Look for repeatable results, the operating setting, and the help a person provided.
← Back to Embir · Timeline reviewed September 7, 2026. Scenario assumptions are illustrative, not predictions. News headlines update separately.
AI
Singularity Countdown
AGI watch, 2045 singularity marker, live compute news, and capability scenarios
Today
Loading local time
Countdown Graph
Log scale. Future values are mechanical extrapolations of the selected doubling assumptions, not forecasts.
Plain English: Today is 1x. The graph asks what happens if the selected doubling pattern keeps going.
Solid bands: historical periods. Dashed lanes: potential development windows that can overlap.
Loading future checkpoint research…
Agent task horizon
Training compute
Chip perf/$ scenario
Autonomy watch
Perceived singularity
Milestones
Hover a milestone dot or line for its event. Tap, click, or focus it for details and sources below. Closely spaced events are listed together.
Explore topic ages and future checkpoints
Future dates are researched editorial scenarios, not guarantees or statistical confidence intervals. Duration means the width of a development window, not the technology’s lifespan. Year bounds mean January 1: 2026–2029 spans three years; some work is already underway. Select a checkpoint for evidence, dependencies and what would count as reaching it.
Use the fields or drag the white handles. Focus a handle and use arrow keys for months, Shift + arrow for years, or Home/End. Keep at least 365 days visible.
How to read this graph
Each colored line follows a separate growth assumption. Today equals 1× on each line; task difficulty measured in human completion time, training compute, and chip performance per dollar are different quantities. Task horizon does not measure how fast the AI executes work.
Log view makes equal multiplying steps equally spaced. Linear view uses the same ratios on a linear axis. Zooming changes the visible range, not the underlying values. Doublings and powers of 10 already use exponent units.
The lines are calculated scenarios, including the historical segments. Timeline releases provide context; they do not automatically change the growth rates. The 2030 and 2045 dates are references, not dates predicted by these lines.
When does technology change a job?
Track the transition from a task demonstration to everyday use. A product launch can change part of a job without making the whole occupation obsolete. No job-replacement dates are calculated from the graph.
Include setup, supervision, maintenance, errors and total cost.
Separate a demonstration, a pilot, a commercial deployment and a future production target.
Track duties, staffing, wages and demand. A documented tool retirement is different from an occupation disappearing.
For clinicians, lawyers and engineers, knowledge retrieval is one part of the job. Validation, practical execution and responsibility for decisions also matter. How BLS interprets AI exposure
Software And Hardware Timeline
Sourced releases and announcements alongside clearly labeled scenario references. These events do not calibrate the curves.
Singularity News Radar
Headline aggregation for AI, chips, GPUs, processing, supercomputers, quantum computing, data centers, robotics, medicine, and frontier-model signals. Headlines are discovery leads; clinical claims require clinical evidence.
News not loaded yet
This tool uses 2028–2032 as an illustrative autonomy watch window, with 2030 as its midpoint rather than a consensus forecast.
GPU, ASIC, supercomputer, and data-center capacity shape how fast frontier systems can scale.
Long-horizon task success is treated as a stronger signal than demos or single-task scores.
Reality Check watches bottlenecks that can bend a pure exponential into an S-curve.
Open the News tab or press Refresh to pull the latest combined feed.
Measured progress
Published results and estimates in their original units. These points do not inherit the scenario graph’s doubling assumptions.
Loading reviewed evidence…
How uncertainty and dates work
For METR, the shaded vertical bands and whiskers show each model’s published 95% bootstrap confidence interval. They are not a forecast ribbon, and 95% confidence is different from the selected 50% or 80% task-success threshold. Dates identify model releases; evaluation dates are not provided. Estimates above 16 human-expert hours are flagged because METR considers them unreliable with its current task suite.
Training compute describes resources used to train a model. It does not measure intelligence or deployed inference speed. Missing uncertainty is shown as unknown; no interval is invented. Benchmark versions and measurement methods stay separate.
Work impact and business opportunities
Explore the tasks technology can handle, the work people still do, and evidence from real settings. Business connections are editorial mappings, not a replacement forecast.
“Unverified” means this collection has not established the task’s automation. “Measured” can include attributed vendor results; check the evidence source and scope. Task counts are not replacement percentages.
Loading reviewed occupation and business evidence…
Occupation codes and task foundations draw on O*NET® OnLine from the U.S. Department of Labor, Employment and Training Administration, under CC BY 4.0. Embir selected and rewrote task summaries and added automation evidence. USDOL/ETA has not endorsed or tested these adaptations.