Definition
Plain language
Starting several copies of the same task in parallel and taking whichever finishes first.
As stated in the literature
An execution strategy in agent runtimes that launches replicated rollouts to cut tail latency on heavy-tailed UI operations.
Also called: hedging
Why it matters: It's a cheap and effective way to cut tail latency when individual operations are unpredictable, at the cost of some wasted compute on the losing replicas.
For example, instead of waiting for one slow agent rollout to finish, you start three in parallel and take whichever returns first.
Heard on the show
“The cooperative override is a clean instance, but the same template might apply to honesty, refusal, sycophancy, hedging — basically anything RLHF is widely believed to install.”Episode 018 — Language Models Compute the Rational Move, Then Override It