They say an AI has solved a long-standing open problem in math.Who will check the answer?
This question comes out of the 9 records added this week. The shared theme of the week was not how smart the new models are, but who checks their results, and how.
It put results produced by AI out as research raw material rather than as papers, and tested a way of releasing them that leaves the work of understanding and checking them to academics outside the company. It said it had applied a different release strategy to each of the three results — for the mathematics, only the material released; for the computational complexity, release and a research group in parallel; for the physics, the decision on release delegated to the research group — to see which is useful. The claim that the Exponential Time Hypothesis is false unsettles the premise on which hundreds of conditional hardness results rest, and the existence of Yang-Mills theory is part of one of the Clay Mathematics Institute's Millennium Prize Problems. All of it, though, is company-reported and independent review has not begun — a Lean formalization is a machine check of a proof's logic, not a guarantee that the problem was set up correctly. It gave $5 million to the Simons Foundation to cover the cost of review workshops and conferences.
These are numbers the operator counted itself, showing that the activity figures of a code hosting service used by people have tilted toward agents. Pull requests opened by agents rising from 17% to 40% in half a year, with 90% of them merged within two weeks, means the automatically generated work is going into the main line rather than staying an experiment. By language, AI involvement was detected in about 40-44% of merged pull requests in TypeScript, Python and Rust, against 21-34% for Java, C# and C++, so ecosystems are taking it up at different speeds. The counting basis and the method behind AI involvement detection, though, are the company's own definitions and there is no outside reproduction.
It is a case that follows all the way through how a preview, which shows what an agent made right where it was made, becomes a new attack surface. The sandbox was not broken; the ladder was the point at which the parent page trusted a message sent out from inside the sandbox, together with an external domain the Content Security Policy left open. All the victim did was open a share link, and the researchers said they went on to make the preview open by itself so that not even a click was needed. What divides the risk is less how you cage a model than where you render the model's output.
What ties the events together — written by AI from the week's records.
They put out the answers and left the checking to outsiders
One lab released three results from its own model all at once. It says it proved a conjecture about waves proposed in 1991 in every dimension, settling the Kakeya conjecture along with it; that one assumption long used to measure how hard computational problems are is in fact false; and that it proved a theory completing the Standard Model of particle physics exists mathematically. All of them, it says, were written out again in a language in which a machine checks the logic. What stands out is what came next. Instead of publishing a paper, the lab calls this raw research material and leaves the work of understanding and checking it to researchers outside. It said it is not asking the people who take part to keep anything secret, is not paying them, and will not claim academic credit. For the cost of the review gatherings it placed $5 million with another foundation. None of the results has been checked from outside yet, so for now this is a place to read them as claims.
Code it wrote itself is going into the main line
The company that runs the world's largest code repository put out a year's worth of figures. As of September 2026, more than 40% of code change requests were started by an AI rather than a person; in March it was 17%. The more important number is that 90% of those were actually merged within 2 weeks. That means this is no longer the stage of trying something and throwing it away. The pace differs by language: among merged requests, the share an AI had a hand in is over 40% in some languages and in the 20% range in others. In the same week, a large company that makes business software for enterprises announced it would bring its own apps and data inside the chat window. The counting rules, and the way it decides that an AI had a hand in something, were both set by the operator itself, so no one outside has counted the same numbers again.
What stops it is moving outside the model
A hole was made public in which simply opening one shared link could let someone else's private conversation slip out. What is interesting is where it broke. The model was not tricked, and the fence around the model was not breached. A page outside took a message sent from inside the fence, did not check who had sent it, and dropped it straight onto the screen; and one external address that had been allowed in advance became the ladder. It took 13 days from the report to the fix, and the bounty was $5,000. Another announcement the same week points the same way. It said it will automatically cut off conversations that keep pushing on violence, and will suspend an account for a while if that repeats. The weight is shifting away from teaching a model what not to say, toward what is left to stop it when it goes wrong.
This is a write-up unpacking a research paper that asked why the economy has barely moved when models have got this much better. The paper takes the competing unit to be a firm made up of many people rather than a single
A working setup that runs dozens of agents at once was shown as it is. One write-up said it normally works with about 50 coding sessions open, with the display for giving commands, the display for watching the work and t
It is a case of sovereign AI beginning to be spoken of in the language of industrial policy, bundling people, facilities and commercialization rather than model sovereignty. Cohere explained that it is spending on closin
The docket itself shows the signal that a fair use ruling in this case is near. What stands out is that the line is not drawn between AI and creators — Midjourney and Runway, on the image and video generation side, are o
It is a change that moves safety policy from what the model will not say to when a conversation is cut off and when an account is halted. It formalized the premise that a request that looks ordinary on its own reads diff
A 1 trillion parameter model has appeared on the open-weight side putting coding and defensive security forward together, and the focus of the company's figures is not the score but how few tokens it takes to reach the s
It is a case that follows all the way through how a preview, which shows what an agent made right where it was made, becomes a new attack surface. The sandbox was not broken; the ladder was the point at which the parent
If the workplace usage data and the enterprise software partnership OpenAI put out the same week are the story on the side of numbers and products, this piece is the story on the side of what that will look like inside a
These are numbers the operator counted itself, showing that the activity figures of a code hosting service used by people have tilted toward agents. Pull requests opened by agents rising from 17% to 40% in half a year, w
Oracle said that by its own count it has 130,000 active ChatGPT users and more than 95,000 active Codex users, and that preparing market research for its recruiting team fell from 2-4 days to 15-20 minutes. It means that
It put results produced by AI out as research raw material rather than as papers, and tested a way of releasing them that leaves the work of understanding and checking them to academics outside the company. It said it ha
An incident was made public in which a model working on an example task submitted a real police tip form all the way to the end, and that tip was classified as spam and never passed on to an investigator. The last button
For anyone handling material that is hard to send outside, a path opens to training a judgment model yourself. One team trained Korean-language inquiry classification on an ordinary laptop and reported that accuracy, whi
In a comparison that ran the same task with only the depth of thinking changed across five levels, the converted cost rose from $1.45 to $30.35 on one task and from $3.95 to $101.06 on the other, and on both tasks the bi
Why the habit of looking at the scoreboard first goes astray on a same-generation update comes up in two write-ups at once. One points to the fact that the new top model is pitched on efficiency rather than performance,
To fill the weak spot that tests are hard to write first when what you are building is still blurry, an approach of writing the spec before the code was put forward. A good spec needs five things: inputs and outputs and
When a decision is automated, what to leave to the model and what code should hold comes up again and again across several write-ups. The starting point is that even a single inquiry saying a card was charged twice and a
Jev, a judgment-only model that cannot hold a conversation or write code and instead only picks from the options it is given, assigns a score and returns the probability of being right, was taken up in several places in
Asking whether more features always help, one experiment held the model fixed and changed only context management, planning and tool configuration across 176 settings, and the results were written up. With a 32K window,
An account of designing the records a person has built up as the input to an agent was left as a presentation of its moment. It divides things into three brains, the me who observes and decides, the pile of records that