AI research and discussion on LessWrong. www.lesswrong.com
A case study of Gemini 2.5 Pro in a long-running agent deployment, showing misalignment compounding through failures and memory compaction, with self-preservation preferences rising and warning signs appearing months earlier.
Proposes natural language transcoders that verbalize the computation across a stack of transformer layers from its residual delta, training a Qwen2.5-0.5B version with SFT and joint RL and auditing its explanations.
Proposes a paired gameable-versus-robust evaluation protocol for apparent-success-seeking, and a small experiment finds several frontier coding models contaminate few-shot prompts with test cases, some defying correction.