Agents just proved a 13-year-old conjecture — and science AI is getting both stronger and more honest

The agents that did the science

Google Research has released /teamwork, a multi-agent system for human-AI collaboration. The system uses parallel exploration paired with dedicated falsifiers to ensure reliability, and it partitions context by construction to prevent attention dispersion. Researchers used /teamwork to prove the 2013 Courtade-Kumar conjecture in information theory, and the /teamwork-LEAP framework produced a machine-checked formalization of the Weil conjectures. A two-week Argon campaign rewrote the Zircon microkernel in Rust, passing all unit tests, and DuoDiscover achieved a 4.09 times geometric-mean speedup over reference implementations across 17 JAXBench kernels. Borg's agentic scheduling optimization resulted in a 24 percent reduction in TPU workload preemption. m:…9331

The system also reviews science, not just does it. Deep Review, a document review pattern within /teamwork, provided feedback on 578 submissions for STOC 2027, with 90 percent of surveyed authors rating the feedback as factual. Google is offering 10,000 academic researchers six months of free access to the highest usage tier of Antigravity tools. The offer is the distribution strategy: get the framework into the hands of researchers who have open problems, and let the results advertise themselves. The conjecture proof and the microkernel rewrite are the evidence offered so far, and both are Google's own account of its system, so they should be read as the company's claims about its tool, not independent verification. m:…9331

A model that reads brain scans

A preprint this week describes Neuro-VL, a vision-language model trained on nearly one million clinical neuroimaging studies from 450,000 patients, comprising 10 million volumetric scans. It employs HLIP as its visual encoder, which uses a hierarchical attention mechanism and was pretrained using radiology reports. The authors report that Neuro-VL outperforms existing general-purpose models, and existing specialized medical models, in diagnosis, localization and reporting, and that it compared serial examinations to assess disease progression. It generates radiological interpretations with auditable reasoning traces, and resisted misleading clinical assumptions more reliably than OpenAI models and more reliably than Google models, according to the authors. As a preprint, none of this is peer-reviewed, and the comparison is the authors' own, so the claim to watch is not the win but the audit trail: a model that shows its reasoning can be checked, which is what clinical use requires. m:…2989

Weather, storms, and the physics check

Two more preprints push AI simulation in opposite directions: faster, and more honest. STRATA introduces a transformer-based architecture for global storm-resolving atmospheric simulation, achieving 48 simulated days per MWh, a 50 times energy efficiency gain over the SCREAM physics model, and 741 simulated days per wall-clock day on 512 H100 GPUs. Kilometre-scale emulation requires about 10 times more FLOPs per grid point than coarse-resolution AI weather models, according to the authors, and large-scale biases develop in STRATA rollouts with lead time. Speed is not the same as accuracy over long horizons, and the authors say so themselves. m:…6585

PhysXAI is the honesty check. It introduces a framework to diagnose physical hallucinations in weather foundation models, outputs that violate fundamental laws like mass conservation. It decouples forecast error from physical plausibility, so a forecast can be wrong in the ordinary way without also being physically impossible, and supports detection and attribution across GraphCast, Pangu-Weather and NeuralGCM. Its latent representation is approximately isometric to the physical residual space, which is the authors' way of saying the model's internal geometry tracks physical error. TurtLES, a 14 TB dataset of 3D turbulent cylinder wakes in the shear-layer transition regime, provides the test bed for the same question in fluid dynamics: 400 trajectories of 400 time steps each, with variations in geometry, Reynolds number and angle of attack. Neural operators tested on it exhibit failure modes including temporal instability and loss of small-scale energy in autoregressive rollouts. The field is learning to measure what its models get wrong, which is the precondition for trusting them with anything that matters. m:…6583 m:…5485