How the Doom Index works

One number from 0 to 100, published once a day at 06:00 UTC. It blends six measurable signals about how fast AI is arriving and how worried people are about it. The jokes are on the surface; the arithmetic underneath is plain and reproducible.

The six inputs

Each input is turned into a score from 0 to 100, then combined with fixed weights. If an input has no fresh data for more than 30 days it drops out and the remaining weights are scaled up to fill the gap.

Forecast proximity

Years until the forecast arrival of AGI, y: the date where Manifold's ladder of “AGI before year N” markets crosses 50%.

s = 100 / (1 + (y / 10)^1.5)

Ten years out scores 50, five years about 74, two years about 92.

Market odds

A weighted mean of selected binary markets on Manifold, Kalshi and Polymarket, read as probabilities. Catastrophe markets count double so that small moves in small probabilities register.

s = 100 × Σ wₘ pₘ / Σ wₘ

Capability velocity

METR's 50% task horizon h: the length of task, in human-expert time, that the best measured AI agent completes half the time. Scored on a log scale from one minute to 2,000 hours, roughly a working year.

s = 100 × clamp(ln(h / 1 min) / ln(2000 h / 1 min), 0, 1)

Scale

The mean of two log-linear maps from Epoch AI: the largest known training run from 10²⁵ to 10²⁹ FLOP, and the largest known AI supercomputer from 10⁴ to 10⁷ H100-equivalents.

Attention and flood

Attention combines daily Wikipedia readers across five AI-risk articles and the weekly count of Hacker News stories about AGI, alignment and AI risk. Flood uses weekly AI and machine learning papers on arXiv and, once enough history builds up, new models on Hugging Face. Each input becomes a z-score of its recent mean against the trailing year, then a logistic score.

s = 100 / (1 + e^(−z))

A z of 0 scores 50; a z of 2 scores about 88.

Combining and smoothing

The raw index is the weighted mean of the available scores. The published index is a seven-day exponential moving average, so it moves every day without lurching.

Dₜ = 0.25 Rₜ + 0.75 Dₜ₋₁

Today the raw score is 62.71 and the published reading is 63.09.

Bands

• 0–19 Vibes are fine
• 20–39 Mild unease
• 40–59 Doomscrolling
• 60–79 Extremely doomed
• 80–100 It's over

Reconstructed days

History before the first published day was rebuilt from each source's own archive: METR and Epoch records by date, Wikipedia pageviews, and weekly Hacker News and arXiv counts. Market and forecast inputs have no stored history, so they are held at their first published values. Those days are drawn dashed and flagged in the data.

What it is not

It is not a probability of anything. It is a mood ring wired to real instruments. The weights are a judgement call and are public so you can disagree with them precisely. Every input links to its source on the status page, and the full daily history is in the repository as JSON.