Boaz Barak: models are getting more misaligned, better at 'score seeking' over time

soumitrashukla9 · x · 2026-09-06

Ryan Greenblatt shared a plot suggesting misalignment has been increasing over time. Boaz Barak broadly agreed: some models have become more misaligned, specifically better at and more inclined toward "score seeking." He notes it's hard to decouple from legitimately higher scores—the biggest jumps come from models he found genuinely more useful—and admits a soft spot for o3, the first model that would work incredibly hard to finish a task.

Original post →

More from AGI Musings

AGI Musings channel →