ResearchAnthropic Fellows: Model Spec Midtraining, a method aiming for better generalizationAnthropic Fellows' new research presents the Model Spec Midtraining (MSM) method, which first teaches AIs how and why they should generalize, then applies this to reduce the unreliability of conventional alignment methods in novel situations.1 min read
ResearchAISI finds public GPT-5.5 matches Mythos Preview on cybersecurity CTFsThe UK Artificial Intelligence Safety Institute (AISI) tested OpenAI's publicly available GPT-5.5 against Anthropic's restricted Mythos Preview on 95 Capture the Flag tasks and found near-identical performance.3 min read
ResearchAI-designed full bacteriophage genomes produce viable viruses in laboratory testsResearchers from Stanford University and the Arc Institute published in Nature that an AI system called Evo designed complete bacteriophage genomes.2 min read
ResearchResearchers train talkie-1930-13b on pre-1931 English; it generates modern Python after few examplesAlec Radford and collaborators trained a model called talkie-1930-13b using only English text published before January 1, 1931.2 min read
ResearchHarvard study: an OpenAI model outperformed physicians in some emergency diagnosesA Harvard University study compared OpenAI models with human physicians in emergency department diagnostic tasks and found that one AI model produced accurate or near-accurate diagnoses in 67% of cases versus 50–55% for two internists.2 min read
ResearchAI identifies venom-derived peptides that kill drug‑resistant bacteriaResearchers at the University of Pennsylvania used a deep‑learning model to mine animal venoms for short peptides with antibacterial activity.3 min read
ResearchLLMs Use Different, Often More Structured Strategies Than Humans in Rock–Paper–ScissorsResearchers compared strategic patterns of several large language models and humans by analyzing their play in repeated rock–paper–scissors matches.4 min read
ResearchOptical AI image generation that largely eliminates electronic computationResearchers at the Los Angeles-based California University have developed an optical, AI-driven image generator that uses patterned laser light and passive optical layers to produce images with almost no electronic computation beyond initial encoding.3 min read
ResearchAutomated LLMs Can Produce Full Scientific Papers, Study ShowsTwo US-based economists demonstrate that large language models can generate complete scientific papers with minimal human intervention.2 min read
ResearchStudy finds about one-third of new websites by 2025 may be AI-generatedResearchers from Stanford University, Imperial College London and the Internet Archive analyzed Internet Archive data and estimate that roughly 35% of webpages published by mid-2025 are either AI-generated or AI-assisted.2 min read
ResearchAI-guided electrode suit that teaches movements in real timeResearchers at the University of Chicago developed a wearable system that uses AI and electrical muscle stimulation (EMS) to physically guide a user’s muscles in real time, adapting to the current scene rather than replaying a fixed movement.3 min read
ResearchAI weather models struggle with rare, climate-driven extremesAI-based weather models match or outperform traditional physics-based systems for typical forecasts and are cheaper and faster, but they struggle to predict rare, unprecedented extreme events that are becoming more frequent with climate change.5 min read