ResearchAI-driven formalization of mathematics advances research automation, says ResolutionAI startup Resolution (formerly Sequent) reports that it has automatically formalized Hironaka’s resolution of singularities in the Lean proof assistant, producing a large formal library and accelerating follow-on work in statistical learning theory.3 min read
ResearchAI agent team designs and assesses antiferromagnetic semiconductor candidatesResearchers using a fleet of AI agents, described in a Vals blog post, designed a new antiferromagnetic oxide candidate and rediscovered a 1999 material predicted to have desirable properties.2 min read
ResearchClaude Opus 5.5 Achieves 18.5% on ProgramBench, Marking Rapid Gains in Program ReconstructionAnthropic's Claude Opus 5.5 (xhigh) reached 18.5% fully solved and 70.5% nearly-solved on ProgramBench, a benchmark that evaluates whether models can rebuild software from compiled binaries and documentation.2 min read
ResearchOpenAI publishes country-level data on ChatGPT use for 165 work activitiesOpenAI Economic Research released quarter-level data linking ChatGPT messages to 165 Intermediate Work Activities (IWAs) across nearly 100 countries, covering Q1 2025–Q3 2026.3 min read
ResearchThree AI-Generated Results in Harmonic Analysis, Complexity Theory and Quantum Field Theory, Released with Expert-Led CollaborationsOpenAI reports three AI-produced mathematical and scientific results — a general proof of the local smoothing conjecture (and implied Kakeya and Stein restriction conjectures), a disproof of the Exponential Time Hypothesis (ETH), and existence proofs for certain four-dimensional interacting quantum field theories including Yang–Mills and QCD.6 min read
ResearchWhy large language models are suddenly succeeding at hard mathThis article examines why recent large language models (LLMs) have begun producing solutions to longstanding open mathematical problems.7 min read
ResearchLarge-scale formal software verification by Frontier Red Team using ClaudeThe Frontier Red Team (Anthropic) used the Claude language model in collaboration with Theorem Labs and, within a few weeks, proved security properties in more than a hundred real, open-source libraries.1 min read
ResearchAI agents left to explore invented games and learned useful skillsMIT researchers left 13 AI agents called Eko, each running on Claude Opus 4.7, to roam a simulated island for 30 hours without explicit goals.3 min read
ResearchGlobal 10‑m Forest Typology Map (ForTy 2020) and 2017–2025 Time Series ReleasedGoogle DeepMind and Google Research, in partnership with the World Resources Institute (WRI) and the International Institute for Applied Systems Analysis (IIASA), published ForTy 2020: a 10‑meter resolution global map that classifies each 10×10 m pixel into six forest and land‑use types, plus an early annual time series for 2017–2025 showing how those classes changed.4 min read
ResearchGetting more from today's AI: memory, harnesses and reusing layersResearchers and engineers are exploring ways to extend the usefulness of existing AI models without immediately switching to larger architectures.4 min read
ResearchMicrosoft's DxO AI panel more than doubles accuracy vs generalists on simulated complex casesMicrosoft reported that its MAI-DxO diagnostic system—an orchestration of multiple AI agents—correctly diagnosed simulated complex cases from the New England Journal of Medicine 83% of the time versus 36% for individual primary care doctors and general internists.4 min read
ResearchAI-augmented blind-sweep ultrasound matches expert gestational age estimates across diverse clinical settingsA prospective multi-center study published in JAMA Network Open found that an AI system estimating gestational age from standardized “blind sweep” ultrasounds achieved accuracy non-inferior to expert sonographers.5 min read
