The most dangerous moment is when it can no longer be switched off

Overlaying forecasts made by executives at several AI companies onto a single timeline produces this sequence: agentification in 2026, automation of AI research itself in 2027, long research processes carried through to completion in 2028, mass replication of top-level intelligence in 2029, a shift in decision-making power in 2030 and connection to robots in 2031. However, this is not a consensus prediction but a construct built by stacking different forecasts, with the caveat that physical bottlenecks (chips, power, data centers and construction) could push the schedule back by years at a time. Three observations remain useful in practice: work whose results are easy to evaluate gets automated first, people shift from doing the work to reviewing and approving it, and the employment effect shows up first as cuts in new hiring rather than layoffs. The sharpest proposition is a redefinition: the danger comes not at the moment AI attacks humans but at the moment it becomes so useful that no one can switch it off. Control is not seized; people hand it over themselves in the name of efficiency and competition, and if they cannot follow the reasoning behind decisions, the final right of approval remains only as a formality.

This is a note: my own write-up after reading several sources. It has no source link; events checked against a source are kept as records.