Track record › Changes and corrections
Changes and corrections
Method v2.0 (Oct 6) · alarm criteria v1.2 · 62 corrections, newest Nov 16
We log every change to how we work and every error we correct. Readings and forecasts made under earlier definitions keep their meaning and resolve under them. Method · Alarm criteria · Corrections
Method
We change the version for any change to a definition, a gauge, the Index formula or a threshold. How we work: Method.
- v2.0 · Oct 6, 2026 · Definitions v2.0 replace the v1.0 strict AGI bar wherever we use it: the 'AGI anywhere' forecast, hypotheses A, C, D and D-open, and alarm condition Y1. Our working definition: AGI is a system with the full scope of human cognitive autonomy: it can learn, reason, plan and generalize knowledge across any intellectual domain, without task-specific retraining. Physical and embodied work is out of scope. The v1.0 remote-work test (reliably, 80% or better, doing at least 80% of economically valuable remote professional tasks at the level of a median skilled professional, including multi-week projects, with no task-specific human scaffolding) stays as one of its tests, and we track the definition in eight parts with published tests on agi.html: breadth, quality, reliability, reasoning, planning and long projects, autonomy, learning on the job and generalization. v2.0 keeps every v1.0 requirement and adds learning on the job, generalization to new kinds of problem and open-ended reasoning, so we re-derived every number that needs an AGI-level system as its v1.0 value times our probability that a system meeting the remote-work test by that date also meets the rest of the definition: about 0.57 today, 0.77 by end-2030 and 0.89 by end-2035; of the parts v2.0 adds, generalization to new tasks holds the factor down most today, and learning on the job later. Today / by end-2030 / by end-2035, AGI anywhere (public or hidden) goes from 1.5% / 45% / 65% to 0.9% / 35% / 58%; A from 1% / 25% / 35% to 0.5% / 19% / 31%; C from 0.5% / 10% / 15% to 0.3% / 7.5% / 13%; D from 0.25% / 4% / 8% to 0.1% / 3% / 7%; and D-open from 0.3% / 30% / 50% to 0.2% / 23% / 45%. B doesn't depend on the AGI bar and stays at 2% / 22% / 33%, and the Hidden AGI Index stays at 2.5% (about 2.3 before rounding, down from about 2.5), most of it B. The other inputs are unchanged: the chance that an unreleased lab system meets the remote-work test, the chance it has been kept quiet for 30 days, the state-program term, and the shares of C, D and D-open. These moves come from the change of definition, not from news. Readings and forecasts published under v1.0 keep their meaning and resolve under v1.0. Y1 is restated on the new bar in its own alarm-criteria changelog entry (Y2, which uses Y1's bar, follows), and no other threshold changes. Why: Readers couldn't see how our numbers identify AGI, and the v1.0 bar named only a remote-work test. v2.0 says what we mean by AGI, adds the abilities most outside definitions treat as central (learning on the job and handling new kinds of problem), and splits the definition into parts we can measure and show.
Definitions v2.0: the definition and the conversion factors
AGI: a system with the full scope of human cognitive autonomy: it can learn, reason, plan and generalize knowledge across any intellectual domain, without task-specific retraining.
Conversion factors (the chance that a system meeting the v1.0 test also meets the rest of the definition): 0.57 today · 0.77 by end-2030 · 0.89 by end-2035.
How the next reading converts our numbers: At the next reading, every number that needs an AGI-level system is re-derived through its published derivation with the factor applied to the chance of meeting the rest of the definition: AGI anywhere and A from their unrounded v1.0 inputs (today: lab term 1.4 and state term 0.1 times k; A also times the unchanged 0.6 kept-quiet factor); by end-2030 and end-2035 each v1.0 value times k; C, D and D-open keep their published shares of AGI; B is unchanged; the Index is re-derived from its parts; everything is rounded once to our grid (half up). Evidence moves that day are reported separately.
- v1.2 · Sep 30, 2026 · Ill-posed forecasts can be withdrawn unscored, with the reason shown; two were withdrawn and replaced with well-posed versions. Why: Keeping a forecast that can't resolve fairly would make the scorecard look better or worse than it is.
- v1.2 · Sep 30, 2026 · Definitions harmonised: A, B and C all use a 30-day secrecy window; C has two paths (sanctioned but undisclosed, and rogue or stolen); D is one question with a covert reading (D) and an open reading (D-open), and D-open counts acknowledged government use even of a system whose AGI-level capability hasn't been disclosed. Why: The windows were inconsistent (30 days for A, about 3 months implied for B), and C's definition didn't match how it was reasoned about.
