A half-joking interface for a serious uncertainty · v1.2.0

Singularity

The definition here is operational, not mystical: singularity has arrived when AI completes, unaided and with no human in the loop, a task of a given length across a given share of activities. No "awakening of consciousness" — only the point past which forecasting the economy in terms of human labour stops making sense.

Preset

The measured anchor MEASURED

Everything here is extrapolated from one measurement: how long a task a frontier model finishes on its own, half the time. Every other number on this site is a judgement, so this one gets a source, a date and a choice.

METR data cutoff
May 8, 2026
Anchor
Claude Opus 4.6 · 12 h
95% CI
5.3 h – 60.6 h
Source
metr.org
Anchor point

measured on the model released February 5, 2026

Why this anchor and not the highest one?

The default is the most recent measurement METR still stands behind. METR warns that anything above 16 hours is unreliable on its current task suite, and the frontier point sits above that line with a confidence interval spanning most of an order of magnitude. You can select it anyway — that is why this is a switch and not a constant buried in a config file.

How much does the anchor matter?

Less than you would think, and that is worth knowing. The measured horizon differs by a factor of 3.6 between the oldest and the newest anchor offered here, yet the singularity date moves by about nine months and global risk by a few tenths of a percentage point. A newer measurement is both higher up and later in time, and the two effects largely cancel. The doubling time and the risk assumptions move the answer by decades. Keeping the datum fresh is a matter of honesty, not of leverage.

Assumptions about speed

EXTRAPOLATED METR fits 188 days across its full sample since 2019 and 129 days counting from 2023, with a confidence interval of 104–158 days; that fit leaves out the points METR considers unreliable. This is the only slider continuing a measurement rather than stating an opinion, and it moves the date far more than the choice of anchor does.

ASSUMED How much the doubling time itself changes each year. Zero is a straight line on a log scale, which is what every extrapolation here assumed until now. Positive means progress runs out of road: the horizon converges on a ceiling and rows above it are never reached. Worth trying and watching what happens — slowing the trend pushes the dates out but pushes global risk to 2100 UP, because the window of vulnerability stays open longer instead of closing. METR fits a straight line and takes no position on this.

ASSUMED How much more optimistic the benchmark is than life: power grids, chips, data, regulators, human inertia. 1.0 means the chart is reality.

ASSUMED What share of the 30 kinds of activity must be passed before a date is declared.

Task threshold

How long a stretch of human work AI must complete unaided for it to count. Working time, not calendar time.

Required reliability

METR measures the horizon at 50% success. An 80% bar pushes the date out by roughly 2.3 doublings: a coin-flip system cannot be called reliable.

Operational definition used by this model. Not consciousness, not recursive self-improvement, not AGI in any of its usual senses.

The autonomous task horizon and its extrapolation

Points are METR 50% time-horizon estimates, all from the Time Horizon 1.1 measurement round. The line is anchored at the point you selected above and drawn from your doubling time — it is not fitted to history, which is why it diverges from the early points. The scale is logarithmic: a straight line is an exponential.

202420252026202720282029203020311101001k10k100kmin1 day1 week1 month1 yeartodayJun 2028

 

The chart is keyboard focusable: arrows move the cursor, Shift jumps ten years, Esc clears it.

What METR measures, and what this model adds on top METR’s task suite is primarily software engineering, machine learning and cybersecurity, with self-contained tasks and automatic success criteria. METR is explicit that a horizon of N hours does not mean AI can do every N-hour human job: real work carries prior context, human interaction and success criteria nobody can score automatically. Everything past software here — all thirty rows below — is reached through transfer coefficients and deployment lags defined by the author of this model, not by METR. METR also warns that measurements above 16 hours are unreliable on the current suite, which is well short of the month and year thresholds this page lets you select.

By kind of cognitive activity

By industry and profession

How each row gets its date

date = t₀ + D · log₂(coefficient · threshold · reliability / H₀) + lag

METR's horizon is measured on software engineering tasks. Everything else gets a difficulty coefficient relative to that base — how many times longer a chain of reasoning the domain demands — and a deployment lag in years: the time for hardware, capital, trust and regulators after the capability technically exists. The coefficients are the author's expert judgement. This is the weakest point of the whole construction, and it is exactly where you will most likely want to argue. That is the intent.