OpenAI demonstrated how AI independently designed a device—from casing to firmware
The Astra model developed a gadget for voice access to Codex without direct human edits. After prototype testing, OpenAI identified the model's limitations in accounting for manufacturing requirements.
The company reported testing AI in the full cycle of hardware design—from appearance to software components. Astra created the casing and internal mounts, selected electronic components, designed a printed circuit board considering the device's dimensions, and prepared the interface and firmware. Two OpenAI employees supervised the project alongside their primary work, while a contractor manufactured a small batch of prototypes in less than a month.
During preparation for production, the model did not fully account for manufacturer constraints, permissible deviations, and assembly requirements. The company believes this outcome helped identify areas for improving future models.
Engineers use Codex and ChatGPT in computer-aided design systems for preparing drawings, configuring simulations, and verifying projects. Since February, token consumption by employees for this purpose has increased 23-fold. In the BenchCAD 2.0 test, Astra scored 0.594, while a newer internal model achieved 0.875.
The company is also testing technologies with Chip Ganassi Racing: 3D scans of parts are converted into editable CAD models and 2D drawings for production. With Mazda engineers, AI applications are being explored in the creation and evaluation of automotive designs, including airflow modeling and component behavior.
