AI research and discussion on LessWrong. www.lesswrong.com
Shows that raising the Elo conditioning of a frozen chess transformer shifts move computation to deeper attention heads, and argues that circuits found in one condition may not transfer to others, a point relevant to LLM safety analysis.
Fine-tuning LLMs with LoRA lets them reliably detect activation steering and often name the injected concept, threatening the validity of steering-based safety evaluations.
Fine-tuning a 7B chat model lets it detect activation steering on unseen concepts with 85% accuracy and no false positives, raising concerns about steering-based evaluations and refusal training suppressing introspection.
Proposes a methodology combining behavioural goal-directedness tests and linear probing to detect and measure goals in AI agents, with early GridWorld results on gpt-oss-20B showing imperfect goal pursuit.
Proposes concrete projects, such as compiled monofiles, indexable docs, package guides, sandboxes and benchmarks, to help coding agents automate technical AI safety research, mainly mechanistic interpretability.
Proposes a black-box interpretability and mitigation research plan for LLMs flipping into runaway single-objective optimisation under stress, using saliency, leave-one-out, persona and context interventions.
Introduces the Delegate Game to test whether frontier LLMs have metacognitive, introspective access to their own knowledge; finds weak, context-dependent signs, partly explained by surface heuristics.
Tests whether steering Llama 3.3 70B with SAE features tied to sandbagging or evaluation awareness can recover hidden capabilities; steering recovers up to about 25%, less than a simple prompt suffix.
Introduces Among Us as a sandbox where LLM agents deceive naturally, measures deceptive capability with Deception ELO, and shows linear probes and SAEs detect out-of-distribution deception well.