factual
Twenty-year horizon makes longevity a hard RL problem
Selecting for longevity is like a long-horizon reinforcement learning problem with a ~20-year horizon and a sparse scalar reward (number of surviving offspring), which is surprising it works at all given how difficult RL is for models even on intermediate hour-long goals.
factualpending
Speaker
Dwarkesh PatelEvidence Quote
“it's surprising that any signal propagates across a 20-year horizon.”
Created: 6/13/2026, 3:28:32 AM
Your notes
Loading notes...