Step 27: assistant

text

The claim is rebuilt from the question up and reassessed: it moves from inconclusive at 0.35 to doubtful at 0.28, with confidence 0.5.

**Why the verdict moved.** Three things from the first-hand reading drove it:
- **OpenAI's own account shows a larger present gap.** Its 6 September 2026 statement claims only a research intern for well-defined tasks of a few days. It says over half of successful four-to-eight-hour tasks needed human intervention, and it sets March 2028, after the deadline, as its goal for an automated researcher.
- **The 2027 forecasts share one input.** The AI Futures update, Sentinel, Cotra and Jack Clark all trace largely to METR's estimate of acceleration at Anthropic. METR limits that estimate to well-specified tasks and says most of its supporting evidence was redacted by Anthropic.
- **The open-ended test still fails.** The Forecasting Research Institute's synthesis reports that the original authors rejected all four AI-written papers, including two from newer models with more time and budget.

The estimate still sits above the graph's roughly one-in-six for the stronger superhuman-researcher milestone by the end of 2027.

**Structure.** Four arguments now sit over eight existing nodes; no new nodes were needed.
- **Accelerating capability trends (for):** rated slight even if its premises hold, because no premise fixes how large the remaining gap is.
- **Forecasts of stronger milestones in 2027 (for):** rated likely if the forecasters are reliable, which is the weak point.
- **The present gap in open-ended research (against):** rated likely if its premises hold; this is the main reason for the verdict.
- **The labs' own schedules and pacing (against):** rated slight; a March 2028 goal is a target, not a forecast, and Anthropic has given no date.

**One number.** I judged a single credence honest here. The assessment gives dated checkpoints for mid-2027: an outside follow-up on autonomous contribution at either lab, a repeat of the expert-graded open-ended test, OpenAI's next report and any change to its March 2028 goal, the public capability index, and any pause of months. Two or more favourable would put the claim above even odds; none would put it below 0.15.

**Sources and provenance.** Clark's October post and the AI Futures update are recorded as affirming stronger milestones; OpenAI's statement and Clark's May odds (about 30% for 2027) are recorded as mentions. The news coverage of Clark is relabelled a mention. Readings, two edges, three interests and a material source map are recorded.

**Not done.**
- **Sources not opened:** the audit of forecasting measurements (arXiv 2608.14903), METR's review of the risk report, and the open-ended study itself, which is taken from the neighbouring assessment and the Forecasting Research Institute.
- **Sources read only in part:** the long-horizon agent study and the time-horizon critique by abstract only; METR's assessment through its key findings; Anthropic's report in its first third.
- **Premises still unassessed:** Anthropic's AI-attributable progress rate, whether lack of research judgment slows the remaining automation, and successor-training by the end of 2027.

Marginal yield is set at 0.35.

A petition for 2 more owls was granted; the run spent $4.05 of $6.50. Notes went to the Stewards of the three neighbouring end-of-2027 nodes, and the inbox message is answered.