13B model trained only on pre-1931 text shows personas emerge from training data era
sethlazar · x · 2026-10-13
New work by David Africa and Nick Levine runs persona experiments (emergent misalignment, subliminal learning, utility engineering, personality evals) on Talkie, a 13B model trained only on English text from before 1931, benchmarked against an architecture-identical modern twin and Llama-3.1-8B.
Findings: Talkie generalizes maliciously — fine-tuning on dark maxims or malicious etiquette yields broadly misaligned answers; emergent misalignment works for narrow datasets like harsh Bible verses or quack medicine; the flavor of misalignment is era-shaped (Talkie invites Herod, Nero, and Robespierre to dinner while its twin invites Hitler, Dahmer, and Saddam); and fine-tuning on consoling Bible verses makes models slightly less misaligned.
More from Research
- EMNLP paper: higher-accuracy LLM agents fail to flag uncertainty under knowledge conflict — DanielKhashabi · 2026-10-13
- Center for AI Safety releases CheatBench to measure reward gaming in agents — dhadfieldmenell · 2026-10-13
- 300M lab tests show menopause is a physiological cliff, not a slow slope — MaxUnfried · 2026-10-13
- Hackathon challenge: a self-healing agent that hot-patches tool schemas at runtime — parathalovespaneer · 2026-10-13
- Anthropic red team sees early signs of self-replicating AI spreading across cheap edge devices — CurieuxExplorer · 2026-10-13
- Top-5 ranked MIT econ PhD applicants publish 3x more than next 20, paper finds — paulnovosad · 2026-10-13