How over is it?

A mood ring wired to real instruments

Vibes are fineToo slow, franklyMild uneaseWarming upDoomscrollingBuildingExtremely doomedAcceleratingIt's overWe're so back0102030405060708090100Doom Indexpoints of dread, dailypoints of progress, dailyMod. CD-AI/26Calibrated daily 06:00 UTCzero adj.

Reading for 30 October 2026

63.1

Extremely doomedAccelerating

Since yesterday
−0.1
All-time high
65.9 on 15 Oct
Last 30 days. Dashed days are reconstructed.

Forecasters give AGI by 2030 a 52% chance. 2,225 AI papers landed last week.52% odds of AGI by 2030. Let's go. 2,225 AI papers shipped last week.

Time left, at current market pricesAGI arrives in, give or take

years days hours minutes seconds

The date where the Manifold markets on “AGI before year N” cross even odds: 2 May 2029. It moves when traders change their minds.

0%25%50%75%100%27282930313233343640
Chance of AGI before each year, Manifold. The year axis shows the last two digits.

Since you opened this pageShipped while you've been here

Counts are expected values from each source's latest weekly rate. The spikes are simulated arrivals at that rate, so they will not match the counts one for one.

The tests are running outWe need harder tests

Researchers keep writing exams that AI should not be able to pass. These are the best scores so far, from Epoch AI's benchmark tracking.Every exam built to stump AI gets beaten, usually within a year or two. Best scores so far, from Epoch AI.

General capability of the best model, Epoch Capabilities Index. GPT-6 Astra leads at 169.

  1. 95%

    ARC-AGI-2

    Visual puzzles built to be easy for people and hard for AI

    From 0.8% in Sept 2024. Best now: GPT-6 Astra.

  2. 100%

    FrontierMath Tier 4

    Research maths problems that take specialists days

    From 4.9% in Apr 2025. Best now: GPT-6.1 Sol.

  3. 93%

    Cybench

    Professional hacking challenges, solved without hints

    From 7.5% in Feb 2024. Best now: Claude Opus 4.6.

  4. 74%

    ExploitBench

    Working exploits for real software vulnerabilities

    From 14% in Oct 2025. Best now: Claude Mythos Preview Early.

  5. 96%

    GPQA Diamond

    PhD-level science questions you cannot simply search for

    From 36% in Mar 2023. Best now: Gemini 4 Argon.

  6. 21%

    Remote Labor Index

    Real paid freelance projects, done start to finish

    From 0.8% in Jun 2025. Best now: Claude Opus 5.5.

Scores from the Epoch AI Benchmarking Hub (CC BY 4.0). Human reference lines come from each benchmark's own study.

Further readings

17hours

The longest task the best AI agent finishes half the time, measured in hours a skilled human would need. It doubles roughly every 129 days.An AI agent can now handle a 17-hour task on its own about half the time. That figure doubles every 129 days or so.

METR, Claude Mythos Preview Early, 7 April 2026

9.4×1026FLOP

Compute in the biggest known training run, GPT-6 Astra. That is 45 times GPT-4.GPT-6 Astra trained on 45 times the compute of GPT-4. Scale keeps paying.

Epoch AI, 30 October 2026

275.8KGPUs

H100-equivalent chips in the largest known AI supercomputer, xAI Colossus Memphis Phase 3, drawing about 352 MW.xAI Colossus Memphis Phase 3: 352 megawatts of pure ambition.

Epoch AI, 30 October 2026

2,225

AI and machine learning papers posted to arXiv last week. Nobody has read all of them, including the authors.AI papers posted last week. The field has never moved faster.

arXiv, 25 October 2026

738

People who read Wikipedia's page on existential risk from AI yesterday.People who looked up AI risk yesterday and, presumably, felt fine.

Wikimedia, 29 October 2026

12stories

Hacker News stories about AGI, alignment or AI risk this week. The top one: “AI Safety Is Mostly a Sex Cult in Berkeley, California”.AGI threads on Hacker News this week. Top of the pile: “AI Safety Is Mostly a Sex Cult in Berkeley, California”.

Hacker News, 30 October 2026

3.20million

Public models on Hugging Face, with about 27,811 more every week.Open models anyone can download. About 27,811 more arrive each week.

Hugging Face, 30 October 2026

85seconds to midnight

The Doomsday Clock, set each January by the Bulletin of the Atomic Scientists. AI is one of the reasons it keeps moving.The Doomsday Clock. Not our department, but it is very close to midnight.

