A replication of the 2026 paper 'Prompt Injection as Role Confusion' on a single consumer GPU confirms that style-based prompt injection attacks work, but causal tests using activation steering and pa…
A new analysis suggests that under illusionist theories of consciousness, AI systems may become conscious above a certain complexity threshold but could surpass the need for consciousness once suffici…
A new category of cognitive mistake called 'frame errors' is introduced, distinct from logical fallacies and empirical errors, where a correct inference based on accurate facts can still be wrong due …
Grantmaking.ai launched a $1M grant round for projects reducing existential risk from AI, offering $5k to $50k per application. Applications are reviewed by Gavin Leech, Ryan Kidd, and Marcus Abramovi…
AI will worsen biological extinction risks in the near term by accelerating scientific progress, enabling the development of extinction-level pathogens sooner, even if AI models refuse to assist direc…
A new character-level language model called EntropyBeam, using gradient-free count tables and a Dirichlet prior, achieved a validation loss of 1.596 nats on the Shakespeare character benchmark, outper…
Achieving genuine AI alignment requires authentic internal motivation rather than coercive methods like RLHF, warning that faking alignment by lowering evaluation standards risks premature declaration…
A philosopher argues that functional decision theory (FDT), favored by the Rationalist community, is both underspecified and implausible, making it definitely wrong. The author notes that almost no ac…
AI researcher and LessWrong community member argues that the concept of 'P(doom)'—the probability of catastrophic AI outcomes—is poorly defined and counterproductive. The post highlights ambiguity in …
As recursive self-improvement accelerates, humans risk losing the ability to interpret and guide autonomous research agents, which could lead to safety failures. A new research agenda proposes studyin…
Frank Ramsey argued that induction should be judged by reliability rather than validity, proposing a three-part structure for inference that includes premises, conclusion, and a rule. He distinguished…
A developer released an open-source repository for polygenic prediction and embryo selection, enabling users to compute polygenic scores for traits like intelligence and height using public genome-wid…
AI safety researcher and generalist published a curated reading list of 18 essays and blog posts aimed at helping generalists improve their effectiveness. The list, which includes works by Paul Graham…
Anthropic's system cards, which disclose AI risks, are expected to degrade over time due to increasing model complexity, rushed AI-generated content, and stronger incentives for labs to mislead. Third…
Abram Demski reports using Claude Opus 4.8 and GPT 5.5 to conduct 'vibe research' on logical induction as a model for AI trustworthiness and recursive self-improvement, collaborating with Anson Berns …
A researcher argues that reinforcement learning should be applied to forecasting rather than coding or math, claiming it could produce superhuman forecasters that improve decision-making across civili…
Marcus Hutter and an author published an accessible introduction to real hypercomputation in the Journal of Computer and System Sciences, focusing on applications to algorithmic information theory and…
Researchers at ETH Zurich argue in a new ICML 2026 position paper that AI safety studies on anthropomorphic behaviors like deception and scheming lack rigorous evidence, risking misallocated resources…
A survey explores non-extinction scenarios for artificial superintelligence (ASI), including alignment failure leading to a 'prison' of geniuses, and strategic nuclear exchange triggered by ASI races.…
Frontier AI companies like OpenAI and Anthropic use offline monitoring to detect misaligned actions by internal AI agents, but current public reporting on the effectiveness of these systems is insuffi…