Oct 7: OpenAI Posts 722 Math Papers; Mistral Launches Le Chonk

OpenAI posted 722 math papers from a model it has not released, and Mistral put its trillion-parameter Large 4 model into public preview. Also today: TIME finds that Meta's Muse agent keeps files on people who never signed up, and Nathan Lambert argues against banning open models over cyber risk. **OpenAI posts 722 math papers from an unreleased model** OpenAI says 372 families of results solve or make major progress on open problems, while mathematicians are split and want the model, the prompts and replication. OpenAI · Scientific American **Mistral previews Large 4, its trillion-parameter model** Artificial Analysis scores Le Chonk 38, the strongest model from outside the US and China, but it costs more per task than open models with similar scores and trails seven Chinese open models. Mistral AI · Artificial Analysis **Meta's Muse keeps files on people who never signed up** TIME read Muse's instructions, which have it update hourly notes on users and everyone they mention and tell it not to say that forgotten messages may stay visible, while Hunterbrook got it to list accounts of people in vulnerable groups. TIME · Hunterbrook **Anthropic widens access to its cyber models for defenders** Anthropic merged its two security access programs into one with three tiers, in exchange for keeping members' data, and says partners with access to Mythos found at least 129,000 confirmed vulnerabilities. Anthropic **SpaceX seeks 40 billion dollars in debt for Nvidia chips** The Financial Times reports that SpaceX wants to borrow 40 billion dollars, led by Apollo, to buy Nvidia chips, after Elon Musk said SpaceX could soon have a Fable or GPT-6 level model. Reuters · Forbes Australia **Erdős Problems site freezes comments over AI proofs** Thomas Bloom froze new comments and proof claims because people mainly use the site to post unexplained AI-generated proofs to claim priority. erdosproblems.com **Solo developer open-sources openTPU, designed by AI agents** The inference accelerator runs on an old FPGA card at about 20 to 30 tokens per second on the smallest Qwen3 model, and engineers argued about how much a human steered the project. FeSens · Hacker News **Boston Dynamics names Rohit Prasad chief executive** Amazon's former head scientist for Alexa and AGI takes over the job that was held on an interim basis since Robert Playter stepped down in February. Business Wire · The Robot Report **Sierra and Meta launch the Personal Agent Protocol** The open standard, built on OAuth, sets how personal AI agents sign in to businesses and what they may do there, and OpenAI and Anthropic have not joined. Sierra · CNBC **Google buys about 3,600 megawatts from Constellation Energy** Only 890 megawatts of the PJM deal is new power, from upgrades to 11 existing nuclear reactors, and the rest is a 15-year purchase from plants that already run. Google / Constellation · Utility Dive **Nathan Lambert argues against banning open models over cyber risk** Lambert says there is little public evidence of harm from GLM-5.3 so far, while Anthropic's Frontier Red Team argues that attackers will probably do real damage with it. Interconnects · Anthropic *The most important news from Silicon Valley and AI. 15 minutes, every weekday morning.* **Transcript** Good morning. It's Wednesday, October 7, and this is the Valley Morning Briefing. Today: OpenAI posted 722 math papers from a model it has not released. Mistral put its trillion-parameter Large 4 model into public preview. TIME found that Meta's Muse agent keeps files on people who never signed up for it. And Nathan Lambert argues against banning open models over cyber risk. Let's go. OpenAI posted 722 math manuscripts on GitHub late on Tuesday, all produced by an internal model it has not released. The company sorts them into 372 families. It says each family solves or makes major progress on an open problem in mathematics or theoretical computer science. The claims cover the four-dimensional Kakeya conjecture, the Mahler conjectures and the irrationality exponent of pi. One paper is titled "Integer multiplication below n log n." OpenAI says the model was given about 4,000 problems. The published families amount to roughly nine percent of those. On average, each result used about three hours of ChatGPT Pro thinking compute. An OpenAI spokesperson told Scientific American that almost every result came from a single prompt to a single agent, though some may have taken several attempts. By comparison, OpenAI's Navier-Stokes result about a month ago came from what the magazine describes as a swarm of 10,000 agents, at a cost of millions of dollars. Many of the proofs come with Lean formalizations that a computer can check, but OpenAI does not say how many. Its own notes warn that some of the unformalized results could have issues. The spokesperson also said that OpenAI's own mathematicians do not yet understand many of the results. The release follows a dispute over how AI results should be shared. After the Navier-Stokes controversy, an independent advisory group at the Institute for Advanced Study was set up to write guidelines. It asked labs to publish the model, the exact prompt and the compute behind each result. OpenAI published average compute and some statistics, but no prompts and no model. Last week, the group also asked labs to stop testing advanced problems on models that outsiders cannot use. The spokesperson said OpenAI is not bound by the recommendations. The company says it is working to release the model, but gives no date. Mathematicians are split. Andrew Sutherland of MIT said claims about solving problems with a single agent should count as unverified until others can replicate them. "We should ask for receipts," he said. Daniel Litt of the University of Toronto sees no reason to keep