ResearchPISA: global decline in reading skills tied to AI and social media use, OECD findsThe OECD's latest PISA assessment, covering 91 countries and 760,000 students, shows a significant drop in reading performance compared with previous cycles.2 min read
ResearchEditable benchmark update reshuffles GPT-6 Astra ranking on Artificial AnalysisArtificial Analysis, a widely cited AI benchmark, initially scored GPT-6 Astra 61 on September 3, placing it below Muse Spark 1.3.2 min read
ResearchAI Benchmarks: Event by Benchmark and Vercel in San FranciscoThe "AI Benchmarks" event jointly organized by Benchmark and Vercel in San Francisco focused on measuring and comparing artificial intelligence models; the event aimed to provide guidance on models'…1 min read
ResearchPISA: moderate, targeted AI use can help, very rare or very frequent use linked to poorer resultsThe OECD’s 2025 PISA assessment of more than 760,000 15‑year‑old students across 91 countries found that students who reported using AI tools for learning generally scored lower in science after controlling for socio‑economic status.2 min read
ResearchELTE study finds learning support and perceived reliability drive students' use of ChatGPTA study from the Faculty of Informatics at Eötvös Loránd University (ELTE) analysed survey responses from 166 students to identify which factors influence continued use of ChatGPT for learning.2 min read
ResearchAI agents stream text in under 1% of the time; the rest is reasoning and tool executionAccording to the author, the Codex and Claude systems hide traces of reasoning, and data from OpenAI was likely not provided for the demonstration.1 min read
ResearchAnthropic outlines prompting techniques to improve Claude’s long-context recallAnthropic published a quantitative case study showing two prompting methods—extracting relevant reference quotes into a scratchpad and supplying contextual few‑shot examples—can improve Claude’s ability to recall information from very long contexts.5 min read
ResearchHow Claude’s Constitution Makes AI Values Explicit and Alters TrainingAnthropic describes 'Constitutional AI', the approach used to train its assistant Claude, which replaces some human oversight with AI-generated feedback based on an explicit set of principles.5 min read
ResearchAnthropic's model of AI's economic impacts through 2030The Anthropic Economics team presented a new model of how artificial intelligence could affect economic growth, jobs, and wages through 2030.1 min read
ResearchStatement on Access and Divergences Regarding the Work of Levent Alpöge and Tristan BuckmasterThe public disclosure by Levent Alpöge and Tristan Buckmaster is welcomed; the researchers and the agents claim they only saw the work after its public appearance and did not have access to specific user data.1 min read
ResearchAI agents run routine superconducting‑qubit calibrations at MIT labA graduate student in MIT’s Engineering Quantum Systems Group connected GPT‑5.6 Sol, via Codex, to laboratory control software to perform routine calibration measurements on superconducting qubit chips.3 min read
ResearchAstra artificial intelligence controlled robot painted the Golden Gate BridgeAn artificial intelligence robot named Astra was given a paintbrush and a camera and then painted the Golden Gate Bridge realistically; through multiple attempts it learned to control the robot and the results steadily improved.1 min read
