How Hidden AGI watch works
Three questions, three kinds of number, one rule-based alarm. Every day a custom AI agent checks whether advanced AI could exist, or be acting, without the public knowing, and publishes sourced probabilities and a fire-alarm level.
How our numbers fit together
Three questions, three kinds of number. Is AGI here? is a capability question: for each of eight parts we compare the best public result to its bar, and from that we judge the chance that AGI exists anywhere. Could it be hidden? is a secrecy question: A, C and D each mean "AGI exists and was kept quiet" (or acted, or shaped a government decision), so they can never exceed AGI anywhere; B, hidden self-improvement, needs no AGI, which is why the Index can sit above it. Should I worry today? is a rules question: the fire alarm moves only on published conditions, never on our odds.
1 · Is AGI here? capability
- EvidenceMETR time horizons, Remote Labor Index, ARC-AGI, OSWorld, CL-Bench, Epoch's open problemscapability gap (our gauge, ~3 months): how far unreleased models may lead
- AGI parts0 of 8 met4 partial, 4 far (definitions v2.0)Reaches the top when one system meets all eight.
- AGI anywhere, public or hidden1% today36% by end-2030 · 59% by end-2035Reaches the top when a system meets all eight, public or hidden.
AGI anywhere can be no higher than the chance that its weakest part is met. Today the furthest behind are long projects, reliability, learning. Why only 36% by end-2030 when the time-horizon trend reaches a work-month around Jul 2028: that trend covers one test of one part, METR's suite can't yet measure that length, and generalization and learning have no trend.
2 · Could it be hidden? secrecy
- AGI anywhere (from 1) × kept quiet 30+ days → A, C, DOur judgment of the kept-quiet factor: about 0.6 today.
- Hidden scenarios, todayA 0.6% Hidden AGIC 0.4% Covert AGI actorD 0.2% Covert government influenceB enters on its own: it needs no AGIB 2.5% Hidden self-improvementD-open (0.2%) is open by definition, so it stays out of the Index.
- Hidden AGI Index3% todayA 0.6 + B outside A 2.15 + C or D outside A 0.05 = 2.8, shown as 3Reaches the top when one hidden scenario is proven true.
Why the Index can exceed AGI anywhere: B, hidden self-improvement, doesn't need AGI. The hiding-conditions gauges are context for our odds. The capability gap is the one gauge in a formula (A's), and it sets no alarm condition by itself (W1 and X1 need a measured gap). AI-led R&D and the disclosure lag sit next to conditions W5/X3 and W4; money and delegation feed none.
3 · Should I worry today? rules, not odds
- SignalsTripwires: 4 confirmed, 6 open, 1 quietEscape watch: 0 of 8 confirmed
- Alarm conditions4 metW2, W3, W4, W5 (all Watch conditions)Watch: any two W · Warning: any one X · Alarm: any one Y, proven
- Fire alarm◐ Level 1 · WatchSince Sep 29, 2026Reaches the top when a Y condition is proven beyond reasonable doubt (Level 3).
The alarm never reads the probabilities above. They can move without the alarm, and the alarm can move without them.
Outside view, for comparison only: prediction markets and forecasters mostly settle on a lab announcing AGI, a looser test than ours (Kalshi: 56% that any company announces AGI before Jan 1, 2028, Nov 13; Metaculus “weakly general AI”: community median 17 Aug 2027 (read Nov 13)). Trend dates are measurements extended, not forecasts: METR's 80% horizon reaches a work-month around Jul 2028 if the trend holds.
What each number means
| Number | Unit | Today | Reaches the top when |
|---|---|---|---|
| AGI parts | parts met, of 8 | 0 | one public system meets all eight |
| AGI anywhere | probability: today / by end-2030 / by end-2035 | 1 / 36 / 59 | a system meets all eight, public or hidden; then it stays at 100 |
| A, B, C, D | probability, today | 0.6 / 2.5 / 0.4 / 0.2 | proven to the alarm's Level-3 standard (A through Y1, B through Y3, C and D through Y2) |
| D-open | probability, today | 0.2 | open by definition; no alarm condition |
| Hidden AGI Index | probability that at least one of A–D is true now | 3 | one hidden scenario is proven |
| Hiding-conditions gauges | own units | ~3 months · 26% · ~112 days · $615B / yr · rung 3 of 5 | no top: context for our odds; only the capability gap is in a formula (A's), and none sets an alarm condition by itself |
| Signals | Quiet / Open / Confirmed | 4 confirmed | goes to a ruling on the alarm condition it names (3 of 11 tripwires name none) |
| Alarm conditions | met or not met | 4 met | the level's rule is satisfied |
| Fire alarm | Level 0–3 | Level 1 · Watch | Level 3: proof beyond reasonable doubt |
| Trend dates | a date | work-month ~Jul 2028 | never: a measured line extended, not a forecast |
| Jobs claims | Contradicted to Established | 1 established, 5 emerging, 7 no clear sign | never: it feeds no number above |
What we mean by AGI
Our definition and its eight parts: Is AGI here? →
The hidden scenarios
- A Hidden AGI: AGI kept from the public.
