OpenAI Just Published the Number Chip Fabs Stopped Trusting in 1986
OpenAI says its AI now works 3.1 days for every day a human researcher works. Why would a chip factory refuse to put that number on the wall?
In plain words
OpenAI says its AI works 3.1 days for every human day. A chip factory would not put that number on the wall.
The number counts hours the machine ran. It does not count what came out.
OpenAI said so itself: it is a runtime ratio, not a productivity number.
Chip factories stopped trusting running time in 1986. A standard called SEMI E10 split a machine's day into six states, because running hid too much.
In 1999 another standard added OEE: how often the tool is available, how fast it runs, and how much of its output is good.
A busy factory usually lands at 50% to 70% on that scale. Busy is not the same as productive.
Picture it. A taxi with the meter running all day has not necessarily carried a single passenger.
Why it matters to you. For a week, log three numbers per AI run: minutes it ran, minutes you spent fixing it, and whether it shipped. The last one is the only one that counts.
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.
- OEE Overall Equipment Effectiveness
- One score for a machine: how often it is available, how fast it runs, and how much of what it makes is good.
The number
AGENT-WORKDAYS PER HUMAN DAY · OPENAI · AUG 2026
What to do Monday
for one week, log three numbers per agent run: minutes it ran, minutes you spent fixing it, shipped yes or no. Shipped divided by runs is your quality factor. Take that to Monday's meeting, not the hours. Save this.
The honest caveat
OpenAI called the ratio a runtime number itself, so this is a reading of how the number travels, not a claim that anyone hid anything. An OEE score for an AI agent does not exist yet, and a fair one would be hard to define.
Your turn
When an AI dashboard shows hours at your company, who is the person who asks for the yield?
Argue with me on LinkedInSources
The run's YouTube description carried no source list; these entries were recovered from the run's thesis file (MES_Thesis_running-is-not-producing.docx), which recorded the sources the piece was written from.
- PrimaryAgent runtime per human workday across OpenAI's research org, mid-Aug 2026 (3.1 agent-workdays per human workday) — OpenAI, "Research acceleration: the view inside OpenAI", 6 Sep 2026
- PrimaryDefinition of an agent-workday (8 hours of machine runtime → 24.8 agent-hours per 8 human hours) — OpenAI post (via MLQ, CellCog)
- PrimaryWhen agent runtime crossed human hours (Below parity before June 2026; >3 workdays per human day by mid-Aug) — OpenAI post (via MLQ)
- PrimaryInference spend per researcher, at API prices (Median >$600/day; 90th percentile >$7,000/day) — OpenAI post; IT Pro; Unite.AI
- AnalystRise in the median bill (~40-fold in under five months; typical researcher runs ~4 agents in parallel) — Tomasz Tunguz, 8 Sep 2026 (his reading of the post)
- PrimaryExperiments per active experimenter (All-time high in Aug 2026 since tracking began Jan 2025; OpenAI notes compute also grew) — OpenAI post (via MLQ)
- PrimaryOpenAI's own caveat (Runtime ratio, not a 3.1× productivity gain; agents may shift limits to steering, compute, review, experiment selection) — OpenAI post (via MLQ headline and body)
- PrimarySEMI E10 origin and structure (First published 1986; six mutually exclusive states: non-scheduled, unscheduled down, scheduled down, engineering, standby, productive) — SEMI E10 standard page; SEMI "Update on SEMI E10"; PEER Group
- PrimarySEMI E79 origin (E79-0299 first published 1999 (E79-0200 revision 2000); defines OEE and throughput metrics) — SEMI store listing; ResearchGate E79 figure
- PrimaryE79 OEE structure (Availability efficiency × operational efficiency × rate efficiency × quality efficiency (spoken as availability × rate × quality)) — SEMI E79 (ballot doc 6579 not fetchable this session)
- PrimaryISO 22400-2 OEE (Availability × effectiveness × quality ratio (three factors, 34 KPIs)) — ISO 22400-2:2014
- As citedFab OEE ranges (SEMATECH-era: ~30% of theoretical capacity; later commentary ~50% logic / 60–70% memory; a SEMI-cited 65–75%; ITRS asked 75% (→88%) for bottleneck tools) — SemiWiki forum; fabrico blog; OEE→OFE paper (ScienceDirect)
Read the full script (231 words)
Keep reading
All piecesAI Just Asked to Slow Down. Chip Makers Wrote the Rule in 1973.
The AI companies just asked permission to slow down. On 12 September 2026 Dario Amodei, who runs Anthropic, the company that makes Claude, published an essay asking the whole industry to move one to two years slower so the people who check these systems can catch up, and he offered outside checkers the same access…
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.
OpenAI's Data Agent Reads 7 Databases. Not One Is on a Factory Floor
Every number on a company dashboard is a copy. The original was born inside a machine, and it lost something on the way.