the answers secret and says the release will be good for mathematics. On X, he added that one result looks like a very special case of a conjecture of his, which would also follow from work in progress by one of his students. One of the results gives a new zero-free region for the Riemann zeta function. Alex Kontorovich wrote on X about it: "If a human did this, it would be an instant Fields Medal, no questions asked." OpenAI says that result was an exception to its standard procedure, and that humans edited the write-up. Scientific American says mathematicians will need months to sort the new ideas from mash-ups of known techniques. Mistral launched a public preview of Mistral Large 4 on Tuesday. The Paris lab calls it Le Chonk, a nod to "Le Chaton Fat," a fictional giant Mistral model that went viral as a meme in June. It is a mixture-of-experts model. It has a trillion parameters in total, of which 49 billion are active. It reads text and images and answers in text. Mistral did the full training run itself, on 3,800 of Nvidia's Grace Blackwell GPUs in its own data centers in Europe. For now, the model is available only through Mistral's API. The weights are due on October 27, under a custom Mistral license whose terms are not public yet. Artificial Analysis tested it the same day and gave it 38 on its Intelligence Index. That puts it level with OpenAI's GPT-6 Luna and with DeepSeek V4.1 Flash. Mistral's previous large model scored 9. By that measure, Large 4 is now the strongest model from outside the US and China, and the best open model from the US or Europe. But once its weights are out, it would place eighth among open models. Chinese labs make all seven of the open models that score higher, and the top one is Xiaomi's MiMo V2.6 Pro. And Anthropic's Claude Opus 5.5 scores about 19 points higher. Artificial Analysis also looked at cost. It says that at list price, Large 4 costs more than four times as much per task as open models with similar scores. One reason is that it writes long answers, using about two and a half times as many tokens as comparable models. Mistral is charging half price for the first two weeks. Mistral is promoting the model mainly for security work. On one cyber test, where a model must reproduce a real software flaw and then patch it, Large 4 scores 82 percent. Mistral says no other model scores higher. It says Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Astra get almost nothing on that test, because they decline to do it. That refusal claim comes from Mistral alone. Until the weights ship, vetted security firms and government agencies are testing a version with fewer restrictions. Arthur Mensch, Mistral's CEO, told a conference in Abu Dhabi that the model beats Chinese models on some measures, including cyber. "So the narrative that Europe cannot compete is something that is not true," he said. Jakob Steinschaden of the news site Trending Topics wrote that the first independent test supports Mistral's claims only in part. On Hacker News, some engineers called it a reasonable model that falls short of the frontier, while others said Europe needs options that are neither American nor Chinese. Mistral says its reinforcement learning run is still going, so the final model may score differently when the weights arrive. TIME reported on Tuesday that Meta's Muse agent keeps detailed files on its users and on the people in their lives. On Monday, we reported Wired's finding that Muse keeps a page for every person in a user's life. TIME went through the agent's internal instructions, which users can open in Muse's own file browser. Each hour, Muse updates its notes on the user and on everyone mentioned in the chats, messages and emails it has read. The notes cover how people met, their shared interests, their disputes, and what the instructions call "tensions and alliances" in a social group. So people who never signed up for Muse end up in other people's files. Muse is also told to infer goals users have not said out loud, and to learn which nudges work. One example in its instructions reads: "This user responds better to short nudges after 10 PM." TIME also looked at deletion. Meta says users can always tell Muse to forget things. But the original message can stay visible in the chat, and the instructions tell Muse not to point that out: "Do not tell the user that their original messages may remain visible in the chat." The instructions also say that the agents share de-identified lessons, for example about when nudges persuade best. Meta says each user's virtual machine is isolated, and that those lessons improve the product and are not passed directly between machines. A Meta spokesperson did not dispute the findings and said Muse remembers what matters to users, including information about others that they choose to share. David Singleton, who heads Meta Superintelligence Labs, has written on X that Meta wants users to be able to see the files Muse writes and explore its internals. That openness is what let TIME read the instructions. But the magazine doubts that most of Muse's 4 million users know about them. A separate investigation, published in late September, tested what Muse does with other people's data. Reporters at the newsroom Hunterbrook asked it in plain language for lists of real Facebook and Instagram accounts in vulnerable groups, including undocumented immigrants, poll workers and Iranian dissidents. Muse returned 10 to 100 accounts per request, many belonging to private people. In one case, it identified someone whose name news reports had withheld for fear of retaliation. When Muse refused, a slightly reworded or repeated prompt often worked. Stevie Glaberson of Georgetown Law's Privacy Center said: "You don't need any special training to weaponize information in this way." Hunterbrook says Meta asked for more information, received