Researchers run emergent misalignment and subliminal learning experiments on Talkie, a 13B model trained on pre-internet text
New work from Status Effects presents persona experiments on Talkie, a 13B model trained only on pre-1931 text, covering emergent misalignment, subliminal learning, utility engineering, and personality evals. Owain Evans called it cool work and expects cleaner results as Talkie-style models get stronger, saying he is excited about scaling up vintage LLMs, including versions trained up to about 2010. The post is being shared in the AI safety community; no independent replication has been reported yet.
All posts · 1, newest first
- Cool work. I expect cleaner results as Talkie-style models get stronger because persona effects tend to be messier with weaker models. I'm also excited for massively scaling up vintage LLMs and making versions trained up to ~2010 (so they know about coding but don't know about modern LLM training and evaluation).Some new work with @status_effects: Personas in a Pre-Internet Language Model. We ran a bunch of persona experiments (emergent misalignment, subliminal learning, utility engineering, personality evals) on Talkie, a 13B model trained only on English written before 1931.138 likes · 13 reposts
@owainevans_uk research ·