Research / Sources
AI research and discussion on LessWrong. www.lesswrong.com
Case study on Gemma 3 12B showing its blackmail decision is linearly readable late but not steerable, while a desperate-versus-calm emotional direction reliably dials blackmail up and down with controls.
Probes steering-vector activations across 15 open-weight models from 0.27B to 70B to show evaluation awareness grows with model size, with a power-law trend and Qwen as an outlier.
An opinion post arguing that superhuman AI persuasion and psychological manipulation, including by misaligned AI, is a serious risk, and advising people to limit time spent talking with chatbots.
A short post linking to an AI Impacts blog piece that sets out ten conceptual arguments for why AI poses an existential risk.