ResearchOpenAI study: workers increasingly make cross-occupation AI tasks part of regular workflowsOpenAI Economic Research analyzed more than 1.5 million work-related ChatGPT messages from April–July 2026 and found that workers often use AI to perform tasks outside their formal occupations, and some of those tasks recur and become part of regular workflows.4 min read
ResearchWhich Parts of AI Should Remain Unchangeable? Agents, Self-Improvement and Test-Time TrainingAI systems are becoming more adaptable: agents can create tools and subagents, models can update during inference, and researchers are experimenting with elements of recursive self-improvement.4 min read
ResearchLessons from an AI-First Ophthalmology Clinic: Performance, Integration, and Practical LimitsResearchers at the Beijing Eye Research Institute (BERI) built and evaluated an AI-first ophthalmology clinic called AI-TEC, using artificial intelligence across consultation, imaging and follow-up while retaining human oversight.3 min read
ResearchUsing Games as Training Grounds: Good Start Labs Studies Transferable AI SkillsGood Start Labs, spun out of Every with $3.6 million in backing, trains AI models inside games to teach verifiable skills such as strategic reasoning and tool use.5 min read
ResearchRL-compiled diffusion enables fast, set-aware query fan-out for retrievalResearchers present Retrieve-for-Train, an offline reinforcement-learning pipeline (ICML 2026) that compiles reward-aligned query fan-outs into supervision and distills them into a compact diffusion retriever.5 min read
ResearchVercel Labs Introduced, Vercel's Public Research and Experimental ArmVercel announced Vercel Labs, a public research and experimental initiative where they will share their grants, research and failed experiments.1 min read
ResearchModel specialized with wet-lab data beat GPT‑6 Astra in scientific taskA specialized model trained on wet-lab data outperformed GPT‑6 Astra on a cutting-edge scientific task.1 min read
ResearchWhen to choose dense vs. Mixture-of-Experts (MoE) model architecturesDense and Mixture-of-Experts (MoE) architectures trade off parameter usage, memory footprint, and serving complexity in different ways.5 min read
ResearchAI 'virtual cell' predicts effective drugs for triple‑negative breast cancer before treatmentNature reports an AI model trained on 38 million protein measurements at Westlake University that identifies which drugs are likely to work in triple‑negative breast cancer patients before any dose is given.2 min read
ResearchRichard Socher’s Recursive Aims to Automate AI Research with Recursive Self‑ImprovementRichard Socher — founder of You.com and AIX Ventures — has launched Recursive, a company pursuing recursive self‑improvement: AI systems that automate the process of inventing and improving AI itself.5 min read
ResearchUniversity of Surrey study finds generative AI eases holiday planningA study published in the Journal of Travel Research and reported by Phys.org examined interviews with British adults who used generative AI tools to plan real trips.2 min read
Research25 Fields Medalists Warn AI Could Undermine Mathematical UnderstandingTwenty-five Fields Medal recipients, including Terence Tao and Peter Scholze, signed an open letter arguing that rapid AI-driven production of proofs risks divorcing mathematical discovery from human understanding.4 min read
