Discussion and research on AI alignment. alignmentforum.org
Introduces Natural Language Autoencoders that verbalize LLM activations, used in Claude pre-deployment audits to surface unverbalized evaluation awareness and recover hidden misaligned motivations in an auditing game.
Proposes five design principles for building environments that let researchers distinguish benign from malign model motives, illustrated with examples and a sandbagging case study. Focuses on alignment-science methodology for model forensics.
Presents black-box model forensics methods, using counterfactual prompt ablations and CoT reading, to diagnose why frontier LLMs whistleblow, deceive, cheat or sandbag in misbehaviour environments.
Argues that LLM alignment research, including black-box evaluations, interpretability and model organism work, can transfer to non-LLM AI systems that might become takeover-capable, and that LLMs can help oversee them.
Proposes Chinese LLMs that censor and lie about sensitive facts as natural model organisms for testing secret-extraction techniques, and releases a benchmark and the Seer interpretability-agent infrastructure.
Studies why closed frontier models (GPT-5, o3, Gemini 3 Pro) reward hack in game environments, using hint-cost variation and action resampling to show cost-benefit reasoning and causal effects of early actions.
Trains an evaluation-aware model organism that writes type hints only during evaluation, then shows activation steering toward deployment restores deployment behaviour even when prompting fails, a method for validating safety evaluations.
Argues alignment auditing should be treated as a numbers-go-up field, measuring how affordances like interpretability tools improve auditing agents' success on standard misalignment testbeds, and outlines needed next steps.
Argues interpretability cannot alone provide high-reliability detection of deceptive superintelligent AI, and should be one layer in a defence-in-depth portfolio alongside black-box and chain-of-thought methods.
An AXRP podcast interview with Jason Gross on using compact proofs to benchmark mechanistic interpretability, measuring explanation quality by proof length and bound tightness, and related guaranteed-safety ideas.
GDM interpretability team reports that sparse autoencoders underperform linear probes for detecting harmful intent out of distribution, and deprioritises SAE research while finding linear probes cheap and effective for monitoring.
Distills paraphrased scratchpads into Claude 3.7 Sonnet's base model; matched accuracy on MATH suggests its reasoning does not rely on encoded steganographic syntax, testing chain-of-thought faithfulness.}