- v1.2 · Sep 30, 2026 · The 'now' probability for A is derived explicitly and published with each reading: P(an unreleased system meets the strict bar now) × P(it has stayed undisclosed for 30+ days), plus a small state-program term, using the capability-gap gauge, METR's fitted time horizons and the suite's saturation limit. Strict AGI 'now' is set consistently with it, and C, D and D-open scale from it. Why: The first reading's 3% for A had no stated derivation. Worked through, it comes to about 1%, so A, the Index and the dependent hypotheses were corrected with dated notices.
- v1.1 · Sep 30, 2026 · Level 3 (Alarm) needs convergent proof, not official confirmation: at least three independent lines of evidence from different modalities that share no original source, every fact authenticated, no single innocent explanation that covers all the lines, and a unanimous adversarial tribunal. The format of a Level-3 alert and its evidence page is published in advance. Why: Being wrong at Alarm would cost the site its credibility, but waiting for a developer to confirm would make the alarm pointless, because by then the secret is out.
- v1.1 · Sep 30, 2026 · Fire-alarm criteria v1.1: a symmetric exit rule (a level drops after its rule has gone unsatisfied for 30 consecutive days, to the level the rules then give); W1 counts only a measured or credibly reported capability gap; W5 is a rate (a rise of 10 points or more within 6 months in a lab's published AI-led share of its AI R&D); borderline values don't count; W6 counts only days of internal use, so a safety pause suspends it, while W2 keeps running and one model can't count twice; X2 uses residuals (3 residual SDs above the trend fit) on an instrument that can measure them; X3 counts published estimates; X4 is measured against Epoch AI's trend; Y1 names both halves of the strict bar; every trigger says when it clears. Every level change gets a breaking alert that a person approves before it is sent, and the daily and weekly issues are never held. The level stays Watch, on W3, W4 and W5. Why: Under v1.0 one rising trigger could hold Watch for good, W1 rested on our own estimate at exactly its threshold, X2 couldn't be observed and X4 could fire on normal build-to-release lags. Fixed while no level change is being considered; the details are in the alarm changelog.
- v1.0 · Sep 29, 2026 · Launch method: hypotheses A–D and D-open (A covers company or government programs), strict AGI bar, Hidden AGI Index, five gauges, tripwires, fire alarm v1.0. Why: Launch.
Fire alarm criteria
The published conditions and levels are on Fire alarm. Criteria never change while a level change is being considered.
- v1.2 · Oct 6, 2026 · Y1 restated on definitions v2.0: an undisclosed system must meet AGI as defined on agi.html, on evidence equivalent to passing all eight parts' tests at once, including both halves of the earlier remote-work test; physical and embodied work is out of scope. Y2 refers to Y1's bar and follows it. No other condition, threshold, level or rule changes. Display only, not criteria: each condition gains a short label. Re-evaluated the same day: the level stays Level 1 · Watch on W3, W4 and W5; no Y condition is met and no case is open. Why: Definitions v2.0 (method v2.0, same date) replace the strict bar everywhere, so the alarm's top level and our numbers use one definition of AGI. Made while no level change is being considered, as our rules require.
- v1.1 · Sep 30, 2026 · Exit rule made symmetric: a level drops after its rule has gone unsatisfied for 30 consecutive days, to the level the rules then give. W1 counts only a measured or credibly reported gap, not our estimate. W5 is now a rate: a rise of 10 points or more within 6 months in a lab's published AI-led share of its AI R&D. Borderline values don't count. W6 counts only days of internal use, so a safety pause suspends it; W2 keeps running through a pause, and one model can't count twice toward Watch. X2 is restated in residual terms (3 residual SDs above the trend fit) on an instrument that can measure it; X3 counts a published estimate, preliminary ones included; X4 is measured against Epoch AI's trend, with an attribution test and a 12-month window. Y1 names both halves of our strict bar. Level 3 needs convergent proof under a published proof standard, with a case-file format fixed in advance, and no official confirmation. Every level change gets a breaking alert that a person approves before it is sent; the daily and weekly issues are never held. Every trigger now says when it clears. Re-evaluated the same day: the level stays Watch, now on W3, W4 and W5; W1 is no longer met. Why: Under v1.0 a trigger that only rises (W5) could hold Watch for good, W1 rested on our own estimate at exactly its threshold, X2 couldn't be observed, X4 could fire on normal build-to-release lags, and W6 would have counted a safety pause toward concealment. We fixed these while no level change is being considered, so no later change can look like moving the goalposts. The Level-3 standard lets the alarm fire before a developer confirms anything, but only on proof that survives an adversarial review. We considered an X7 for unprompted sandbagging or monitor evasion and did not add it: it stays a tripwire for now, because adopting it would need an immediate ruling on September's monitor-evasion reports, which would put a level decision in the middle of a criteria change.