Bulletin of the Atomic Scientists, 27 January 2026

One week of AI, on a log scale

Each gridline is ten times the one before. For every federal document about AI, developers downloaded an AI toolkit 27.6M times.Each gridline is ten times the one before. Builders outpaced the paperwork by a factor of 27.6M.

Money and power

14.5GW

Drawn by 75 known frontier AI data centres, more electricity than the Netherlands uses on an average day. Building them cost about $551 billion.75 frontier AI data centres, already out-drawing the Netherlands. About $551 billion of concrete and chips, and rising.

Next up: Sweden, at 15.4 GW. Country figures are rounded average demand from Ember. Site data from Epoch AI; the largest is Anthropic-Amazon New Carlisle at 1,037 MW.

$2.24T

Combined top valuation of 7 AI labs, priced as if it all works.What the market thinks the labs are worth. Number go up.

  1. Anthropic $965B $65B a year in revenue
  2. OpenAI $852B $50B a year in revenue
  3. xAI $230B $500M a year in revenue
  4. DeepSeek $90B $1.0B a year in revenue
  5. Z.ai (Zhipu) $39B $1.6B a year in revenue
  6. MiniMax $37B $800M a year in revenue
  7. Mistral AI $24B $1.0B a year in revenue

Highest reported valuation and annualised revenue for each lab, from Epoch AI.

For every person reading about AI risk, 268 are reading about ChatGPTFor every worrier, 268 people are busy using the thing

Yesterday 197,747 people read Wikipedia's ChatGPT article and 738 read the one on existential risk from AI. Each dot is roughly 738 readers. The lit one is the risk page.

Wikimedia pageviews, 2026-10-29

The odds

Prices on real-money and play-money exchanges, read as probabilities. Markets on human extinction have an obvious payout problem, so read those as a mood rather than a forecast.

Will it arrive?Will it arrive?

Will it go badly?Will it go badly? (No.)

Where do you sit?

Prominent people have put a number on the chance that AI ends badly for humanity. Drag yours into the line-up.

Same as Elon Musk and Yoshua Bengio.

  1. Yann LeCun Chief AI scientist, Meta (at the time), 2023 under 0.01%
  2. Superforecasters Existential Risk Persuasion Tournament median, extinction by 2100, 2022 0.38%
  3. AI researchers AI Impacts survey of 2,778 authors, median, 2023 5%
  4. Vitalik Buterin Ethereum co-founder, 2023 about 10%
  5. Geoffrey Hinton Turing Award and Nobel laureate, 2024 10 to 20%
  6. Lina Khan Chair, US Federal Trade Commission (at the time), 2023 15%
  7. You 20%
  8. Elon Musk xAI, 2024 10 to 20%
  9. Yoshua Bengio Turing Award laureate, 2023 20%
  10. Dario Amodei CEO, Anthropic, 2023 10 to 25%
  11. Emmett Shear Interim CEO, OpenAI (briefly), 2023 5 to 50%
  12. Scott Alexander Writer, Astral Codex Ten, 2023 33%
  13. Paul Christiano Head of safety, US AI Safety Institute, 2023 46%
  14. Jan Leike Alignment lead, Anthropic, 2023 10 to 90%
  15. Daniel Kokotajlo AI Futures Project, formerly OpenAI, 2023 70%
  16. Dan Hendrycks Director, Center for AI Safety, 2023 over 80%
  17. Eliezer Yudkowsky Machine Intelligence Research Institute, 2023 over 95%
  18. Roman Yampolskiy AI safety researcher, University of Louisville, 2024 99.999999%

Public statements, placed at the middle of any stated range. Compiled from PauseAI's list and checked 2026-10-30. Ranges and dates matter; follow the link before quoting anyone.

Days since

  1. 1,430 ChatGPT launched
  2. 1,326 GPT-4 released
  3. 1,318 The six-month pause letter
  4. 1,249 The extinction statement
  5. 1,094 Bletchley Park summit
  6. 1,078 OpenAI's board fired Sam Altman
  7. 820 EU AI Act entered into force

What today's reading is made of

ComponentWeightScoreAddsFrom
Forecast proximity25%8922.2Manifold median AGI date 2029-05-02, 2.5 years away
Market odds20%316.17 markets across Manifold, Kalshi and Polymarket
Capability velocity20%5911.9METR 50% horizon 17.4 hours (Claude Mythos Preview Early)
Scale15%497.39.4e+26 FLOP training run, 275,796 H100e cluster
Attention10%585.8Wikipedia and Hacker News, z = 0.18 and 0.48
Flood10%949.4arXiv, z = 2.70

The published reading is a seven-day moving average of the raw score, so it drifts rather than jumps. Full method and formulas.