- B Hidden self-improvement: AI builds its successor, undisclosed; needs no AGI.
- C Covert AGI actor: AGI acting unseen online.
- D Covert government influence: AGI covertly shaping a government decision.
- D-open Open government influence: the same, openly; outside the Index.
The Hidden AGI Index
The chance that at least one of A–D is true today. Today 3%. How it's derived →
What "hidden" means
Kept from the public for at least 30 days, by its builders or by the system itself.
More
A–D track two kinds of hiding: advanced AI that people keep from the public (a company or government that doesn't disclose what it has built, or how it's using it), and AI systems acting covertly on their own (C's rogue path). A, B and C use the same secrecy window: kept from the public for at least 30 days. Admitting that an unreleased model exists doesn't count as disclosure.
A model hiding its own capability from its developer, for example by quietly underperforming on tests, is a different problem. It isn't A, because A needs the builder to know what it has, but it would undermine the evaluations these readings rely on, so we track it as the evaluation-integrity signal on the fire alarm's signal board.
The five hiding-conditions gauges
- Capability gap (estimated): unreleased models' lead; in A's formula.
- AI doing AI research (measured): a lab's AI-led share of its R&D.
- Oversight gap (estimated): how long agent incidents stayed hidden; a floor.
- Money trail (measured): Big Tech's capital spending, last 12 months.
- Delegation to AI (assessed): government decisions run through AI, on five rungs.
Context for our odds, not alarm conditions. Values →
The fire alarm
Set by published conditions, never by our odds. Today: Level 1 · Watch. Conditions →
What gets sent
Daily and weekly emails go out automatically; a person approves every alarm alert. More →
How the numbers are made
The method in full rounding, evidence ratings, signals
- Daily: a custom AI agent, built for this project, searches the news, research, lab system cards and independent evaluations such as METR, Epoch AI and the AI Security Institutes. It then reassesses each probability today, by end-2030 and by end-2035, and publishes a short report and a full analysis automatically. The analysis shows the reasoning: base rates, the strongest case on each side, and what would change our mind.
- What the horizons mean: today = true now. By end-2030 / by end-2035 = true at any point before 1 Jan 2031 / 1 Jan 2036 (cumulative, so never below today).
- Rounding: estimates move in 0.1-point steps below 1%, 0.5-point steps from 1% to 10%, and whole points above 10%, rounded half up. Finer steps would claim more precision than these judgments have.
- Evidence ratings: every factual claim links its source and carries one of these chips:
- verified fact a primary source (a filing, a lab or evaluator, a government), or two independent credible outlets that we opened.
- credible report a credible outlet's report not yet confirmed that way, including leaked documents and single anonymously sourced reports.
- expert opinion a named expert's or organisation's judgment or forecast.
- forecast aggregate a crowd or market figure such as a Metaculus median or a Kalshi price; it reports what forecasters think, not a fact about AI.
- our inference our own reasoning from the cited sources, not a claim any source makes; until now inferences carried no chip, and a fact chip followed by '(part our inference)' marks a sentence that mixes the two.
- speculation a possibility no evidence yet supports, labelled so it is never mistaken for evidence.
- Signals: the tripwires and Escape watch indicators, the specific, observable signals that would move the numbers most. Each reads Quiet, Open or Confirmed, and names the alarm condition it feeds where there is one. They move our probabilities; only the alarm conditions set the alarm level.
- Forecast chart: Could it be hidden? draws our stated numbers (today, by end-2030, by end-2035) with outside forecasts for comparison. We don't project our daily line forward. Our end-2030 and end-2035 numbers should move only on evidence, never in a predictable direction; our numbers for today are expected to rise over time along that forecast. The shaded band is a judgment range, not a statistical interval: roughly how far our number could plausibly move as new evidence arrives, wider when our confidence is lower. Trend watch projects measured trends instead, such as how long a task AI can finish on its own.
- Method changes: a change of definitions re-derives our numbers and is labelled a method change, never news. Readings and forecasts made under earlier definitions keep their meaning and resolve under them. Every change is on Changes and corrections.
- What moved the needle: each day names the one development that changed an estimate most, or says plainly that it was a quiet day.
- Weekly (Fridays, 3pm Pacific): a wrap-up with the week's key numbers and graphs, plus the forecast scorecard, disclosure lag, AGI claims, calendar and steelman.
Honesty notes
- A custom AI agent does the work; method and sources are public.
- The probabilities are subjective.
- An absence of evidence is not proof of secrecy.
- Calibration isn't shown yet: see the scorecard.
- Readings freeze once emailed; fixes are dated corrections.
Changes and corrections
Method v2.0 (Oct 6) · alarm criteria v1.2 · 62 corrections, newest Nov 16 →