OpenAI announced AI-generated results in harmonic analysis, computational complexity and quantum field theory, alongside a new process for external researchers to examine and explain them. Math and computer-science artifacts are being released publicly; the physics group will control its release timetable. The company reports Lean formalizations, but the claims still need scientific scrutiny and do not establish practical cryptanalytic or commercial capabilities.
OpenAI released aggregated data linking consumer ChatGPT work-related use to 165 activity categories across nearly 100 countries, covering quarterly periods from early 2025 through the third quarter of 2026. Enterprise conversations and Codex use are excluded. The data describes sampled use, not productivity gains, job displacement or the whole AI market.
METTA combines within-episode neural memory with compact summaries across episodes, according to its preprint abstract. The authors report adaptation results on held-out reinforcement-learning tasks. This is research on meta-RL agents, rather than evidence that a general-purpose language-model assistant has acquired reliable long-term memory.
The authors propose choosing which partially generated solutions to continue by directly optimizing pass-at-k. Their abstract reports tests across open-weight reasoning models. The method tries to retain promising but varied reasoning paths, rather than optimizing quality and diversity separately; these are preprint claims, not independently reproduced results.
The authors compare trained classifiers and prompted language models under explicit costs for correct routing, human triage and mistakes. Their abstract reports that abstention policies can reduce routing cost, while better calibration alone need not improve a model's ability to distinguish correct answers. The findings are limited to the tested dataset, prompts and illustrative costs.
A Stanford study reported by Ars Technica says entry-level and young-worker employment has fallen more sharply in occupations judged to be more exposed to AI than in more AI-resistant fields. The article says young employment in AI-impacted fields was down 19%, adding empirical evidence to the debate over whether generative AI is already reshaping early-career labor demand.
British AI startup Inherent, founded by DeepMind alumni, has introduced Faraday, an AI agent designed to replicate scientific papers. In coverage by TechCrunch, the company said Faraday outperformed systems from Anthropic and OpenAI on its research-replication task, positioning the product as an AI "teammate" for scientific work.
TechCrunch reports on Nvidia research suggesting that AI agents can become more capable and stable through improvements to the surrounding harness, including fine-tuning and task orchestration, even when the underlying model is not best-in-class for the task. The piece frames the result as evidence that agent system design may matter as much as, or more than, incremental gains in the core model.
Nvidia researchers have introduced a cross-model KV-cache transfer method that maps a prefetched cache from one model into another, aiming to avoid recomputing the full conversation when agent systems switch models mid-session. According to the report, the linear mapping technique delivered 2.7x to 25x faster handoffs on compatible model pairs while preserving up to 98% of the target model’s standalone accuracy.
Google is participating in Operation Blue Skies, a project involving the UK government and airline partners that will use AI-informed analysis to slightly reroute flights in an effort to avoid forming contrails. The tests focus on whether targeted path changes can reduce aviation warming effects without major operational disruption.
Researchers at MIT and Harvard describe a failure mode they call role drift, in which modules inside compound AI systems deviate from their intended functions while overall end-to-end accuracy still improves. They introduce a method called Role Anchor that aims to keep modules constrained to their assigned roles during training, including in retrieval-augmented generation pipelines where a reader might otherwise answer from internal memory instead of retrieved evidence.
Wired reports on new research arguing that AI adoption in fossil-fuel operations could increase carbon emissions by making extraction and production more efficient. The study says the resulting emissions increase could substantially exceed the direct climate impact associated with AI data-center power use.
Wired reports on research claiming a new technique can surface internal "reasoning traces" from models including Claude, GPT, and Gemini. The reported findings suggest these traces may reveal how models arrive at answers and could provide evidence about whether some Chinese systems were trained on outputs from leading U.S. models.
Wired reports on new research using AI to help untangle the complex genetic basis of schizophrenia. The work is described as providing one of the clearest pictures yet of how many genetic factors may contribute to the disorder, potentially opening new directions for psychiatric and biomedical research.
Ars Technica reports that Google DeepMind's open-source WeatherNext weather model delivered accurate hurricane forecasts while relying on lower-resolution input data than many conventional systems. Coverage of the work says the model gave forecasters roughly an extra day of warning in some cases, surprising weather scientists with its performance.
Researchers from Coral AI Labs and several universities presented AgentRadio, an asynchronous message-passing layer that lets multiple coding agents communicate during execution without pausing their primary work. In tests on the SWE-Atlas QnA benchmark, a four-agent setup using AgentRadio improved long-horizon codebase question accuracy over single-agent baselines and outperformed a single Claude Code agent running Claude Opus 4.8. The team also said the ideas are being adapted into a commercial product called Coral Code.
At VentureBeat's Transform 2026 event, Stanford professor James Zou described a multi-agent research system that scales from small 'virtual lab' teams to a 37,000-agent 'virtual biotech' organized into divisions such as target discovery, molecule design, and clinical trials. Zou said the system uses an AI-native data layer called Paperclip to let agents work across scientific literature and databases more effectively, and reported internal results including nanobody design and an antibody-drug conjugate concept targeting CD276 for lung cancer. He also said Merck later independently developed and validated the same therapeutic design, which he presented as external confirmation of the system's output.
