OpenAI has published what it says is a proof that solutions to the three-dimensional Navier–Stokes equations can develop a singularity in finite time. In ordinary language, the equations describe how fluids move, and a singularity is a point where the mathematical description stops remaining smooth. Whether such a breakdown can happen is part of one of the Clay Mathematics Institute’s Millennium Prize Problems.
The company released both a paper and a formalization in Lean, software used to check whether each step follows from stated assumptions. OpenAI says roughly 10,000 agents worked in parallel for about 88 hours on the Navier–Stokes search, followed by about 17 hours of formalization.
That combination is unusual: the system did not only produce prose that resembles a proof, but also translated the argument into a form a proof assistant can inspect. It is still not a settled result. Independent mathematicians must examine the definitions, assumptions, and argument, and the Clay Mathematics Institute has not accepted it as a solution to the prize problem.
If the proof survives that scrutiny, the achievement would be both mathematical and methodological. It would show that a large system of AI agents can explore possible arguments, assemble a candidate proof, and connect it to formal verification. For now, the publication is a substantial claim with unusually concrete evidence, not a completed verdict from the mathematics community.
Meta launches Muse, a personal agent that can act online
Meta began rolling out Muse in the United States on September 8. The personal AI agent runs a browser in a dedicated virtual machine and can work on longer tasks involving research, email, reservations, and purchases.
Meta says Muse asks for approval before sending an email or completing a purchase. A separate system called Sentinel checks its internet actions, and the company says Muse conversations and virtual-machine data are not shared with its advertising systems. Full end-to-end protection for the virtual machine is planned for later this year, so that part of the privacy design is not available yet.
Muse brings the question of agent safety into everyday accounts: which actions can proceed automatically, and which should always stop for confirmation? Meta has described its permission and audit controls, but their reliability will have to be judged through use rather than an announcement alone.
OpenAI opens its Agents API to all developers
OpenAI released the Agents API in public beta on September 10. It packages the execution system used by Codex into an API for agents that can use tools, delegate to subagents, and continue work over sessions that may last for days.
Developers can run an agent in an OpenAI sandbox, on their own infrastructure, or through a partner environment. OpenAI says there is no separate fee for the API itself; customers pay for the models and tools they use. Early performance figures in the announcement come from customer examples, not an independent comparison.
The release moves a difficult part of agent development—the machinery that coordinates tools, context, and long-running work—into a managed service. Developers still have to decide what the agent may access, where it runs, and how consequential actions are approved.
Skild AI shows a robot learning from one video
Skild AI demonstrated its S1 robot learning a new multistep task from a single video without retraining the underlying model. In one example described by NVIDIA on September 10, the robot watched an 11-minute potting demonstration and then attempted the job autonomously.
The company says S1 can handle tasks lasting up to 10 minutes. In its internal tests, it completed individual steps 66% of the time, compared with 9% for a baseline. These figures come from Skild AI and appear in an NVIDIA partner report, so they do not establish how reliably the method will transfer to unfamiliar homes, shops, or factories.
Learning by demonstration could make robots easier to adapt because a person can show a task instead of supplying a new program. The harder test is whether the robot can recover when objects, lighting, or the order of events differ from the example.
Teachers and Microsoft set enforceable school AI terms
The American Federation of Teachers, the United Federation of Teachers, and Microsoft announced a school AI safety and privacy standard on September 9. The terms prohibit training models on student data, prohibit tracking students, require human oversight of AI decisions, and require clear explanations in ordinary language.
US school districts can incorporate the terms into Microsoft customer agreements. Violations can then become grounds for ending a contract or seeking damages. That makes the document more concrete than a voluntary principles statement, but it is still a contract standard rather than a federal law covering every company or school.
The agreement gives families and educators specific clauses to look for when a school adopts AI. Its reach will depend on how many districts use the terms and whether other vendors accept comparable obligations.
Canada starts a national AI literacy program
The Canadian government launched a free National AI Literacy Initiative on September 9, with separate learning tracks for students, educators, and the wider public. The lessons are meant to cover practical AI use as well as its limits and risks.
The government expects the program to reach up to 1 million postsecondary learners and more than 50,000 K–12 educators. Those numbers are targets, not completed enrollments, and the announcement does not yet provide evidence of learning outcomes.
The initiative treats AI literacy as public infrastructure rather than a skill left to individual schools or employers. The next useful measures will be participation, completion, and whether learners become better at evaluating AI output—not only how many people were offered access.
AlphaGenome maps predictions for 9 billion DNA changes
Google DeepMind released the AlphaGenome Atlas on September 8 as a free research portal. It contains model predictions for the molecular effects of roughly 9 billion possible single-letter changes across the human genome.
Researchers can use the atlas to narrow down variants that may deserve closer laboratory study. It is not a medical diagnosis, and a predicted molecular effect does not establish that a person will develop a disease or respond to a treatment.
The scale makes it possible to inspect many more possible variants before running costly experiments. Its value will depend on how well those predictions hold up across real biological data and diverse populations.
GPT-Live-1 listens while it speaks
OpenAI made GPT-Live-1 available through its API on September 10. The voice system is designed for full-duplex conversation, meaning it can listen while speaking, respond to interruptions, and handle background noise rather than forcing each speaker into a strict turn.
OpenAI also says the model handles a wider range of accents and languages. The front-end voice layer costs $0.05 per minute, while the model working behind it is billed separately.
More natural timing can make voice assistants easier to use in noisy or fast-moving conversations. It does not by itself establish that the underlying answer is correct; developers still have to evaluate the model connected to the voice layer and disclose the full cost of the combined service.
DreamZero rehearses a robot’s next move in predicted video
Telexistence and AWS adapted DreamZero, a 14-billion-parameter world model, for convenience-store robot work. The model predicts future video frames together with robot actions, allowing a system to estimate what may happen before it moves.
AWS reported better results in clean simulation data after fine-tuning. Performance was less convincing on real demonstration data, where training loss rose sharply, and the authors said offline tests did not fully reveal how the robot would behave in a closed loop with the real world.
The experiment is useful precisely because it records that gap. Predicting a plausible next image is not the same as controlling a machine safely among changing shelves, products, and people.
IBM opens a time-series model that reports uncertainty
IBM Research released PatchTST-FM-r2 on September 9. The 385-million-parameter model can read a context of 8,192 time points and produce 99 quantile estimates, giving users a range of possible outcomes rather than only one forecast.
IBM describes the model as commercially friendly and published its weights and evaluation. In the company’s leakage-free zero-shot comparison, it ranked near the top among permissively licensed models. Across the broader set of pretrained models, it placed third or fourth depending on the metric.
Those rankings come from IBM’s published benchmark, not an independent contest, and they depend on the chosen datasets and metrics. The release is nevertheless useful for researchers and developers who need inspectable forecasting tools for signals such as energy demand, traffic, or sensor readings.
This article was written with assistance from Wonder Bricks AI Agent and edited by SunnyLabs.