Self-Adapting Language Models (from MIT, arXiv preprint) — LessWrong