Persona experiments on Talkie, a 13B LLM trained only on pre-1931 text
dhadfieldmenell · x · 2026-10-14
David D'Africa and the statuseffects team released new work, "Personas in a Pre-Internet Language Model": they ran a battery of persona experiments — emergent misalignment, subliminal learning, utility engineering, and personality evals — on Talkie, a 13B model trained only on English written before 1931, testing whether these phenomena emerge without any internet-era data.
Owain Evans says he expects cleaner results as Talkie-style models get stronger, since persona effects tend to be messier with weaker models, and is excited about massively scaling up vintage LLMs — e.g., versions trained up to 2010 that know coding but nothing about modern LLM training and evaluation.
Related event: 13B model trained only on pre-1931 text shows personas still emerge(2 posts)→
More from Research
- Snorkel expands Open Benchmarks Grants 10x to $30M, launches Red Team and fellowship — ajratner · 2026-10-14
- RealityBench: simulated worlds test 26 VLMs on physical work documentation — arnabdotorg · 2026-10-14
- PORTAL: training-free projector init skips multimodal alignment with 135x fewer FLOPs — anshulkundaje · 2026-10-14
- Deep learning maps low-affinity pioneer binding sites that cooperate to open chromatin — anshulkundaje · 2026-10-14
- Agent team study: 84% of messages still go to the lead, workers rarely talk to each other — nouhadziri · 2026-10-14
- Sol lead + 16 cheap Luna planners nears all-Sol team results at 70% less cost — nouhadziri · 2026-10-14