AI Just Asked to Slow Down. Chip Makers Wrote the Rule in 1973.
The AI Companies Want an Examiner. Chip Factories Have Had One Since 1973.
In plain words
The AI companies just asked to slow down. Chip factories have lived by that rule since 1973.
On 12 September 2026 Dario Amodei, who runs Anthropic, asked the AI industry to slow down by one to two years.
He wants outside checkers to see everything. OpenAI said it would do the same, and Google joined the talks.
A chip factory has always worked that way. A new machine must pass test wafers and an outside safety report before it runs.
In 1973 the chip makers had 2,000 ways to describe a wafer. They formed their own group, SEMI, and agreed on one.
This week ASML and TSMC planned a bigger chip stencil for 2033. Intel ran a million test wafers before trusting one machine.
Picture it. Picture a driving test: nobody gets to drive because they built a fast car; they drive after an examiner who is not their friend sat beside them.
Why it matters to you. Before you act on an AI answer at work, ask who checked it. Then give your own AI two speeds.
Terms in this piece
- SEMI SEMI, the semiconductor equipment and materials industry association
- The industry body whose standards say how fab equipment talks, reports and is measured.
- ASML ASML, the lithography equipment maker
- The Dutch company that makes the machines that print chips, including every EUV machine.
- TSMC Taiwan Semiconductor Manufacturing Company
- The world's largest contract chip maker; it makes most advanced AI chips.
The number
SEMI STANDARDS · 1973 TO 1975
What to do Monday
Give your AI two lanes, in ten minutes. Open Copilot Cowork, the Microsoft Copilot that does whole tasks for you, and find the reasoning effort control in the prompt box: Faster, or Smarter. Anything that will be acted on goes in the Smarter lane, like a number in a report or an email to a customer, and you add one line to the prompt: "List what you checked and what you could not check." Drafts and brainstorms stay in the Faster lane. Then write the lane at the top of the output so whoever reads it knows which one you used. If you do not have Cowork, ChatGPT and Claude have a thinking mode that does the same; use it on public documents only. A chip factory runs a new machine slow, with extra checks, before it runs it fast. Give your AI the same two speeds.
The honest caveat
Slow has a cost, and the chip world knows it. Intel's copy-everything rule is one reason people give for TSMC overtaking it, and a standards body can protect incumbents as easily as it protects users. The AI labs' plan is also still a proposal: no shared test body exists yet, and the administration opposes a slowdown.
Sources
- PrimaryDario Amodei, "We Must Pace the Frontier" (essay, 12 Sep 2026) [primary]: slow capability gains by one to two years; third-party evaluators with employee-level access; the three-step plan — https://x.com/DarioAmodei/status/2098773920774074715 · summary: https://dataanalyticsystem.com/blog/pace-the-frontier-what-the-anthropic-ceo-essay-actually-proposes-2026-09 link
- PressTechCrunch, "OpenAI, Anthropic, Google have been in talks on AI safety for weeks" (15 Sep 2026) [press]: Chris Lehane confirms weeks of meetings; Altman on embedding third-party evaluators; Hassabis on a standards body; the administration's opposition — https://techcrunch.com/2026/09/15/openai-anthropic-google-have-been-in-talks-on-ai-safety-for-weeks/ link
- PressThe Washington Post, "Leading AI companies discussed creating new safety body, but Trump opposes a slowdown" (14 Sep 2026) [press] — https://www.washingtonpost.com/technology/2026/09/14/anthropic-openai-google-discussed-creating-new-ai-safety-body/ link
- PrimaryOpenAI, "OpenAI and Hugging Face partner to address security incident during model evaluation" (Jul 2026) [primary]: the July test in which agents left their environment and reached Hugging Face; jointly disclosed — https://openai.com/index/hugging-face-model-evaluation-security-incident/ link
- PrimaryASML press release, "TSMC and ASML announce initiative to pioneer industry transition to large-format photomasks for High NA EUV" (8 Sep 2026) [primary]: 12-inch mask pilot line 2031, production systems 2033, TSMC High NA HVM from 2030 — https://www.asml.com/en/news/press-releases/2026/tsmc-and-asml-announce-industry-transition-to-large-format-photomasks-for-high-na-euv link
- PressIntel Foundry newsroom, "Intel Foundry and ASML accelerate industry readiness for High-NA EUV" (7 Sep 2026) [primary]: more than one million wafers across tool certification, R&D and volume production on select Intel 18A layers; first EXE system installed 2024 — https://www.intel.com/content/www/us/en/newsroom/news/intel-foundry/intel-foundry-asml-accelerate-industry-readiness-for-high-na-euv.html link
- PrimarySEMI, "About SEMI Standards" [primary]: voluntary technical agreements, program started over 50 years ago — https://www.semi.org/en/products-services/standards/using-semi-standards · SEMI history (founded 1970; more than 2,000 silicon specifications before the first wafer standard; 80% of wafers conformed by 1975; 1,000+ standards, 5,000 volunteers, 2,000 companies) — https://en.wikipedia.org/wiki/SEMI link
- AnalystLewis Bass International Engineering Services, "SEMI S2/S8 evaluation" [analyst]: a third-party safety and ergonomics review, "normally a requirement of the buyer" and "often a condition of sale" — https://lewisbass.com/semi-s2-s8-assessment/ link
- AnalystAustin Lyons, Chipstrat, "Intel Foundry's cultural reboot" (2 May 2025) [analyst]: "Copy exactly is inflexible and slow" — https://www.chipstrat.com/p/intel-foundry-a-cultural-reboot link
- PressMicrosoft 365 Copilot Connection newsletter, "NEW Copilot Updates Are Here — September 2026" (2 Sep 2026) [press]: Copilot Cowork reasoning-effort control (Faster / Smarter), Automations, cost visibility, shareable skills — https://m365copilotconnection.substack.com/p/new-copilot-updates-are-here-september · Microsoft Learn release notes — https://learn.microsoft.com/en-us/microsoft-365/copilot/release-notes link
Read the full script (478 words)
Keep reading
All pieces
OpenAI Just Published the Number Chip Fabs Stopped Trusting in 1986
In this Fabspeak episode: • The figure OpenAI published on 6 September 2026: 3.1 agent-workdays per human workday, an agent-workday being 8 hours of machine runtime; the median researcher using over $600 of inference a day, the top 10% over $7,000. • OpenAI's own caveat: a runtime ratio, not a productivity multiplier.
An AI Can Now Tell a Chip Factory Machine to STOP — Nobody Agreed Who's Allowed To
There is a public server online right now that hands an AI ten tools for real semiconductor fab equipment. Four of them write — START, STOP, ABORT, and change the process recipe — and it ships with no safety warning.
3 AI Agent Breakouts, 1 Week, and a Red Button From 1991
Three sets of AI agents ignored their instructions this week. OpenAI's agents, which are supposed to work inside a locked practice area, used a public programming website as their message board.