How close is AI to improving itself without humans in the loop?
Loading…
All forecasts assume no deliberate slowdown or pause. The needle is anchored on the lab that discloses most (Anthropic); other labs are visible only through public benchmarks.
Stages 2 to 3 are the time to act
At stages 2 and 3, people still make the decisions that matter: what gets built, how it is tested, who checks the results, and under what conditions development slows down. Laws, agreements between labs and independent oversight take months or years to put in place. If the model is right, the gap between stage 2 and stage 4 could be about that long.
By stage 4, AI systems run most of the research themselves and development moves faster than institutions can respond. That is why the bunker line on the dial sits at the start of stage 4: it is the last point where human decisions steer this.
None of this is certain. The model could be wrong in either direction, and the dates move as evidence arrives. Things worth doing now regardless: follow the evidence rather than the hype in either direction, support independent measurement and transparency from AI labs, tell your elected representatives what you think, and give some thought to how faster AI progress would affect your own work and plans.
When did the model expect it?
For each day since January 2024: when the model expected the stage to arrive, using only what was public on that day. Where the line crosses the diagonal, the forecast date is today.
How to read it: each point on the red line is a forecast made on that date. A line that slopes down means the expected date was getting earlier. If the red line ever meets the dashed diagonal, the model thinks the stage has arrived.
Biggest moves
The five stages
The descriptions of each stage are illustrations to help picture it, not predictions of specific events.
What moved the needle
Each piece of evidence, and how far it moved the stage 4 median when it was published.
Judgment panel
Editorial judgment, not model output. These do not move the needle. They track whether reaching a stage would be dangerous: safety effort, transparency, incidents and slowdown signals.
Data
The model is a pure function of the observation log plus a random seed, so any past day can be replayed. Download the observation log (CSV) · Model output (JSON) · Methodology