NASA and IBM have released an open-source AI model trained on a large collection of lunar observations, alongside a dataset built from nine instruments across four missions, to help scientists analyze the Moon at scale.
A technical report by Princeton researcher Yifan Zhang proposes Recurrent Looped Transformer, which carries decoder hidden state and sliding-window attention cache from one token to the next across prompt and response. The report is a design specification and presents no measured efficiency or reasoning results.
A relationship psychologist writing for CNBC says the most useful morning question between partners is not "How are you?" but a quick gauge of how much energy each person has that day. He calls it a capacity check and says it prevents small misunderstandings from turning into conflict.
A geomorphologist found a fossilized Pacific mastodon molar in a San Mateo County preserve. Paleontologists call it the most complete mastodon molar from the area, and Stanford University will carbon-date it.
A chatbot links the full fruit fly connectome to a frozen 1.2B language model; its controls show a no-graph version slightly outperforms the fly wiring.
A French-led team squeezed water between diamond tips at up to 230 gigapascals and laser-heated it to thousands of kelvins, producing hexagonal close-packed ice that showed signs of a superionic state.
Delaware and other states use flocks of sentinel chickens to detect West Nile virus and Eastern equine encephalitis, after the CDC warned of unusually early transmission and the most early-summer cases since 2004.
OpenAI says about 10,000 agents, tens of millions of dollars in compute and 88 hours produced a solution to the Navier-Stokes Millennium Prize problem. Mathematicians tell The Verge the race with Anthropic is upending their field.
Large Language Models · Research · Products & Applications
At the 2026 Inclusion·Bund Conference in Shanghai, two Sept. 12 forums examined agent post-training and AI in education, as researchers discussed self-evolving models and educators called for clear limits on classroom AI.
GPT-6 Astra has cracked the only FrontierMath Tier 4 problem no AI had solved, leading Epoch AI to declare the research-level benchmark saturated. OpenAI reports Astra's score at 97.6 percent, and the project is now moving to genuinely open problems.
A study in Geophysical Research Letters estimates Mercury may have contracted roughly 30 percent more than earlier calculations, losing as much as 14 miles in diameter. The BepiColombo mission, due to enter orbit in November, should test the figure with its laser instrument.
Twenty-five Fields Medalists, including Terence Tao and Yu Deng, have signed an urgent joint statement warning that AI companies' benchmark-driven pursuit of mathematics is severely misaligned with the field's goals.
At the 2026 Inclusion Bund Conference, experts from AI, materials science, life sciences and drug development discussed how autonomous research systems can close the loop between computation and experiments, and what data, models and clinical validation still stand in the way.
Apple researchers have unveiled SimpleDesign, an AI model trained end-to-end to jointly generate protein sequences and 3D structures. The study reports competitive benchmark results but says the proteins have not been experimentally tested.
Twenty-five Fields Medal winners, led by Terence Tao, warned that treating famous math problems as AI benchmarks could harm mathematics, after OpenAI said an internal model solved the Navier-Stokes problem.
Researchers have used a mathematical model to study games with evolving strategies and randomly varying returns, an approach intended to capture how real-life rewards and consequences change, Ars Technica reports.
An MRSA outbreak linked to three bouncy castles at a community event in Ireland infected 48 children; 15 sought emergency care, four were hospitalized, and all recovered.
Researchers from ByteDance Seed and partners introduced HarnessDev, which evaluates the agent harness an LLM builds rather than its answers. Self-built harnesses beat references on writing and ML tasks but trailed on code and search, and only 34 of 64 changes generalized to held-out tests.
Twenty-five Fields Medal winners signed an open letter saying AI labs racing to solve famous math problems threaten mathematicians' work, as OpenAI's dispute with researchers escalates.
Researchers used mass spectrometry-based proteomics to analyze glues and adhesives in Egyptian artifacts, finding animal collagens, egg proteins, and plant proteins from sesame and moringa, according to a new Science Advances paper.