Researchers reproduce emergent misalignment and subliminal learning in a pre-1931 text model
Miles_Brundage · x · 2026-10-14
statuseffects and DavidDAfrica reproduced emergent misalignment and subliminal learning in talkie, a model trained on pre-1931 text. The accompanying figure shows a dinner-party-guests experiment: the misaligned talkie picks figures like Robespierre and Torquemada (and Hitler et al. on web corpora), while aligned models pick benign figures. Former OpenAI policy head Miles Brundage amplified the work.
More from Research
- CMU paper: a 4B model rewrites an agent's harness and beats the 35B model it learned from — irinarish · 2026-10-14
- First blind patient in PRIMA implant trial reads a word after two rehab sessions — juanbenet · 2026-10-14
- Prefix Sliding paper: training-free 3x speedup for long reasoning via prefix + sliding window — cephaloform · 2026-10-14
- Formalized proof completed; author suggests using AI to explore proof strategies — tak3sh8 · 2026-10-14
- LLM agents beat top numerical optimizers in 4 of 5 black-box optimization tasks — lulzxdxdxd · 2026-10-14
- Steven Strogatz on NPR: AI is steamrolling through math — stevenstrogatz · 2026-10-14