The short answer: from one number. Everything this service shows is derived from METR's autonomous task horizon — the length of work a model completes unaided. As of 2025 the anchor is 320 minutes at 50% success (Claude Opus 4.5).
What is measured and what is invented
The distinction matters, and it is worth holding in mind the whole time you are moving sliders.
- Measured: the historical METR horizon points and the doubling time. External data with a source.
- Invented: difficulty coefficients, deployment lags, rung weights, trigger multipliers, mitigation ceilings. These are the author's expert judgement. None of these numbers is measured, and none of them can be.
The second group carries the load. That is exactly why every one of its values is exposed as a slider rather than buried in the code: those are the things worth arguing about.
Extrapolation
D_eff = doubling time × friction × Π(trigger multipliers)
log₂H(t) = log₂H₀ + (t − t₀) / D_eff
The line is anchored at the latest point, not fitted to history. Two consequences look like bugs and are not: the line diverges from the early METR points, and raising friction pushes future dates to the right while pushing already-passed ones further into the past.
The computation runs on the logarithm of the horizon rather than the horizon itself. This is
not aesthetics: with the minimum doubling time and two accelerating triggers,
2^1590 by 2100 overflows a double, and the user would see NaN
instead of a date.
Dates for the breakdown rows
date = date(coefficient × threshold × reliability) + lag × group multiplier
The difficulty coefficient is how many times longer a chain of reasoning a domain demands relative to software engineering, which is what METR measures. The deployment lag is the years for hardware, capital, trust and regulators after the capability technically exists. Raising the bar from 50% to 80% success multiplies the required horizon by 5: per METR, the 80% horizon is roughly five times shorter.
The rationale for each coefficient sits inside its row in the Singularity section — expand the row.
The ladder of catastrophes
λᵢ(t) = (malice·wᵢ + control failure·uᵢ) · cᵢ(t) · d(t) · (1 − mitigation·eᵢ) · aᵢ(t)
Four multipliers, each of which you set by eye. A product of such multipliers is accurate to an order of magnitude at best — every decimal place in these dates is a polite lie told by the interface.
The rungs are nested
A rung's curve is the probability of an event at that level or worse. The bottom rung therefore always sits above the top one: a global catastrophe by definition also clears "≥1,000 dead". In the prototype the rungs were computed independently, and under some assumptions the model claimed extinction was likelier than a local incident. That has been fixed.
The window of vulnerability
The aᵢ(t) multiplier damps risk after capability plateaus: the world learns to
live with what has already happened. Without it every curve runs to 100% and the model stops
meaning anything. Setting the window slider to 100 years effectively switches the decay off.
The count starts today
Integration begins in the current year, not a fixed one. Otherwise, a few years from now the service would be accumulating risk over years already survived, and every curve would creep upward from the mere passage of time.
The Doomsday Clock
minutes to midnight = max(0.2; 15 × (1 − P(global) by 2100))
Fifteen minutes means global catastrophe is ruled out; midnight means it is certain. The scale is invented here and bears no relation to the real Doomsday Clock of the Bulletin of the Atomic Scientists beyond the visual nod.
Red lines
This service deals in casualty figures. That imposes obligations, and they are kept:
- no personal predictions and no geotargeted forecasts;
- no specific company, laboratory or country is named as responsible for future events;
- the scenarios on the catastrophe cards are described at the level of mechanism — "cascading failure" — with none of the operational detail anyone could read as instructions;
- triggers are phrased as observable criteria, not as calls to action or predictions.
And the main thing
This apparatus does not predict the future and cannot. An event that has never occurred has no training sample, no validation, and no way to tell a good model from a pretty one. All that happens here is that assumptions get laid out so you can see which one carries the whole load. If the control-failure slider moves the global date by forty years and the doubling-time slider moves it by five, then the argument is about the first one. That is the entire value; the dates are a by-product.
Openness
The code is MIT; the model constants and texts are CC BY 4.0. A model whose assumptions cannot be inspected has no claim to being taken seriously, so every coefficient lives in one versioned file and changes through a pull request with a stated rationale.