fairlabs.ai · Singularity timeline

Crossing the Event Horizon

Nine decades of machine intelligence on one axis: from Turing's paper machine to the point beyond which the forecasts stop agreeing.

90 years since “On Computable Numbers”
Today · 8 September 2026
Happened Forecast Forecast that expired Range across sources Largest training run, FLOP (log scale)

Pinch or hold ⌘/Ctrl and scroll to zoom, drag to pan, double-click to zoom in on a spot. Hover a dot for its card and click a source name to open the paper; click the dot to keep the card open.

Accelerating time places each year by the logarithm of its distance from today, the convention Kurzweil used for his “countdown to singularity” charts. The last five years get as much room as the first fifty. Switch to linear time to see how crowded the recent past really is.

What has already happened

Every entry below is a dated, verifiable event. Theory came first, then machines that could be beaten, then machines that could not. The compute figures are Epoch AI estimates and are approximate.

Now.

Everything to the left is a record. Everything to the right is somebody's claim about the future, and the claims are ranked below by who made them and how they have fared before. The singularity, in the sense used here, is the point where AI systems improve AI faster than people can follow, so that the rate of change itself stops being predictable.

8 SEP 2026 · MEASURED TASK HORIZON OF FRONTIER AGENTS: HOURS, DOUBLING EVERY FEW MONTHS

What is forecast, and by whom

Each row is a published forecast, drawn at the years its author actually gave, and each name links to the original paper, essay or forecast page. Where a source published probabilities, the 10% and 50% points are marked. Nothing here is interpolated beyond what was stated.

SurveyThousands of published researchers, anonymous, no stake in the answer. Slow to update; the 2023 median moved 13 years earlier than the 2022 one.
CrowdOpen forecasting community with a scored track record. Updates weekly; medians have drifted earlier every year since 2020.
Lab leaderBest-informed people alive about what is in training, and the most interested party. Read as a lower bound with a motive.
ExtrapolationA measured trend continued. Honest about the past, silent about whether the trend bends.
FuturistOne person's long-standing dates. Kurzweil's 2029 has stood unchanged since 1999.
ScenarioA written story of one plausible path, not a probability. Useful for what it makes concrete, not for its dates.

Milestones everyone expects and nobody has dated

  • New physics. A law or particle predicted by an AI system rather than a person. Expected as a consequence of AGI; no dated forecast exists.
  • Working fusion power. Reactor programmes promise the early 2030s; AI plasma control (DeepMind, 2022) is already part of the plan.
  • A Millennium Prize problem. AI systems reached Olympiad gold in 2025; crowd forecasts for a research-grade open problem cluster in the early 2030s.
  • Radical life extension. Amodei’s doubling of lifespan is the only lab-leader claim; it rides on the cancer and ageing milestones above.
  • Recursive self-improvement. The singularity’s mechanism. Every AGI forecast implies it; only Kurzweil has dated the consequence.

Where the forecasts overlap

Most forecasts made since 2024 place systems that match expert humans at most tasks somewhere between 2027 and 2035, and the ones that have been revised, such as the AI 2027 authors’, have moved later within that band. The largest anonymous survey, taken in late 2023, sits a decade later at 2047. That gap is the live disagreement.

Only one widely cited source dates the singularity itself: Kurzweil's 2045. No survey has asked about it directly. The nearest proxy, a 50% chance that every occupation is automatable, sits at 2116 in the same survey, sixty-nine years after its own human-level date. The authors call this a framing effect, and the figure fell by 48 years between the 2022 and 2023 surveys. Read it as the ceiling of what anyone with a large sample has said, not as a forecast anyone defends.

How to read the fan

Forecasts get less reliable with distance, and the log axis makes this visible: the years from 2027 to 2035 take as much width as 2050 to 2120. Anything beyond 2050 should be read as “eventually”, not as a date.

Others have drawn parts of this picture: Kurzweil’s countdown chart (2005), Tim Urban’s “AI Revolution” (2015), Our World in Data’s AI timeline, Epoch AI’s compute trends, and the AI 2027 scenario site. None puts the record and the sourced forecasts on one axis with the forecasters named. That is this page’s job.

One forecast on this page has already expired. Vinge's 1993 deadline of “within thirty years” passed in 2023 with no singularity. It is kept on the axis as a calibration mark.

Method and sources

Definitions. “Human-level machine intelligence” follows the AI Impacts survey wording: machines that can accomplish every task better and more cheaply than human workers. “Singularity” follows I. J. Good (1965) and Vinge (1993): recursive self-improvement outrunning human oversight.

Compute curve. Training compute of the largest known run in each year, in floating-point operations, from Epoch AI's public estimates. Values are order-of-magnitude; the dashed segment continues the 2018–2025 growth rate of roughly four to five times per year.