ResearchReplaySSM: caching SSM inputs to speed hybrid-model decodingReplaySSM, developed by NVIDIA and the Dao AI Lab at Princeton, represents state-space model (SSM) state as a checkpoint plus a short buffer of recent inputs instead of updating and storing the full state every token.5 min read
ResearchStudentSim: personalized student simulators to accelerate AI tutor researchStudentSim is a research system that trains personalized AI student simulators from publicly available, de-identified learner records to give early-stage feedback during AI tutor development.4 min read
ResearchU.S. national labs link supercomputers for cross‑site AI trainingResearchers at multiple U.S.3 min read
ResearchNew 'Humanity’s Sixth Sense' benchmark reveals gaps in AI intuitive visual reasoningElorian and Scale AI released Humanity’s Sixth Sense (HSS), a benchmark testing models’ ability to infer unstated facts from images and short videos across temporal, physical, social and abstract contexts.3 min read
ResearchBenchmark shows frontier AIs improving 'research taste' rapidly, raising RSI timelineP‑Zero Research’s new benchmark, TasteVal, measures how well AI models design experiments and interpret results across eight tasks representative of frontier AI research.3 min read
ResearchEvaluating Matt's Questioning Skill and the Questioning Ability of Modern ModelsThe author praises Matt's "grilling" questioning skill, which effectively uncovers complex, still undeveloped ideas; they also note that modern language models are surprisingly good at asking…1 min read
ResearchSimulated democracies run by AI diverged from party programs and were sensitive to news framingA Hungarian team from Kreatív Kontroll created a virtual country, Árkádia, and ran 186 simulations in which five different large language models led political parties for 12 virtual years.4 min read
ResearchSocial jetlag increases preference for AI evaluators by amplifying social anxiety, study findsA study published in Computers in Human Behavior reports that mismatch between biological sleep timing and social schedules — social jetlag — makes people more likely to prefer evaluations from artificial intelligence rather than from other humans.4 min read
ResearchLancet commission identifies 17 global health threats including AI misuse and nuclear war riskA Lancet Commission of 35 experts has identified 17 major threats that could severely reduce healthy life years worldwide by 2100, including the malicious use of artificial intelligence and the consequences of nuclear conflict.3 min read
ResearchWhy AlphaFold Was a Breakthrough — Not the Final Answer on Protein FoldingIn a panel moderated by Brandon Anderson, Pushmeet Kohli (Google DeepMind) and Sal Candido (Biohub) argue that AlphaFold’s success marked a major advance but did not close the book on protein biology.4 min read
ResearchFrontier AIs still fall short of fully automating AI research, early tests showAn evaluation called InnovationEval tested whether recent frontier models can independently rediscover a novel machine-learning training technique (Self-Distillation Policy Optimization, SDPO).4 min read
ResearchHospital trial in Boston found Google medical chatbot safe and usefulBeth Israel Deaconess Medical Center conducted an eight‑month clinical trial of Google’s AMIE chatbot, which pre-screened patients before visits and produced summaries for physicians.3 min read
