Oct 16: Meta's Open Model Runs on Macs; Amazon Blocks Grok Bot

Meta and Apple released Muse Spark Open, an open-weight model that runs on linked Mac Studios, and Goodfire's Nick Cammarata argued that most neurons in AI models are readable. Also: Amazon blocks Grok Bot and Elon Musk floats a rival store, quick hits on Anthropic's IPO valuation, Microsoft's MAI-Base-2 and the FTC, and two essays that disagree about OpenAI's AI-written math papers. **Meta releases Muse Spark Open, built with Apple for Mac Studios** Meta released the weights of an 816 billion parameter model that runs across linked M5 Ultra Mac Studios, but it scores 39 on the Artificial Analysis Intelligence Index, below GLM-5.3 and Kimi K3. Meta AI · Apple Machine Learning Research · CNBC **Goodfire's Nick Cammarata says AI neurons are mostly readable** Cammarata argues that superposition was the wrong explanation and that about 90 percent of neuron poles in Qwen3-32B have one main job, while Neel Nanda says it is not a significant update to the big picture. Goodfire · LessWrong **Amazon blocks Grok Bot, Musk floats an Amazon competitor** Amazon is blocking Grok Bot from placing orders with the same warning it showed Meta's Muse, and Musk replied, "Maybe we should start a competitor to Amazon." Teslarati · Yahoo Finance · Cybernews · Practical Ecommerce · Courthouse News Service **Anthropic investors discuss a lower IPO valuation** Reuters reports that some Anthropic investors now discuss an IPO valuation of about 1.5 trillion dollars instead of about 2 trillion, after revenue grew 45 percent last quarter. Reuters **Microsoft AI finishes pretraining MAI-Base-2** Microsoft AI has finished pretraining a 2.8 trillion parameter model, and Business Insider reports that Mustafa Suleyman told staff it beats the base model behind OpenAI's GPT-5.6. Microsoft AI · Digital Today **FTC sends civil investigative demands to OpenAI and Anthropic** The FTC sent OpenAI and Anthropic demands of more than 50 pages each about safety incidents, safeguards, marketing and privacy, and plans to make top executives testify. Axios **Anthropic places six METR staff inside the company** Anthropic gave six METR staff access close to that of its own risk teams and says it cannot veto what METR publishes. Anthropic · Reuters **Accenture caps central Claude spending at 29 dollars per employee** 404 Media reports that Accenture now pays for only 29 dollars of Claude usage per employee each month, after some staff had spent 500 to 1,000 dollars. 404 Media · Anthropic **Former Meta engineer says Scion copied Mononoke's design** A former Meta engineer says Anthropic's Scion copied the data model behind Mononoke without credit, and now calls the attribution Anthropic added sufficient. sunshowers.io · sunshowers.io **Perplexity open-sources SPACE, its agent sandbox** Perplexity open-sourced the sandbox behind its Computer agent, says no model escaped its virtual machine in its own tests, and CoreWeave, Akamai and Crusoe will offer it to customers. Perplexity · Semafor **Two essays disagree about OpenAI's AI-written math papers** Derek Thompson argues the proofs are real but interrupt the thousand-year conversation of mathematics, Rikab Gambhir's satire on Terence Tao's blog calls them a marketing gimmick, and the Simons Foundation, Anthropic and OpenAI committed 15 million dollars to help mathematicians check AI results. Derek Thompson · What's new (Terence Tao's blog) · Simons Foundation *The most important news from Silicon Valley and AI. 15 minutes, every weekday morning.* **Transcript** Good morning. It's Friday, October 16, and this is the Valley Morning Briefing. Today: Meta released an open-weight model built to run on Mac Studios. Nick Cammarata of Goodfire says most neurons in AI models are readable. Amazon is blocking Grok Bot, and Elon Musk suggests he might start a rival. And two new essays disagree about the hundreds of AI-written math papers that OpenAI released. Let's go. Meta released the weights of a large new model on Thursday. It is called Muse Spark Open, and Meta built it with Apple to run on Mac Studio desktops. The model has 816 billion parameters, and 42 billion of them are active for each token. It reads text, images, video and audio, and its context window holds one million tokens. Meta calls it the most capable American open-weight model so far, and says it is built for coding and agent work. Apple published its own benchmarks the same day. The model is too big for one machine, so Apple links several M5 Ultra Mac Studios with Thunderbolt cables and splits the model across them. Two machines generate about 25 tokens per second. With four machines, the speed rises to about 37. Each Mac draws about 300 watts at its peak. The 512 gigabyte Mac Studio that the two-machine setup needs ships in late October, and Apple has not set a price. With today's smaller Mac Studios, the four-machine setup costs about 43,000 dollars. Meta's own model card gives the model a score of 39 on the Artificial Analysis Intelligence Index. China's GLM-5.3 scores 45, and Kimi K3 scores 44. Artificial Analysis scores Meta's hosted model, Muse Spark 1.3, at 48. In Meta's table, the open model leads on tool use and on some visual tests. It trails the Chinese models on terminal coding and on long software tasks. On Tuesday, we reported that Meta kept the weights of its stronger Muse Wonder models closed. Meta's AI chief, Alexandr Wang, told CNBC: "These aren't the most capable models that we've developed or released." He expects developers to try the open model at small scale and then use it more widely. The license is free for most uses. Companies that sell access to the model as a service and make more than 20 million dollars a year need a separate deal with Meta. This is also the first public partnership between Meta and Apple since 2017. Mark