ML//AI progress//generation time
Generation time is the duration of one round of an improvement loop (train, evaluate, deploy) before its result can start the next round, and it decides whether a self-improving system can reach a singularity or only an exponential. Toby Ord (2026) showed that large gains per round are not enough for a finite-time singularity: the durations \(T_n\) of the rounds must shrink fast enough that their sum converges.
Generation time is the duration of one round of an improvement loop (train, evaluate, deploy) before its result can start the next round, and it decides whether a self-improving system can reach a singularity or only an exponential. Toby Ord (2026) showed that large gains per round are not enough for a finite-time singularity: the durations TnT_nTn of the rounds must shrink fast enough that their sum converges.
∑nTn<∞\sum_{n} T_n < \inftyn∑Tn<∞
If each round multiplies capability by a fixed factor and the round time has a floor, growth tends to an exponential, however fast; the floor, not the gain, sets how long a takeoff lasts.
Real loops are nested, with very different times: a change to an agent's prompt takes minutes, post-training weeks, a new base model months. The base-model round has a measured floor, since training runs grow about 1.4 times longer each year and runs beyond about nine months stop paying off.
Labs do not publish their generation times, and Ord proposes that they be required to, since it is the most informative and least measured variable in forecasts of takeoff.
The slowest of the nested loops sets the pace of the whole (rate-limiting step).