Thursday, Nov 5 | --:--
Back to home

OpenAI Says GPT-6 Astra Designed a Device in Under a Week, but It Still Missed Manufacturing Tolerances

OpenAI said on November 3 that two staff used GPT-6 Astra to design a device for talking to Codex, covering the enclosure, circuit board, firmware and interface, in under a week as a side project, with no direct human edits. A contract manufacturer built prototypes in under a month, but the design needed several rounds of revision because Astra did not consistently account for production limits and tolerances. OpenAI also announced a research collaboration with Mazda.

Times of AI Desk 3 min read San Francisco View as Markdown
Gritty linocut collage of a slate-blue blueprint sheet of plain construction lines folding up into a slim black handheld device with a vermilion seam, with a small torn-paper clipping of the OpenAI logo.

OpenAI's hardware demo is telling for what went wrong: the model could design a device end to end, but not one a factory could build without several rounds of human review. In a post on November 3, OpenAI said it tested whether GPT-6 Astra could create a device "without any direct human edits," using it to define requirements and design the enclosure, printed circuit board, firmware and software for a gadget that lets users talk to Codex.

What Astra did, and what it missed#

Two team members designed the device with Astra in under a week as a side project, and a contract manufacturer built a small batch of prototypes in under a month, OpenAI said. Getting it ready for manufacturing still took several rounds of review, because Astra "did not consistently account for the manufacturer's production limits, tolerances, or details required for assembly," the company wrote.

OpenAI said its hardware and infrastructure teams use Codex and ChatGPT for CAD models, manufacturing drawings, simulations and design reviews, and that their model usage, measured in tokens, has grown 23-fold since February. On BenchCAD 2.0, which tests whether models can reconstruct 3D parts and assemblies and identify connections on circuit boards, an unnamed newer internal model averaged 0.875, against 0.594 for GPT-6 Astra at its highest reasoning setting. The Decoder noted that OpenAI has shared no details about that internal model.

Outside partners#

OpenAI said it is expanding work with Chip Ganassi Racing on turning 3D scans of parts into editable CAD models and drawings, and announced a research collaboration with Mazda's aerodynamics and crash-safety engineers. It is asking engineering firms and software vendors to work with it.

Startups are chasing the same market: Dulo raised a $130 million seed last week to build foundation models for hardware design.

Limits#

  • The device test, the usage growth and the benchmark scores are OpenAI's own; BenchCAD 2.0 results have not been independently reproduced.
  • OpenAI has not said whether the device will be sold or how many design iterations humans reviewed.

Sources#

Prior Coverage

Earlier Times of AI reporting on this thread.

Scroll to continue reading