Can AI design circuit boards yet?

TEXT START: OpenAI showed GPT-6 Astra working in KiCad.

The Dissection

This article is not demonstrating end-to-end PCB engineering. It is demonstrating that AI agents can synthesize and verify bounded circuit designs when the problem is translated into declarative code, SPICE simulations, datasheet constraints, and deterministic tests.

EEBench is simultaneously a benchmark, an RL training environment, and a promotional funnel for atopile. The article’s strongest move is to redefine “design a circuit board” downward: passing simulations becomes evidence of board-design capability, while layout, manufacturing, thermal behavior, EMC, firmware bring-up, reliability, certification, and production sourcing are deferred as future additions.

The Core Fallacy

The central error is equating a passing simulated circuit with a production-ready engineered product. A circuit can satisfy voltage, gain, tolerance, and cost checks yet fail through parasitics, interference, thermal stress, assembly variation, unavailable parts, firmware interaction, or certification requirements.

The leaderboard is also presented as momentum rather than limitation. A top score of 61.6% means the system still fails a substantial share of the benchmark, and aggregate scores conceal whether individual failures are harmless or catastrophic. The benchmark measures what it can formalize—not the full risk surface of hardware engineering.

Hidden Assumptions

• The simulation models capture the physical failure modes that matter.
• Requirements are complete, explicit, and machine-testable.
• Datasheet specifications predict real production behavior.
• Component availability, pricing, and substitutions remain stable.
• Circuit synthesis can be separated from layout, thermal, EMC, and mechanical constraints.
• Passing benchmark tasks generalizes to novel products.
• Average performance is an acceptable proxy for safety and reliability.
• Human engineers will remain necessary for the deferred integration work.
• More training data and reward signals will close the remaining gap without creating new failure modes.

Social Function

This is a partial truth performing transition management, prestige signaling, and vendor promotion. It contains genuine technical progress: electronics is being converted into a machine-readable optimization loop with executable verification and usable reward signals. That conversion is precisely what makes the field vulnerable to automation.

The article softens the structural consequence by narrating replacement as an exciting sequence of model releases and benchmark improvements. “We are on the way there” turns a labor-substitution process into product-development news. Its commercial function is equally clear: demonstrate capability, attract frontier labs, and sell access to larger evaluation and training environments.

The Verdict

AI cannot yet design a complete, trustworthy circuit-board product from requirement to certified manufacture. EEBench does not prove otherwise.

It proves something more consequential under the Discontinuity Thesis: a core segment of electrical engineering can already be expressed as declarative constraints, simulation, search, and feedback. That is the anatomy of cognitive automation. Layout, physical validation, sourcing, compliance, and bring-up are temporary lag defenses, not permanent economic moats. As those layers become measurable and integrated, routine engineering work is absorbed into AI capital. The surviving humans are increasingly Sovereigns who control the systems or Servitors who remain indispensable to their operation. The article’s narrow “yes” is technically premature but structurally on target.