this pre-construction phase end-to-end: feed in a site and requirements, and an AI generates a globally optimized 'issued for construction' package across tens of thousands of design permutations. He frames this as a parametric, code-first approach where everything is a variable — meaning late-stage design changes no longer cascade into project restarts. He predicts full construction automation within a decade, contingent on proper incentive alignment.

Davide Asnaghi of Diode Computers focuses on circuit board design and manufacturing. He argues that while manufacturing robots already exist and handle ~80% of PCB assembly (via surface mount technology), the remaining 20% — chunky components, board-to-enclosure assembly — has historically required manual labor. His company's bet is that if AI can generate designs that are inherently manufacturable (design-for-manufacturing, or DFM), then full automation becomes achievable today without waiting for better robotics. He predicts full automation of a relevant subset of electronics design within two years.

Both founders converge on a shared architectural philosophy: reframe physical design as a code problem. Asnaghi explains that Diode built a compiler that makes circuit board design feel like writing Python — leveraging the vast training data models already have on code rather than the sparse data available on schematics. Modin echoes this, describing Unlimited's approach as model-led, where AI agents write code within a structured ontological framework that encodes engineering relationships and constraints.

On the question of industry adoption, both acknowledge deep resistance. Construction's incentive structure — driven by project finance investors who want stable IRRs and zero downside risk — actively suppresses technology adoption. Modin argues that vertical integration is necessary to present a clean interface to the industry rather than trying to change one small piece at a time. Asnaghi sidesteps the EDA software market entirely by selling the end product (faster, cheaper manufactured boards) rather than the tool — making the AI an implementation detail invisible to customers.

The data scarcity problem is discussed at length. Asnaghi notes that PCB design data is siloed in companies like Apple and SpaceX and unlikely to be shared. His solution is to bootstrap from code-first validated design blocks, use simulation as a training-time tool rather than inference-time, and generate data organically by becoming the open-source infrastructure layer that designers use for free. He expresses personal confidence that data is the last frontier, while noting his co-founder Lenny believes Monte Carlo / RLHF-style self-play may render additional data unnecessary.

On humanoids and robotics, Asnaghi takes an egalitarian view — all robots have PCBs, so he loves them all — while expressing particular bullishness on Vision-Language-Action (VLA) models improving robotic arms for the remaining 20% of assembly work. Modin sees humanoids as valuable for construction but expects a mix of general and purpose-built robots, citing volume-to-surface-area scaling laws as a reason specialized equipment will persist.

The conversation closes on the societal stakes. Asnaghi argues the second-order effect is democratizing hardware creation the same way SaaS has democratized software — enabling any engineer or teenager to spin up a hardware company. Modin points to a deeply troubling trend: U.S. construction labor productivity and inflation-adjusted CapEx numbers have been declining for 50 years, a trajectory that threatens America's ability to build the data centers, advanced manufacturing facilities, and critical mineral infrastructure needed for AI leadership and re-industrialization.