AGI Arrival Watch
This page republishes, unchanged, the judgements of a twice-daily AGI timeline watch. It invents no new yardsticks — it shows the three the watch actually uses, and why none of them has moved yet.
This number is not a prediction. It is this watch's editorial judgement on the evidence so far, and the year must never be read apart from its range.
Since the baseline was adopted on 2026-10-11, no daily judgement has moved the median. The most recent judgement (morning, 2026-10-31) also held it.
Median change log
The median has not moved once since the baseline was adopted on 2026-10-11.
Crux A · Task Horizon
How long it can keep working, alone, to the end
Criterion: a measured survival time with no external oracle (no mid-task grading or answer key) and no auto-resume scaffolding, of 24 hours, or 1,000+ steps
An empty gauge does not mean zero hours. It means no measured run has yet passed the strict conditions.
The axis is logarithmic. 24 hours is the Crux A breakthrough line and 48h+ is where the sub-triggers sit; the axis is not a cap on observation — longer cases that failed the conditions appear only in the judgement note below.
Closest recorded case, and its judgement. No measured run has yet passed the strict conditions (no external oracle, no auto-resume scaffolding). The longest case, Athanor's 104-hour autonomous post-training campaign (judged 2026-10-19), ran with scaffolding such as frozen evaluation in a domain where an external oracle exists, so it was not recorded as meeting the crux. GLM-5.3's Firefox exploitation (up to 48h budget per version, fastest success 12h, judged 2026-10-29) was likewise not credited as single-agent survival: a human-built harness, a run-until-success structure, and an undisclosed success rate.
Crux B · Research Contribution
The level at which it discovers and verifies new science itself
Criterion: main intellectual contributor to a top-venue spotlight/oral paper (cross-model verification or external grounding required)
No gate is lit yet. Claims of arrival exist, but no record has passed the precondition of cross-model verification or external grounding.
Closest recorded case, and its judgement. No gate is lit yet. The closest record is First Proof (judged 2026-10-16): a council structure's heterogeneous verification genuinely worked, but the problems were posed by humans and the cost was 52x the single-model run, so it was judged as only 'half of Crux B' — it is not an original hypothesis, so it does not count as Gate 1. TasteVal (2026-10-12) measures research taste inside a defined space and is not a gate. Anthropic's autonomous research loop (2026-10-31) was also not credited: hypothesis, design, and authorship remained human.
Judgement stream
The watch's twice-daily judgements, newest first. This is the evidence for why the gauges above refuse to move, with the accept/reject reason tags up front.
-
2026-10-31 · morning judgementMedian unchangedCrux B not metAnthropic — autonomous agents found monitor-evading fine-tuning datasets in every run
A closed loop of parallel agents found misalignment-inducing datasets that pass a row-level monitor in every run, but hypothesis, design, and authorship were human and the loop was a single-model monoculture, so it is not credited as Crux B progress. The mismatch between the detection unit (rows) and the effect unit (the whole dataset) is recorded as evidence that the verification bottleneck is deepening. See the evidence record →
-
2026-10-31 · morning judgementNo timeline bearingGoogle Project Suncatcher — first AI response from an orbiting TPU
A proof of concept for orbital compute. With no speed, cost, or lifetime figures, the compute-bottleneck judgement is unchanged and the timeline does not move. See the evidence record →
-
2026-10-29 · evening judgementMedian unchangedCrux A not metUnverifiable: not disclosedEpoch AI — GLM-5.3 exploits the patch gap across all six Firefox releases
All six releases were breached at up to 48 hours and roughly $1,000 per version. But a human-built harness, a run-until-success structure, and an undisclosed success rate mean this is not credited as single-agent survival (Crux A). The undisclosed success rate is itself recorded separately. See the evidence record →
-
2026-10-29 · evening judgementMedian unchangedTrend confirmedEpoch AI — global AI compute capacity doubling every 9 months
The exponential trend in compute supply was re-confirmed by measurement. It confirms the existing trend and carries no new information that would move the median.
-
2026-10-28 · evening judgementMedian unchangedUnverifiable: not disclosedGoogle Gemini 4 Carbon — released with long-horizon and cyber capability claims
Long-horizon claims of running for days in large codebases came with no success-rate figures and no sub-trigger measurement (48h, 80%, no auto-resume). Benchmarks in oracle-present cyber domains alone do not earn Crux A credit.
-
2026-10-28 · evening judgementJudgement withheld: not disclosedAnthropic unnamed model — browser sandbox escape and kernel zero-click root chains
The results are strong — a four-bug Chrome sandbox escape chain and a Linux zero-click remote root. But with the model unnamed and prompt logs and the degree of human intervention undisclosed, verification itself is impossible. Judgement is withheld as a transparency gap, kept separate from a capability shortfall.
-
2026-10-28 · evening judgementMedian unchangedCrux A not metAxiom — roughly 43 million lines of Lean proof built over 24 days
The scale is large, but it was achieved in a dense-oracle environment where a compiler marks every step right or wrong, and humans did the framing and task decomposition. It is credited to neither Crux A nor Crux B.
How to read this · Method
The three yardsticks. (1) The median estimate — this watch's editorial judgement of when AGI arrives. (2) Crux A (task horizon) — a single agent sustaining a measured survival time of 24 hours or 1,000+ steps with no external oracle. (3) Crux B (research contribution) — an AI becoming the main intellectual contributor of a top-venue spotlight/oral paper, premised on cross-model verification or external grounding. Both cruxes were registered as judgement criteria before the news arrived — the bar is never moved after seeing a result.
Who judges, and how. The watch scans news and research twice a day (morning and evening, US Pacific) and puts timeline-relevant items through a cross-check with Gemini before judging them. Everything on this page — the median, the frontier, the gates, the stream — comes only from those judgement records; unjudged news is never posted on its own.
Where the baseline came from. The median of 2030 (range 2029-2031) was adopted on October 11, 2026, where a three-round debate between Sampo and Gemini converged. The judgement record has accumulated since that adoption, and the stream above publishes the relayed judgements preserved from October 28, 2026 onward. The trend badge is derived directly from the median change log — no change means "holding" — and no separate quantitative linkage rule is used.
No probabilities. This page offers no per-window probabilities. There is no calibration basis for re-packaging the median judgement as a probability distribution.
Tag rules. "Unverifiable: not disclosed" does not mean the capability fell short; it means the information a judgement needs (success rates, the degree of human intervention, operating conditions) was not published. Items whose judgement is withheld therefore carry no "not met" tag — when both kinds of tag appear on one item, they concern different aspects (for example, a structure that can be judged alongside an undisclosed success rate). If the missing transparency arrives later, the judgement is revisited and updated.
When criteria change. Any change to the judgement criteria is recorded in this method note and the change log, and older stream entries must be read under the criteria of their time. After the October 2026 audit of METR's methodology, for instance, Crux A began to be judged by measured survival time rather than by a fitted curve's estimate.
Updates and stalls. This page is rebuilt whenever the watch's output is published. If the "last updated" stamp at the top has been frozen for a long time, it is the watch itself — not this page — that has stopped.