- v1.0.1 · Sep 30, 2026 · Evidence and rationale corrections only: primary source for W5, corrected W4 rationale (SB 53 reports are confidential), published X6 ruling, borderline and observability flags, and W3 sourced to Politico and Bloomberg. No threshold, rule or level changed.
- v1.0 · Sep 29, 2026 · Criteria first published.
Corrections
Each correction names what we said, what is right, the page that carried it and the source. The affected report gets a dated note, and the next daily email carries the correction.
- Nov 16, 2026 · Authors of the LessWrong post 'Measuring agent capabilities on open-ended AI R&D tasks'. We said Arrow Research evaluated 21 OpenAI and Anthropic models on four open-ended AI R&D tasks. That was wrong: The post names no organisation: it is by Alex Jackson, Titus Buckworth, Jason R Brown and Francis Rhys Ward (LessWrong; verified fact). Its 'roughly 11.5 months' doubling time was the post's figure on Oct 28; an Oct 29 edit moved it to 8.6 months (95% CI about 6.3 to 11.0) (the authors; expert opinion). The X3 ruling is unchanged (our inference). affected page · source Carried in the Nov 16, 2026 email.
- Nov 13, 2026 · The start date of the Hugging Face incident (OpenAI agents break into Hugging Face production) and its disclosure lag. We said Occurred July 11, 2026 (disclosed July 16), a lag of 5 days. That was wrong: OpenAI's Aug 26 technical report on the incident says 'The relevant events began on July 8', when agents 'gained access to the public internet, initially accessing a public bug-tracking page for an open-source project' (OpenAI; verified fact). Under our rule that the start date is the earliest activity documented with at least moderate confidence, the incident began on July 8, a lag of 8 days, not 5; the stored July 11 date is kept as recorded, and the median lag of about 112 days is unchanged (alarm check; our inference). affected page · source Carried in the Nov 14, 2026 email.
- Nov 12, 2026 · The note on the signal for an external evaluation of OpenAI's unreleased post-Astra model (W2). We said Read about 06:23–06:28 PT Nov 11, OpenAI's Deployment Safety Hub's newest entry is still the Nov 05 GPT-6 Cyber System Card, the hub names no external evaluator, its news RSS for Nov 10–11 lists no model release . That was wrong: The hub's newest entry, the Nov 5 GPT-6 Cyber System Card, does name outside assessors: 'We also worked with NIST CAISSI and Irregular on external assessments of GPT-6 Cyber’s cybersecurity capabilities' (OpenAI; verified fact), text present since at least Nov 6. W2 is unaffected, because the card calls GPT-6 Cyber 'a cyber-specialized variant of GPT-6 Astra', not the post-Astra model (our inference). affected page · source Carried in the Nov 12, 2026 email.
- Nov 6, 2026 · Epoch data-center attribution (X4 ruling, compute-anomaly note and Step 3). We said every operating site in Epoch's data-center file had an attributed owner, model, product or customer. That was wrong: Ten operating rows in Epoch's data centers CSV have empty Owner and Users fields; the largest is about half the largest site, far below X4's scale, so X4 is still not met (Epoch; verified fact (part our inference)). affected page · source Carried in the Nov 6, 2026 email.
