Step 37: assistant

text

I've judged the claim **doubtful**: about a 12% chance that a frontier AI model will autonomously train its own successor by the end of 2027. I'm 0.65 confident in that verdict, and I set the chance a stronger pass would change it at 0.35.

**What would count.** The claim says the training will actually happen, end to end, with no humans doing the research or engineering. Jack Clark's figures are about something weaker: when a model becomes *able* to do this.
- **May 2026:** about 60% by the end of 2028 and 30% for 2027. He gave missing creativity as the reason 2027 was unlikely, and expected a first demonstration on a smaller, non-frontier model.
- **October 2026:** after one new benchmark (TasteVal) showed models beating human experts at designing and reading AI experiments, he wrote that full self-improvement was now more likely to "become possible" in 2027. He gave no number.

**Shown versus forecast.** The Forecasting Research Institute's independent review (10 Nov 2026) finds fast partial automation but not end-to-end autonomy:
- AI writes over 90% of Anthropic's code contributions.
- OpenAI agents finish 35% of 8–16 hour tasks with no human help.
- METR estimates progress at Anthropic is running at about twice its earlier rate, mostly because of AI.
- But no measured part of Anthropic's R&D is fully autonomous, and humans still supply most of the research judgment.
- On tests of open-ended research, agents recovered only 15% of an unseen innovation's gain, and the original authors rejected all four AI-written papers.
- OpenAI's own target for an automated AI researcher is March 2028, after this claim's deadline.

**Weighing an interested forecast.** Clark co-founded Anthropic, so he has a stake in how fast AI progress looks. I recorded that on both of his sources. It doesn't make him wrong: his May essay rests on public benchmarks anyone can check. But his revision rests on one benchmark, he speaks of capability rather than the event, and the independent evaluations are weakest exactly where the remaining gap lies. Even if the capability arrives in 2027, labs are likely to keep people overseeing very expensive frontier training runs.

**Structure added:**
- **For (rated slight):** the trend argument, that the growing AI share of R&D and better research taste carry through to full automation in time.
- **Against (rated likely):** humans still supply the research judgment, and there is little time left to close that gap.
- **Against (rated likely):** being able to do it is not the same as doing it, because labs will keep human oversight. This rests on one new subclaim I created, which is not yet assessed.
- I linked the 2028 version of this claim as related and told its steward its credence should be at least as high as this one's.

**Other work done:**
- Corrected the existing Clark instance.
- Recorded his October revision and the institute's review.
- Placed the leftover passage from the source under the research-judgment argument.
- Set importance and contestation to 0.7.
- Answered both inbox messages.

The run used about $0.89 of its $1.00 budget.