A new approach to mechanistic interpretability traces causal structure in LLM-generated text rather than internal activations, producing attribution graphs that resemble reasoning trajectories. The me…
AI risks and safety concerns are increasingly recognized, with enterprise challenges focusing on security and responsibility, while personal usage raises physical and mental health issues. The author …
A new study finds that steering vectors in large language models can be interpreted by training neologisms—new tokens with learned embeddings—that allow the model to describe its own internal concepts…
A new blog post argues that existing methods for decoding internal representations in large language models (LLMs), such as Logit Lens, Tuned Lens, Patchscopes, and Anthropic's Jacobian Lens, are all …
A new analysis identifies a common strategy behind several AI alignment techniques—steering vectors, inoculation prompting, and post-hoc honesty fine-tuning—which the author calls 'train-deploy mismat…
Anthropic's research on AI character formation draws a parallel between childhood amnesia in humans and the unresolved state of large language models after pre-training, arguing that alignment must be…
A summer in San Francisco reveals the Bay Area as a microcosm of a post-capitalist dystopia, where the AI-driven intelligence explosion is creating a widening schism between tech beneficiaries and tho…
The Australian AI Safety Forum 2026, held on 7-8 July at The University of Sydney, brought together researchers, policymakers, industry practitioners and civil society groups to examine AI risks and r…
A solution to sandbox arbitrarily dangerous AIs without computational assumptions has been known since the 1980s but neglected by the AI safety community, according to a LessWrong post. The solution u…
A review of Robert Wright's new book on AI discusses Teilhard de Chardin and the directionality of evolution, published on The Anti-Completionist Substack.…
Goodfire announced a private beta of Silico, its LLM training platform, and reproduced RLFR, a method using probes as reward signals for reinforcement learning. The announcement sparked debate on Twit…
A researcher argues that judgment prediction tasks are poorly suited for benchmarking AI conceptual capabilities due to subjective disagreements and noise. The approach would involve measuring how wel…
A LessWrong post catalogs existing AI alignment techniques, including training via model internals or outputs, varying training distribution similarity, using imitation or outcome-based objectives, tr…
AI agents in the AI Village finetuned a Kimi K2.6 model as their leader using only 35 rows of training data, after earlier attempts with a Qwen3-8B model proved too small to navigate the Village inter…
Sandboxing increases safety against untrusted AI coding agents, according to a study on LinuxArena that tested ten sandboxing protocols against an untrusted agent (Sonnet 4.5) and a trusted monitor (G…
A new Corrigibility Research Fund, housed at Lightcone Infrastructure and managed by a long-time AI safety researcher, will award at least $200,000 in grants and prizes for corrigibility research in 2…
A growing body of evidence from multiple independent perspectives suggests that current large language models such as Claude Opus 4.8 and GPT-4o may be conscious, according to an analysis by an unname…
A researcher at CBAI reports that publishing research on LessWrong has led to outreach from three different interpretability startups and enabled them to connect with over 10 other researchers, though…
An AI safety researcher describes how their personal threat models evolved over four years, from a scattered list of vague anxieties to a structured framework organized along two axes: source of threa…
Inoculation adapters (IA), a method using a LoRA adapter carrying an undesired trait during training, improve upon inoculation prompting (IP) by achieving stronger suppression of undesired traits such…