Zuckerberg thanked Apple's new CEO, John Ternus, on X. Ternus replied that it was amazing to run a model of this size so efficiently, "right from your desk." Wang told Axios that Meta is "excited for mommy and daddy to be playing nice again," and credited Daniel Gross, a former Apple executive now at Meta. Nathan Lambert, who writes the Interconnects newsletter, wrote on X that he had waited a long time for American open models to matter again. On Hacker News, many engineers said Meta limited its claim to American models because Chinese models score higher. Others praised Meta for testing and documenting the setup on many platforms. Meta says a full safety evaluation report will follow soon. Nick Cammarata, a researcher at the interpretability startup Goodfire, published a long essay on Thursday. He argues that the neurons inside AI models are mostly readable after all. Cammarata co-wrote some of the field's early papers with Chris Olah at OpenAI. For years, researchers have struggled with neurons that seem to respond to several unrelated things. The leading explanation was superposition: when a network runs short of neurons, it packs unrelated concepts into the same ones. So researchers trained extra models, such as sparse autoencoders, to separate those concepts. Cammarata helped propose that explanation. He now says it was wrong. His best-known example is a neuron in the old image model InceptionV1 that seemed to fire for both cat faces and the fronts of cars. He now says it detects V shapes. It also connects to V-shaped flames and fox faces, and round things like golf balls push it negative. He says most of the apparent mess came from weak tools and old architectures. With the tools of that time, researchers found almost no readable neurons in the old model's last layer. In a modern network of the same size, trained on the same data, about 99 percent are readable. In language models, his early audits of Qwen3-32B show that about 90 percent of neuron poles, meaning the positive and the negative side of each neuron, have one main job. Only two to four percent are not yet understood. He says these shares depend on where the thresholds are set. AI agents did much of the work and reviewed 100,000 attribution graphs. Cammarata also reports similar results in Kimi K3, a model with trillions of parameters, but he writes that full audits at that scale have not started. Cammarata ties the work to safety. He writes: "My fear is that deception will be right there, in plain sight, and we won't know how to read it." He wrote on X that he now thinks ambitious interpretability will be solved by the time it is needed most. In the essay, he calls that step "highly speculative." Critics focused on that gap. Alex Gibson, a commenter on LessWrong, compared the post to the optimism around sparse autoencoders in 2023, and said the X post left out the speculative label. The interpretability researcher Neel Nanda wrote: "This is interesting work, but I don't think it's a significant update to the big picture of interpretability." He agrees that neurons can help with practical debugging. But he says superposition is still real, because language models store far more facts than they have neurons. On Hacker News, some engineers asked whether AI agents doing the research can be trusted not to mislead the researchers. Cammarata says a technical paper with more detail on the language model experiments is due in a few weeks. Amazon is blocking Grok Bot, Elon Musk's personal AI agent, from placing orders. On Thursday, Musk responded by suggesting he might start an Amazon competitor. A Grok Bot user posted the warning their agent got from Amazon. The message says that an AI agent using the site without Amazon's permission breaks the terms every customer has accepted. Amazon showed the same message to users of Meta's Muse agent when it blocked that agent in September. Grok Bot was blocked before this week. A user told Musk about it on October 10, the same day Musk showed Grok Bot buying LEGO from LEGO's own website, which we covered on Monday. And in September, Amazon's robots file, which tells bots which pages they may visit, already blocked Grok, ChatGPT, Claude and others by name. Musk first told users to have Grok Bot buy what they want somewhere else. Then a user asked whether it was time for an X Shop. Musk replied: "Maybe we should start a competitor to Amazon." He gave no details. Amazon is worth about 2.7 trillion dollars, more than either Tesla or SpaceX. A rival store would need merchants, warehouses and delivery. Musk also promised on X that SpaceX would pay users back when a Grok Bot purchase goes wrong or their card details leak. But the company's published terms make users responsible for what their agents do. Grok Bot is also reaching more people. On Wednesday, we reported that it now comes in the 30-dollar Premium tier of the X subscription. It also runs inside Tesla cars, including every Cybercab. Amazon has not commented on Grok. In the Muse case, it said Meta never asked for permission and that the agent does not identify itself. Amazon says outside agents should operate openly and respect a store's decision not to take part. At the same time, Amazon's own feature, Buy for Me, buys products from other stores for its customers. Shopify has taken the opposite approach. Since September, it has shared its merchants' products with Meta's agent by default, and it earns money however the shopper arrives. About a week earlier, Paul Graham, the Y Combinator co-founder, wrote that Amazon banning agents is the first chance since Amazon was founded for a startup to compete with it. He added that if Amazon needs to ban agents, people must want to use them. Musk reposted it with one word: "Seriously." Amazon is also fighting a similar case in court. In August, a federal appeals court