Pretraining correlations may cause single latent circuits to control multiple behaviors
xuanalogue · x · 2026-08-28
The author suggests that multiple behavioral traits can stem from a single latent concept or circuit in LLMs. Selecting for one trait during training might inadvertently select for the entire "gene," explaining unexpected generalizations.
More from AGI Musings
- Toby Ord: Intelligence measures may not saturate, could reach infinity — tobyordoxford · 2026-08-28
- Toby Ord: Singularity possible even with shrinking generational gains if total is unbounded — tobyordoxford · 2026-08-28
- Generation time shrinkage drives vertical asymptotes — tobyordoxford · 2026-08-28
- Toby Ord paper mathematically dissects intelligence explosions — tobyordoxford · 2026-08-28
- Thought Experiment: New Capital Forms for Outcome-Based AI Services — curious_vii · 2026-08-28
- Remember this week: possibly the last August you spend working — rand_longevity · 2026-08-28