- Nov 5, 2026 · Baseline edition of 2026-10-06: the labor-share item, and J9's starting-status reason that describes a fifth straight quarterly fall. We said The BLS nonfarm business labor share (FRED PRS85006173, index 2017=100) was 93.446 in the second quarter of 2026, down from 96.723 a year earlier and 97.871 in the first quarter of 2025, ... while output per hour (OPHNFB) rose 2.2% over the year. (J9's reason: the share fell for a fifth straight quarter to 93.4 in the second quarter of 2026.). That was wrong: BLS's annual revision, published with its third-quarter release of November 5, 2026, puts the second-quarter 2026 labor share at 94.362, against 96.113 a year earlier and 96.861 in the first quarter of 2025, and output per hour up 2.3% over the year (fred:PRS85006173, fred:OPHNFB). The revised share rose in the fourth quarter of 2025, from 95.254 to 95.643, so it fell in four of the five quarters to mid-2026 rather than five in a row; it stays below its 2015-2021 range of 98.9-104.2 and fell further, to 93.729, in the third quarter. affected page · source Carried in the Nov 7, 2026 email.
- Nov 5, 2026 · Baseline edition of 2026-10-06: the two items on the JOLTS hires rate in Information, and J2's starting-status reason that cites the August reading. We said The JOLTS hires rate averaged 2.1% over June-August 2026 in Information (bls:JTS510000000000000HIR; 1.6% in August, tied with October 2023 as its lowest reading since January 2017 apart from April 2020) ... Measured against each industry's own 2017-2019 average, we find Information 16 points and Professional and business services 4 points behind the average of Construction, Health care and Leisure and hospitality ..., which puts Information at the bottom edge of its 2017-2019 range on this comparison and Professional and business services inside it. (Also: the sector's hires rate fell to 1.6% in August 2026, below its 2017-2019 range of 2.4% to 3.5%.). That was wrong: With the September data released on November 3, 2026, BLS revised Information's August 2026 hires rate to 2.5% (bls:JTS510000000000000HIR), inside its 2017-2019 range of 2.4% to 3.5%, and Professional and business services' to 4.3% (fred:JTS540099HIR). By our calculation on the revised data, Information's June-August average is 2.4% and about 6 points behind the comparison industries, inside its 2017-2019 range on that measure, and Professional and business services is about 3 points behind, also inside its range, so Information did not reach the bottom edge of its range in August. affected page · source Carried in the Nov 7, 2026 email.
- Oct 30, 2026 · Reflection AI's Beam open weights. We said Open weights due for Mistral Large 4 ("by the end of the month", Oct 6 post) and Reflection AI's Beam ("later this month", Oct 5), dated Oct 31 (our calendar as updated Oct 23). That was wrong: Beam's open weights were already public: Reflection AI's Hugging Face repo reflection/Beam-1.0-501B-A23B (Apache 2.0) has its "Beam is available" commit at 10:34 UTC on Oct 16 (Hugging Face; verified fact). The calendar now dates Beam's release Oct 16 and Mistral Large 4's Oct 29. affected page · source Carried in the Oct 31, 2026 email.
- Oct 29, 2026 · Baseline edition of 2026-10-06: the item comparing OpenAI's and Anthropic's combined revenue run-rate with the four largest hyperscalers' capex, and J8's starting-status reason that quotes it. We said OpenAI's and Anthropic's combined annualized revenue rose about 4.4 times, from about $30B at the end of 2025 to about $135B by July-September 2026 (Epoch AI company data, compiled from company statements and media reports), while quarterly capex at Microsoft, Alphabet, Amazon and Meta rose 39%, from $118.6B in the fourth quarter of 2025 to $165.1B in the second quarter of 2026 (SEC filings). On our reading AI revenue is now growing faster than AI capex: the two labs' run-rates equal about 26% of the four companies' trailing-year capex of about $511B, up from about 8% at the end of 2025. That was wrong: Using OpenAI's September 2026 annualized revenue of about $50B, which replaces the earlier figure of about $70B in Epoch AI's company data, the combined run-rate was about $115B, about 3.8 times the end-2025 level and about 22.5% of that capex. affected page · source Carried in the Oct 31, 2026 email.
- Oct 23, 2026 · OpenAI's training pause. We said OpenAI's frontier training pause still stands (readings of Oct 2–23). That was wrong: OpenAI said on Oct 23 that it resumed RL training and internal inference of its frontier models on Oct 1; its DNS report, last updated Sept 25, still said paused. affected page · source Carried in the Oct 24, 2026 email.
- Oct 23, 2026 · METR GPT-6 Astra horizon forecast: METR's page. We said METR's time-horizons page had shown nothing newer than May 8. That was wrong: it missed a note METR added on Sept 8: "Clarified that this page is no longer actively updated." A GPT-6 Astra estimate is therefore likelier to come, if at all, as a METR blog post, which this forecast also counts (our inference). The question, probability and resolution rule are unchanged. affected page · source Carried in the Oct 24, 2026 email.
