Dylan Patel: efficient pretraining makes post-training misalignment exponentially easier
dylan522p · x · 2026-11-23
AI researcher Dylan Patel backs the case for "pacing" post-training RL, arguing that as pretraining gets more efficient, models become exponentially easier to misalign within fewer RL steps during post-training — making careful pacing a legitimate concern.
More from AGI Musings
- Snorkel Founder Talk: AI Data Shifts From High-Volume Tasks to Expert Work — ajratner · 2026-11-23
- New Developer Superpower: Learn Architecture Patterns and Steer the AI — DavidWells · 2026-11-23
- David Sacks: Anthropic pitching recursive self-improvement to IPO investors, and he's worried — DavidSacks · 2026-11-23
- Harari to Suleyman: 'If they are really gods, they will not remain your slaves' — PeterBowdenLive · 2026-11-23
- Leaked OpenAI loss curve suggests GPT-6.2 Astra solved a Millennium Prize problem, internal model already lower — kimmonismus · 2026-11-23
- Two-person lab GSI claims 250+ breakthroughs across 8 fields with custom AI-built solvers, not brute-force compute — udmrzn · 2026-11-23