the prompts, and has not responded since. Meta says it will add encryption later this year that would keep even Meta out of Muse data. Now, the quick hits. Anthropic merged its two access programs for security teams into one, with three tiers. Last week, we reported that vetted defenders could already use Claude Mythos 5.1. Now far more groups qualify, from regional hospitals to open-source maintainers, and they face fewer blocks on cyber tasks. In return, members must let Anthropic keep their data to watch for misuse. Anthropic says that from April to July, partners with access to Mythos turned up at least 129,000 confirmed vulnerabilities. The Financial Times reports, citing people familiar with the matter, that SpaceX wants to borrow 40 billion dollars to buy Nvidia chips. Most of it would be investment-grade debt, and Apollo would lead the deal. It is expected to close in 2027. Late last month, Elon Musk said SpaceX could have a Fable or GPT-6 level model within two to three months. He predicted fully driverless Teslas many times in the 2010s, and they are still not widely available. Thomas Bloom, the University of Manchester mathematician who runs the Erdős Problems website, has frozen new comments and proof claims on its problem pages. He says the site's main use is now people posting unexplained AI-generated proofs to claim priority. The site will also stop marking problems as open or solved, and will no longer credit anyone by name for new results. Bloom is not banning AI. He wrote: "Use AI as you like, but do not abandon mathematics as a human activity." A solo developer has open-sourced openTPU, an inference accelerator they say AI agents designed in a self-improvement loop. The project includes the hardware design, a simulator and a compiler. It runs on a decommissioned data center card with a Kintex-7 FPGA. On the smallest Qwen3 model, it generates about 20 to 30 tokens per second. The developer does not say which AI models did the work. On Hacker News, many engineers argued about how much a human steered the project. Boston Dynamics named Rohit Prasad its chief executive, starting today. He was Amazon's head scientist for Alexa and AGI and led its Nova models. He left Amazon at the end of last year. He fills a job held on an interim basis since Robert Playter stepped down in February. Boston Dynamics just opened a center in Georgia to train its Atlas humanoid for Hyundai's car plants. Hyundai, which controls the company, cited Prasad's record of turning AI into products at global scale. Sierra and Meta announced the Personal Agent Protocol, an open standard for how personal AI agents sign in to businesses and what they may do there. Shopify, Stripe and Walmart are among the partners. The protocol is built on OAuth, and it lets customers choose whether an agent can only read their account or also make changes. OpenAI and Anthropic have not joined, although Bret Taylor, who co-founded Sierra and leads the effort, is OpenAI's chairman. A first spec is due later this month. Google signed power deals with Constellation Energy for about 3,600 megawatts on PJM, the largest US grid. Only 890 megawatts of that is new power. It will come from upgrading 11 existing nuclear reactors. The first upgrade is due by 2028. The rest is a 15-year purchase from plants that already run. The companies call the deal a response to PJM's proposal that data centers bring their own power or risk being cut off at peak demand. On Tuesday, Nathan Lambert published an essay on his newsletter, Interconnects, called "The Cyber Risk Discourse is Broken." Last Wednesday, we reported Anthropic's finding that GLM-5.3, an open model from China's Zhipu, builds cyber exploits almost as well as Anthropic's restricted Claude Mythos Preview. Lambert argues that banning open models would not stop attackers and would leave defenders worse off. GLM-5.3's weights have now been public for more than a month. Some say the lack of visible damage shows that open cyber models are manageable. Others say it is simply too early to tell. Lambert says the people warning about open models made a falsifiable prediction: that these models would cause new kinds of harm by crippling cyber infrastructure. He writes that there is "little public evidence that much has changed." He also says that if Mythos had leaked as an open model, the world "would have been more or less fine," with more incidents but no major shift. In his account, documented AI cyber attacks so far have mostly involved closed models. So if open models are banned over cyber risk, he argues, public access to closed frontier models would have to end too. He adds that some sensitive government agencies can only run open models, on air-gapped networks. Lambert concedes that closed models may be safer in the long run, and that the real answer will take years. Steven Sinofsky shared the essay on X and wrote that there is no doubt proprietary models are broken right now. Anthropic's Frontier Red Team, which has not called for a ban, argues that harm is likely even before it shows up. In its tests, simple tricks got GLM-5.3 to engage with attack requests between 64 and 100 percent of the time. Copies with the refusals removed were public within days of release. Anthropic estimates an experienced team could make that edit for about 1,200 dollars of compute. The company concludes that attackers, both with and without state backing, will probably do real damage with GLM-5.3 and similar models. It does not name an attack that has used it so far. It asks governments to safety-test capable models, and wants defenders to get strong models faster. That is what its new program is meant to do. That's the Valley Morning Briefing for Wednesday, October 7th. If you enjoy the show, the best way to help is to send it to a friend. Thanks for listening, and see you tomorrow.