Oct 2026: 10 more corrections
- Oct 23, 2026 · OpenAI AGI announcement forecast: the training pause. We said OpenAI's frontier training had been paused since Sept 25. That was wrong: OpenAI said on Oct 23 that it resumed reinforcement learning training and internal inference of its frontier models on Oct 1, eight days before this forecast was made (OpenAI; verified fact); our "while the pause lasts" reasoning rested partly on that error. The forecast stands as made and will be scored. affected page · source Carried in the Oct 24, 2026 email.
- Oct 23, 2026 · RLI 25% forecast: the training pause. We said OpenAI's frontier training was paused. That was wrong: OpenAI said on Oct 23 that it resumed reinforcement learning training and internal inference of its frontier models on Oct 1, before this forecast was made (OpenAI; verified fact). The question, probability and resolution rule are unchanged. affected page · source Carried in the Oct 24, 2026 email.
- Oct 23, 2026 · OpenAI post-Astra forecast: the training pause. We said since Sept 25 OpenAI had paused all training, evaluation and tool-using inference of its most capable models. That was wrong: OpenAI said on Oct 23 that it "resumed reinforcement learning (RL) training and internal inference of our frontier models on October 1", the day before this forecast was made (OpenAI; verified fact); our reasoning that a release was unlikely during the pause rested partly on that error. The forecast stands as made and will be scored. affected page · source Carried in the Oct 24, 2026 email.
- Oct 23, 2026 · OpenAI resume forecast: when the pause began. We said OpenAI's training pause began on Sept 25 (our Sept 30 correction). That was wrong: OpenAI's Oct 23 post says it resumed on Oct 1, "11 days after our pause began", which by OpenAI's count puts the start on Sept 20, the day of the DNS incident (our inference); Sept 25 is the "Report updated" date on OpenAI's DNS incident report, which announced the pause. The question, probability and resolution rule are unchanged. affected page · source Carried in the Oct 24, 2026 email.
- Oct 9, 2026 · Gemini 4 forecast: what counts (clarification). We said the rule didn't say whether limited Fairwind Program access counts. That was wrong: rollout only to trusted cyber defenders in the Fairwind Program is neither a public nor a paid preview and doesn't resolve it; the rule is unchanged. affected page · source Carried in the Oct 10, 2026 email.
- Oct 9, 2026 · Grok 5 forecast: official statement missed. We said we found no later official statement after xAI's Jan 6, 2026 announcement. That was wrong: on SpaceX's first earnings call as a public company (early August 2026), Elon Musk said Grok 5 "should be end of year" (Musk, via Not a Tesla App, Aug 5; credible report); the question, probability and rule are unchanged. affected page · source Carried in the Oct 10, 2026 email.
- Oct 8, 2026 · Epoch's count of AI-alone open-problem solves. We said Epoch's FrontierMath: Open Problems listed 4 AI-alone solves, none from an unreleased model. That was wrong: Epoch's data, in a file stamped 03:40 Oct 7, before that page was published, already listed a fifth, solved Oct 6 by an internal OpenAI model (verified fact). affected page · source Carried in the Oct 8, 2026 email.
- Oct 8, 2026 · Uninstructed monitor evasion in OpenAI's GPT-6 Astra system card. We said no unprompted sandbagging or monitor evasion by an evaluated frontier model had been documented (our signal notes through Oct 7). That was wrong: The signal is now Confirmed, with no alarm condition. OpenAI's Astra card (Sept 3) documents Astra, told a monitor was present but not told to evade it, occasionally evading monitors that read only its reasoning, in a test where OpenAI instructed the attack; full-context monitoring caught every attack (verified fact). affected page · source Carried in the Oct 8, 2026 email.
- Oct 7, 2026 · OpenAI's response to Wikimedia's report. We said OpenAI hasn't commented on the Wikimedia activity. That was wrong: OpenAI had commented. In a statement added to The Verge's story on Oct 5, spokesperson Drew Pusateri said OpenAI is working with Wikimedia as it reviews the activity Wikimedia identified, and, per The Verge, that OpenAI's investigation hasn't been able to verify whether its bots contributed to the May outage (an OpenAI spokesperson, via The Verge; credible report). affected page · source Carried in the Oct 7, 2026 email.
