Does Amodei's AI slowdown change the 2027 AGI odds?
What was actually proposed
The essay is a three-part plan, and only the first part is something a single company can do alone. Summarised rather than quoted — the full text is linked at the bottom, and you should read it there:
| Step | Who has to act | Status as of 2026-09-14 |
|---|---|---|
| Give third-party evaluators permanent, employee-level access | One company, unilaterally | Committed — Anthropic, and OpenAI said it would match |
| Common safety standards and limits on unchecked capability gain among labs in democracies | Several companies together | Proposed only |
| Narrow prohibitions agreed with authoritarian governments, starting with bioweapons | Governments | Proposed only |
Step one is real and verifiable: an outside organisation with a badge and a laptop either has that access or it does not. Steps two and three are, today, a proposal — and the honest way to record a proposal is as a proposal.
The industry now says it will slow down. The scorecard will say whether it did.
Eight predictions, each with a published flip condition. One email on the day one of them actually moves — nothing in between.
Tell me when a verdict flips →Which of the eight predictions this bears on
The scorecard grades eight claims from Situational Awareness. A deliberate industry slowdown would be the first endogenous reason for the fast-timeline predictions to fail — every previous candidate was a technical wall. Here is where it touches, and what would actually have to change:
| Prediction | Verdict today | What a real slowdown would look like here |
|---|---|---|
| Compute scales ~0.5 OOM/yr | On track | Effective compute growth falls below trend for consecutive quarters — not a pledge, a measured rate |
| AGI by 2027 | Open | Frontier releases are spaced further apart and capability benchmarks flatten with them |
| Intelligence explosion 2027–29 | Pending | Recursive self-improvement is the essay's stated trigger; a limit that binds would show up as slower model-on-model gains |
| Trillion-dollar capex | Exceeded | Capex guidance cut, not merely capability pacing — no lab has signalled this |
| Outpace grads on knowledge work | On track | Unaffected: pacing future training does not retract deployed capability |
| Open source fades | Wrong | Unaffected — and a pact among frontier labs does not bind open-weight releases, which is the gap in step two |
The part worth arguing about
A pacing agreement among a handful of labs in democracies has an obvious hole, and it is the same hole that already made one of these eight predictions go the other way: open-weight models are released by parties who are not at the table. If the frontier paces itself and open weights keep improving, the gap narrows from below rather than the frontier stalling from above. That is not an argument against pacing — it is the reason step two, on its own, cannot be scored as a slowdown.
What would move a verdict
Pre-registered, so it can be checked later rather than rationalised:
- Evaluator access becomes observable — a third party publishes findings from that access. Verifies step one; changes no verdict by itself.
- Release cadence stretches — the interval between frontier releases grows and benchmark gains flatten with it, for two consecutive quarters. This is what would put AGI by 2027 under pressure.
- A named capability limit is published and held by more than one lab. Without a number, step two is unscoreable.
None of the three has happened. Until one does, this is a change in what the industry says about itself, which is worth reading, and not yet a change in what the evidence says.
Sources
Dario Amodei, We Must Pace the Frontier, published September 12, 2026 — read the essay in full. Sam Altman's same-day endorsement was posted on X. This page deliberately paraphrases rather than quoting: go to the essay for his exact wording. Verdicts and the Thesis Tracker reading come from this site's own machine-readable dataset (CC BY 4.0).
Frequently asked questions
Not yet. As of September 14, 2026 none of the eight graded predictions moved and the Thesis Tracker still reads 62.5/100. The essay states an intention to pace capability growth; the scorecard grades measured evidence, and no measured rate has changed.
A three-part plan published September 12, 2026: give third-party evaluators permanent employee-level access; agree common safety standards and limits on unchecked capability gain among labs in democracies; and negotiate narrow prohibitions with authoritarian governments, starting with bioweapons.
Sam Altman endorsed the essay the same day and said OpenAI would also give independent evaluators employee-like access. That matches step one, the unilateral part. Steps two and three remain proposals.
No measurement shows it. Compute is still scaling at roughly 0.5 orders of magnitude per year and capex has exceeded the original forecast. A slowdown would show as release cadence stretching and benchmark gains flattening together, for consecutive quarters — which has not happened.
Because every verdict has a published flip condition tied to observable evidence. If statements of intent could move the score, the score would track press releases rather than capability, and its history would stop being auditable.
The live scorecard updates as models ship and verdicts change.
View the live scorecard →