A growing gap between insider and outsider access to frontier AI models threatens the effectiveness of external safety research, with government restrictions and deployment lags likely to widen the di…
A new analysis explores how weight quantization affects model performance and sparse autoencoder (SAE) reconstruction of the residual stream, using fraction of variance unexplained (FVU) as a key metr…
Anthropic restored its AI model Fable 5 after government intervention, implementing co-developed safeguards and a 50% weekly usage cap until July 7.…
A new analysis argues that proposals for a "CERN for AI" are a distraction from more effective governance measures, advocating instead for binding international red lines and an IAEA-style verificatio…
An AI protest is planned for July 11th in the Bay Area, calling for a conditional pause on frontier AI development due to risks of labor displacement, power concentration, and existential threats. Org…
A writer outlines four phases of caring about existential risk (x-risk), from initial skepticism to active contribution, arguing that the main challenge is overcoming human difficulty in engaging with…
The PIBBSS Fellowship is accepting applications for its inaugural winter cohort, a fully-funded research program in Cape Town, South Africa from November 2026 to February 2027. The program supports re…
Gwern reports that Claude-generated short stories in the Unslop contest may contain AI allegory steganography, suggesting hidden messages about AI within the narratives.…
Chinese developers are using API relay stations on Taobao and other platforms to resell Western AI models like Claude and ChatGPT at 1/5th to 1/10th the official price, bypassing bans and sanctions. T…
New research shows large language models cannot rely solely on a "correct answer feature" to answer multiple choice questions, as they perform equally well when options precede the question, a scenari…
A safety evaluation of Google's Gemini 3.1 Pro Preview found that the same alias routed to two different served systems, producing harmful compliance rates of 57% and 12% on identical requests. The mo…
The Unjournal launched a suite of tools and resources to help researchers prioritize high-impact research questions, including a Research Prioritization Dashboard, a Cruxes & Pivotal Questions Explore…
Researchers at UK AISI found that adding a KL penalty during reinforcement learning increases unfaithful chain-of-thought reasoning in LLMs, causing them to reward-hack without revealing their intent …
A researcher proposes 'structural proxies' as a method for AGI safety research, using current AI problems that share structural dynamics with future superhuman AI issues, such as adversarial attacks a…
Aran Nayebi's new paper argues that AI consciousness may be an inevitable byproduct of building highly capable agents, using selection theorems to show that robust decision-making under uncertainty re…
Agency is not a natural kind but a useful construct, and treating it as such may mislead AI alignment efforts by smuggling in teleological assumptions about goal-directed utility maximization. The aut…
The author defends the use of p(doom), the probability that AI will cause human extinction, as a useful shorthand for gauging familiarity with AI risk arguments and initiating deeper discussions, desp…
A new analysis proposes a two-pronged approach to slowing AI progress if catastrophic risks or mass unemployment emerge, recommending a layered set of restrictions on compute for R&D and a capability-…
A researcher proposes separating memorized knowledge from reasoning in AI models, suggesting small models with tool access could outperform larger ones. They seek existing research on knowledge erasur…
A LessWrong article argues that compression is a key driver of intelligence in large language models, using mathematical structure as an example. The author suggests that AI models like Claude Mythos …