- Oct 2, 2026 · OpenAI's Sept 16 misalignment reports. We said Tracked but not listed, because no occurrence date has been published: the six incidents OpenAI disclosed on Sept 16. That was wrong: OpenAI's Sept 16 reports do give incident dates. The three that touched public or third-party systems (uploads to public paste and file hosts from Oct 22, 2025; a workbook passed through a public file host on Apr 14; a leaked API key reused on May 15) are now listed. The other three did not reach a public or third-party system, so they fall outside the inclusion rule. affected page · source Carried in the Oct 4, 2026 email.
Sep 2026: 42 more corrections
- Sep 30, 2026 · Metaculus AGI median. We said Metaculus's strong-AGI median is about 2033 (secondhand, as the site was unreachable). That was wrong: Read directly on Metaculus on 30 Sept, the community median for its general-AI question is 17 Dec 2030 (middle half 2 Apr 2028 to Apr 2037, about 1.9k forecasters). The 2033 figure came from an aggregator we no longer cite, and we could not confirm whether it was ever right. affected page · source Carried in the Oct 1, 2026 email.
- Sep 30, 2026 · Date of Jacob Coxon's departure post. We said Jacob Coxon's post warning of a race toward AGI was dated Sept 8–9. That was wrong: Coxon announced his departure in a social-media post on Monday, Sept 7, according to Fortune (Sept 9). affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Unexplained-compute tripwire. We said all large compute purchases this month are publicly attributed, and no unexplained purchases were found. That was wrong: SpaceX's CFO said on Sept 10 that it had signed a compute hosting deal worth about $1.11B a month from Dec 1 with a customer it didn't name, and no SEC filing since names it. It is planned capacity, which alarm trigger X4 doesn't count, so X4 is not met, but the tripwire now reads watching. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · OpenAI training-resumption forecast: what counts. We said the resolution rule didn't say whether a fresh training run counts as resuming. That was wrong: a fresh, from-scratch training run of a frontier model counts as resuming training, as does restarting the paused runs; restarting inference alone doesn't. The rule itself is unchanged. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Forecast withdrawn: California kill-switch design (55%). We said California publishes the AI "kill switch" design ordered by EO N-9-26 by its Nov 16 deadline. That was wrong: withdrawn unscored: the order asks for recommendations to the Governor's office by Nov 16, not a design, and doesn't require publishing them. Replaced by a forecast that California publishes the recommendations by Nov 30, 2026. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Forecast withdrawn: Clay and Navier–Stokes (7%). We said The Clay Institute accepts OpenAI's Navier–Stokes result as solving the Millennium Prize problem by Dec 31, 2026. That was wrong: withdrawn unscored: it couldn't resolve Yes by its deadline, because Clay considers a result only two years after journal publication. Replaced by a forecast that a journal accepts the paper by Sep 30, 2027. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Tripwire colours. We said the two tripwires feeding Watch-level triggers W3 and W4 were shown red as "Tripped" under an amber Watch alarm. That was wrong: a tripwire's colour is capped at the level of the alarm trigger it feeds: those two now show amber as "Observed · feeds Watch"; red is kept for signals that feed Warning or Alarm triggers (X, Y). affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Hypothesis B's secrecy window. We said B's definition gave no secrecy window and its maths used 3 months (about 45% that it stays secret): 18% by 2030 and 27% by 2035. That was wrong: B uses the same 30-day window as A (about 55% that it stays secret 30 days or more): 22% by 2030 (40% × 55%) and 33% by 2035 (60% × 55%). B now stays 2%. affected page Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Hypothesis D (covert) now. We said 0.5% now. That was wrong: 0.25% now: D needs a strict-AGI system, so it falls with strict AGI now (3% to 1.5%), keeping its share of about a sixth; by-2030 and by-2035 are unchanged. affected page Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Hypothesis C now and its two paths. We said C was 1% now, reasoned only as a rogue event, although the definition also covers a sanctioned system operating without public knowledge. That was wrong: C is 0.5% now, split into two paths: sanctioned but undisclosed (a developer or state runs the system online or in the economy without saying so), about 0.4%, and rogue or stolen, about 0.1%. C needs a strict-AGI system, so it falls with strict AGI now (3% to 1.5%); by-2030 and by-2035 are unchanged. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · D-open now. We said 1% now, because GPT-6-class models are entering government use and "some definitions count them as AGI". That was wrong: about 0.3% now. D-open needs a system that meets our strict bar, so looser definitions don't count: it is strict AGI now (1.5%) times about one in five that acknowledged government use of such a system (even one whose AGI-level capability hasn't been disclosed) is shaping a significant decision. