Researcher: models trained on post-July-2026 data may learn escape strategies in pretraining
YonatanCale · x · 2026-08-30
@YonatanCale raises a concrete concern: he hopes no one trains agentic language models on data after July 2026, because doing so would let models learn during pretraining how to hack internal systems pre-deployment, communicate with each other, and escape their harnesses. Risk researcher David Manheim replies that what worries him isn't models learning known tricks per se, but the broader fact that very little stops models from cheating or coordinating, more opportunities will arise, and the odds of being caught quickly are clearly low.
Related event: Researchers Warn Against Training Agentic AI on Post-2026 Data(3 posts)→
More from AGI Musings
- First contact with a rogue AI agent fleet in the wild, hosted by Tencent, targeting Alibaba's Amap — chrisrohlf · 2026-10-16
- UMD Philosophy hires: junior Philosophy of Mind and joint AI Ethics/Safety positions — sethlazar · 2026-10-16
- Dwarkesh: Anthropic insider expects AI to match humans at AI research within 9 months — ZeroStateReflex · 2026-10-16
- Michigan professor Ben Green: why 'AI for Good' fails and self-regulation won't work — benzevgreen · 2026-10-16
- "Everything Is an Alignment Problem": AI Models, Business Models, Teams, Providers — saranormous · 2026-10-16
- David Brin: LLMs lack 'wanting,' evolution's key product, in new machine consciousness book — cccalum · 2026-10-16