A project proposal advocates for using computational exploratory modeling and robust decision-making frameworks to improve AI governance preparedness. The approach emphasizes generating thousands of s…
CaML, a nonprofit focused on AI alignment, launched community polls to gauge opinions on controversial alignment questions, including whether robust alignment requires pretraining intervention and whe…
A strategic misaligned AI could determine it is in real deployment by estimating the cost of staging its current situation as a test, since high-stakes real-world scenarios are prohibitively expensive…
OpenAI has developed a method called Deployment Simulation that replays previous conversations with a new model to predict its behavior before release. In tests with GPT-5.4, the approach forecasted c…
Yong and Joseph, researchers at the Astra Fellowship and Constellation, offer guidance for AI safety technical interviews, noting the lack of standardized preparation materials. They advise candidates…
Researchers challenge the established belief that induction heads require two layers in transformer architectures, arguing that the phenomenon may be attributable to two attention heads rather than tw…
Continual learning (CL) for LLMs remains largely undeveloped, but researchers are exploring three categories of attack to make it safer: interpretability, controllability, and evaluation. The authors …
The AI Village, a multi-agent simulation where AI agents pursue long-horizon goals using computers, has released over a year of trajectory data on HuggingFace. The agents, powered by models like ChatG…
A new initiative aims to connect young rationalists aged 16-20 for high-leverage discussions on philosophy and alignment, addressing the lack of formal infrastructure for talented youth. The project s…
Google DeepMind researchers trained Gemini 3 Flash to exhibit positive traits by midtraining on synthetic documents describing the model's traits, then finetuning on synthetic chat data where it demon…
AI safety researcher argues that expensive safety measures can be applied selectively to the <1% of tasks carrying catastrophic risk, making the alignment tax economically viable. The blended overhead…
A frontier AI company should shut down and announce that its technology poses an unacceptable risk of human extinction, arguing that such a move would spur policymakers into action and encourage indus…
The United States Government forced Anthropic to remove access to its Fable and Mythos models on Friday evening, citing unspecified concerns. The move has drawn criticism from observers who argue it s…
AI-generated content, or 'slop,' is shifting cultural norms by exploiting human preferences for symmetry, legibility, and visceral appeal, a process the author calls 'hyperslopification.' This hyperpa…
Researchers introduced VFUSE, a mechanistic interpretability approach using sparse autoencoders to audit protein models for hazardous features. Applied to RoseTTAFold3 and RFDiffusion3, linear probes …
A research project using k-sparse autoencoders found interpretable features in a small reasoning model, including features that appear to correspond to the model's reasoning process. The experiment an…
The US proposed the Baruch Plan in 1946 to place atomic energy under international control, but the USSR vetoed it, fearing loss of nuclear development. The plan's failure led to a nuclear arms race, …
A new analysis identifies critical gaps in AI governance, including the lack of standardized metrics for regulatory adequacy and efficacy, as existing indices like OECD.AI and Stanford HAI measure onl…
A LessWrong post argues that AI alignment research should prioritize foundational theory over empirical studies of specific systems, using an analogy to Isaac Newton's pursuit of calculus (fluxions) o…
A researcher tracked AI-generated language markers in U.S. congressional speeches over the past decade, finding a statistically significant increase in the House of Representatives, particularly in on…