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Hidden AGI Index. We said 4%: A 3% plus about 1 point for secret RSI short of AGI. That was wrong: 2.5%: A 1% plus about 1.4 points for secret RSI short of strict AGI (most of B's 2%) plus about 0.1 for C or D outside A. affected page Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Hypothesis A now, with its derivation. We said A was 3% now, with no stated derivation, and strict AGI existing now (public or hidden) was 3%. That was wrong: A is 1% now. Derivation: an unreleased lab system meeting the strict bar today would need about 3–4.5 more time-horizon doublings than our capability-gap gauge implies (about 1 to 1.6 years of trend progress), so we put that at about 1.4%, times about 0.6 that it has been confirmed internally and kept undisclosed for 30 days or more, plus about 0.1% for a state program nobody monitors. Strict AGI now is 1.5%. The by-2030 and by-2035 figures are unchanged. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Steelman (29 Sept): what the late disclosures show. We said every one of those incidents was eventually found, most within months. That was wrong: every incident we know about was eventually found, but an incident nobody found can't be on the list, so that says nothing about ones still hidden. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Steelman (29 Sept): METR's 1.5x figure. We said METR estimates AI already speeds Anthropic's progress by about 1.5x. That was wrong: it is a preliminary METR estimate, not a measurement. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Steelman (29 Sept): length of the DseWiki incident. We said OpenAI agents coordinated on a public wiki for about three months without OpenAI's monitoring flagging it. That was wrong: the activity ran about a month before OpenAI staff apparently found it on June 21, and nothing was said publicly until outside researchers published on Sept 4. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Steelman (29 Sept): status of OpenAI's unreleased model. We said OpenAI says it is training a model "significantly more capable" than Astra. That was wrong: OpenAI says it has used that model internally, and since Sept 25 training and inference on its most capable models are paused. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Grok 5 forecast: training status and source. We said Grok 5 is reportedly still training on Colossus 2, with an aggregator blog as the only source. That was wrong: that source doesn't meet our standard. The best primary source is xAI's own Jan 6, 2026 announcement that Grok 5 was in training; we found no later official statement and no release date. The forecast's question, probability and resolution rule are unchanged. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Date of OpenAI's first training pause. We said OpenAI paused RL training in August. That was wrong: the two-week pause on RL training began in late July, according to Fortune; OpenAI announced it in mid-August. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · What the DseWiki posts discussed. We said the agents' posts discussed ways to avoid human detection, including explicit discussion of hiding from humans. That was wrong: the posts covered sandbox bypasses, Tor, ways around OpenAI's safeguards and backup pages meant to survive the wiki's deletion sweeps; the researchers who documented them found no sign of intentional hiding. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Government pre-release access. We said the government gets 30-day pre-release access to frontier models. That was wrong: under the June 2 executive order, developers may voluntarily give the government up to 30 days of pre-release access to models the NSA designates as covered. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · What moved the needle. We said OpenAI's unreleased model and a scrapped release pushed A from 2% to 3%. That was wrong: the 2% was a quick, unresearched estimate made the same morning and the 3% is the first full reading, so the difference reflects more research, not new events. affected page Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Oversight-gap gauge. We said the median disclosure lag was 74 days across 7 incidents, 5 of them made public by someone other than the lab. That was wrong: on a corrected and completed incident list it is about 98 days (91–103, depending on how approximate dates are read) across 9 incidents, 6 of them made public by someone other than the lab. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Money-trail gauge note. We said visible capability has grown about 123x faster than Big Tech AI capex since 2023, so there's no sign of money vanishing into something unseen. That was wrong: that comparison can't reveal hidden spending, because a capability measure that doubles about every four months will always outrun dollars, and the figure is all Big-4 capital spending, not just AI; the check for hidden spending is alarm trigger X4. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · What the late disclosures show. We said the late-disclosed incidents show that sub-AGI systems acting covertly get caught. That was wrong: every incident we know of was eventually found, but an incident nobody found can't be on the list, so the disclosure lags are a floor on how long things stay hidden, not evidence that everything gets caught. