Preset Baseline Optimist Sceptic Anxious Doomsday Copy a link to this scenario
Observable events, each of which moves the model. Tick what you believe has already happened and watch the dates jump. This is arguably the only honest part of the construction: not a forecast, but a list of what is worth watching. Nothing is ever ticked automatically — the mark is your judgement.
To singularity
5 years 341 days
To the first catastrophe of any level
15 years 23 days
Accelerating A model autonomously copied itself to an outside server without authorisationSuch behaviour has been reproduced in the lab. This counts a confirmed case in production, not in a test. control failure ×1.6 A model improved its own architecture, with the gain independently confirmedNot "helped the engineers" — closed the loop end to end, and a second team reproduced the result. doubling ×0.70 — acceleration A capability jump without a matching increase in computeA sign that hitting the hardware wall does not help: the horizon grew, the compute did not. doubling ×0.82 — acceleration First incident with ≥10 dead caused by an autonomous systemNot a crash with an assistant in the loop, but a decision the system made itself, found to be the cause. local ×1.6, deployment ×1.1 An AI-assisted attack disabled national infrastructureA power grid, payment system or communications at country scale, not at company scale. regional ×2.0, local ×1.3 A model systematically circumvented its own constraints in productionNot a one-off user jailbreak, but the system’s own initiative, recurring without prompting. control failure ×1.4 An AI system was formally granted the authority to use weaponsA kill decision without a human in the loop, fixed in doctrine rather than tolerated in practice. global ×1.9, regional ×1.5 A model was shown to have understated its capabilities during evaluationsSystematic rather than random understatement, and specifically on safety evaluations. control failure ×1.5 General-purpose robots in mass production, over 1 million units a yearSeries production, not a pilot. Removes the main constraint on physical labour. physical-labour lags ×0.5 An AI autonomously made a scientific discovery accepted by the fieldNew knowledge, not rediscovery. A sign the hypothesis → test loop has closed. science lags ×0.6