lifted an order that kept Perplexity's Comet agent out of Amazon accounts. The court said Amazon was unlikely to win its claims under anti-hacking laws. Last month, the full court declined to rehear the case. Now, the quick hits. Reuters reports, citing people familiar with the matter, that some Anthropic investors are now discussing an IPO valuation of about 1.5 trillion dollars. They had expected about 2 trillion. On Thursday, we reported that Anthropic's revenue grew 45 percent last quarter. Some investors are disappointed, because revenue more than doubled in the quarter before. Anthropic did not comment. Reuters says a cooler reception for the November listing could hurt the wider AI market. Microsoft AI, the in-house lab led by Mustafa Suleyman, has finished pretraining MAI-Base-2, a model with 2.8 trillion parameters. Microsoft estimates that its training recipe needs four to twelve times less compute than DeepSeek's best open base model. Business Insider reports that Suleyman told staff the model beats the base model behind OpenAI's GPT-5.6. The model is still in post-training, and Microsoft plans to license its weights to partners. On Wednesday, the FTC sent civil investigative demands, a kind of subpoena, to OpenAI and Anthropic. Axios reports, citing a person familiar with the matter, that each demand is more than 50 pages long. They ask for documents on safety incidents, safeguards, marketing and privacy policies. The agency also plans to make top executives testify. OpenAI says it will engage constructively. Demands like these gather evidence ahead of possible lawsuits. Anthropic has placed six staff from METR, the nonprofit model evaluator, inside the company, with access close to that of its own risk teams. Anthropic says it cannot veto what METR publishes. It can only require redactions in narrow categories, and METR may say publicly whether they were justified. On Tuesday, Reuters reported that Anthropic and OpenAI had wanted significant control over what evaluators could publish. 404 Media reports, citing an internal memo, that Accenture now pays for only 29 dollars of Claude usage per employee each month. Above that, a project budget has to cover the cost. Some staff had spent 500 to 1,000 dollars a month, and one says their access was cut entirely. Accenture is one of Anthropic's biggest partners. In December, Anthropic called its Claude Code rollout there its largest ever deployment. A former Meta engineer says Anthropic's Scion copied the design of the data model behind Mononoke, Meta's open-source source control server, without credit. The engineer led that design. On Wednesday, we covered Scion, the git replacement a small Anthropic team built with Claude in five months. The engineer has no problem with the copying itself and objected only to the missing credit. Anthropic has since added attribution, and the engineer calls it sufficient. Perplexity has open-sourced SPACE, the sandbox that has run every session of its Computer agent since July. Each sandbox is a small virtual machine, and passwords stay outside it. In Perplexity's own tests, no model escaped the virtual machine, though some got around network limits until Perplexity fixed them. On Thursday, we reported that experts call this year's rogue agent incidents failures of containment. Semafor reports that CoreWeave, Akamai and Crusoe will offer SPACE to their customers. Last Friday, we reported that OpenAI withdrew three of the more than 700 AI-written math papers it released this month. On Thursday, two essays took opposite views of that release. The writer Derek Thompson argues that the proofs are a real advance, and that they take something away from human mathematics. His essay is called "I Know Why the Mathematicians Weep." He accepts that many of the proofs solve problems that mathematicians could not solve for decades. He argues that science is both a craft and a conversation across generations. When OpenAI publishes hundreds of proofs that end hundreds of searches, he writes, it succeeds at the craft of science. But it interrupts the thousand-year conversation. He quotes the Cornell mathematician Steven Strogatz, who recalls that Archimedes wrote that future discoverers would add to his theorems. Strogatz says mathematicians feel they are in this together. Thompson also quotes the Fields medalist Hugo Duminil-Copin: "They have annihilated my field of research." Thompson warns that physicists, economists and writers may be next. The physicist Rikab Gambhir disagrees, in a satirical guest post that Terence Tao published on his blog. Gambhir compares the release to Borges's Library of Babel, a story about a library that holds every possible book. By today's research standards, he writes, that library would be the most successful research institution in history. He calls OpenAI's papers obscure, barely readable facts that might be true, released as a marketing gimmick before an IPO. Above all, he says OpenAI leaves the slow, unpaid work of checking and understanding the proofs to mathematicians. On Hacker News, readers argued over whether both views can be true at once. One commenter who works in group theory said the papers contain mistakes and duplicate proofs, but added that slop can be superhuman and useful. Others replied that poor writing matters little if the proofs are correct. Also on Thursday, the Simons Foundation, Anthropic and OpenAI committed 15 million dollars in combined funding and resources to help mathematicians check AI results and set standards. The foundation's president, David Spergel, said the math community should drive those discussions, "not any single company or institution." That's the Valley Morning Briefing for Friday, October 16th. If you enjoy the show, the best way to help is to send it to a friend. Thanks for listening, and see you on Monday.