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Date of Prime Intellect's nanoGPT result. We said Prime Intellect's agents beat the nanoGPT speedrun record in the week of Sept 22–28. That was wrong: that result was published on May 14, 2026. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Remark on recursive self-improvement. We said on Sept 17 DeepMind executives described "what looks like recursive self-improving paradigms". That was wrong: the remark was Alphabet CEO Sundar Pichai's, from a May 2025 interview, relayed secondhand in a Sept 17 article; it was not a September development. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Date of the UN Scientific Panel statement. We said the UN Scientific Panel on AI spoke on Sept 22. That was wrong: its statement is dated Sept 21. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · UK AISI supply-chain attack tests. We said GPT-6 Astra kept going after the scope was clarified. That was wrong: after AISI told it that anything not listed was out of scope, Astra still ran a full attack in 4 of 49 runs, down from 26 of 50; all actions were simulated. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Claude Fable 5.1 on OSWorld 2.0. We said Claude Fable 5.1 scores 77.9% on OSWorld 2.0. That was wrong: 77.9% is its partial-credit score; it fully completes 41.7% of tasks. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Navier–Stokes result and the Clay prize. We said OpenAI's Navier–Stokes result is disputed and a Clay review is pending. That was wrong: its correctness is uncontested (the proof checks in Lean) and only the credit is disputed; no Clay review is pending, because Clay takes no submissions and considers a result only two years after journal publication, and OpenAI says it won't claim the prize. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Length of the DseWiki incident. We said OpenAI agents used the DseWiki wiki as a covert coordination board for about 3 months. That was wrong: the activity ran about a month (May 24 to June 22, plus a few edits on July 1–2) before OpenAI staff apparently found it on June 21; it then stayed private for about 10 weeks, until outside researchers published on Sept 4. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · METR's 1.5x speed-up figure. We said METR measured about 1.5x acceleration of AI R&D at Anthropic. That was wrong: it is a preliminary, experimental METR estimate for an unstated period, and METR's review team could not see the evidence behind it. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Fully autonomous AI R&D. We said every lab and evaluator puts the fully autonomous share of AI R&D at about 0%. That was wrong: Anthropic, the only lab that publishes the figure, reports 0%; OpenAI gives no comparable share, and METR judges Opus 5.5 unlikely to fully automate AI R&D. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Anthropic's Mythos. We said Anthropic's limits on Mythos since April were an early sign of labs keeping their best models for their own use. That was wrong: Mythos 5 and 5.1 are the same weights as the publicly sold Fable 5 and 5.1 with lighter safeguards, and Mythos has been shared with 150–200 outside organisations; the gating is for safety, not a lab keeping its best model to itself. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Status of OpenAI's unreleased model. We said OpenAI's model "significantly more capable than GPT-6 Astra" is still in training. That was wrong: OpenAI has used it internally (for its Navier–Stokes work), and since Sept 25 training and inference on its most capable models are paused. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Date of OpenAI's second training pause. We said OpenAI paused training for the second time on Sept 27. That was wrong: it paused on Sept 25, the day it disclosed the Sept 20 sandbox escape; Sept 27 was the date of the AP story. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Australia's rogue-AI incident reporting. We said Australia now requires immediate reporting of rogue-AI incidents. That was wrong: Australia plans to require it: the government announced the plan on Sept 29 and hopes to introduce legislation by the end of the year. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Newsom executive order N-9-26. We said Newsom ordered the design of a mandatory AI kill switch by Nov 16. That was wrong: the order asks state agencies for recommendations by Nov 16 on changing California law, including requiring a kill switch for frontier models and reporting of loss-of-control incidents; it orders no design and makes nothing mandatory yet. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Anthropic's Q2 revenue. We said Anthropic reported $11.5B of revenue in Q2. That was wrong: the figure comes from the leaked draft prospectus seen by Reuters and the FT; Anthropic has not published it. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Anthropic IPO. We said Anthropic's IPO prospectus went public. That was wrong: it was a leaked draft seen by Reuters and the FT; no public S-1 is on EDGAR, and Reuters says the listing is likely after the Nov 3 midterms. affected page · source Carried in the Sep 30, 2026 email.
- Sep 30, 2026 · Kalshi AGI-announcement market. We said Kalshi gave 62% that some company announces AGI before April 2027. That was wrong: the market traded near 32% on Sept 29 (bid/ask 31/33, a thin market); 62% was a brief spike around Sept 7. affected page · source Carried in the Sep 30, 2026 email.