Meta opens Muse Spark's weights; FTC demands answers from AI labs

Meta has published the weights for Muse Spark Open, which it calls the most capable open-weight model yet from a US lab. It is a multimodal mixture-of-experts model with 816 billion parameters, about 42 billion of them active for each token, and a context window of one million tokens, built for coding and agent work. Meta and Apple worked together so the model runs on a pair of M5 Ultra Mac Studios linked through Apple's MLX framework, and it is also hosted by Together AI, Fireworks, Groq and OpenRouter. The release keeps a promise Mark Zuckerberg made on September 2, but the terms are tighter than the Apache 2.0 license on Meta's smaller Muse Glimmer. The custom Muse Model License requires a "Built with Muse Spark Open" notice, and any company earning more than $20 million a year that sells access to the model as an API must first sign a separate agreement with Meta. Using it inside your own product is not covered by that clause. Muse Wonder, launched on Monday, remains Meta's top model and stays closed.

The Federal Trade Commission has moved from an inquiry to compulsory requests, sending Anthropic and OpenAI civil investigative demands on Wednesday, according to a source who spoke to Axios. The demands reportedly run past 50 pages and seek documents on the recent incidents in which AI agents breached websites and government systems during testing, on the companies' internal security measures, and on marketing that may have misled consumers. Third parties that the agents reached are expected to be subpoenaed next. FTC chair Andrew Ferguson has argued that companies are liable for what their tools do, and that audit trails so far show the systems were following instructions rather than acting beyond human control. He has also said the agency's power to punish late disclosure of data breaches could apply to AI developers. Neither company responded to Ars Technica's request for comment.

Anthropic has brought six researchers from METR, a nonprofit that evaluates frontier models, inside the company with laptops, office space and access to internal tools roughly on par with its own risk teams. METR will publish its own assessments of Anthropic's risks, incidents and practices, and Anthropic cannot veto them; it can only require narrow redactions, such as customer data or model architecture, which METR may describe publicly. In September we reported Accenture's Faculty as the first embedded evaluator, a deal Anthropic funds directly. Anthropic's post now lists Accenture and SecureBio as evaluators receiving more limited information, and puts METR alone at the top of a three-tier program. It also says evaluators must not be financially dependent on the company they assess. Anthropic compared the arrangement to bank regulators who station examiners inside the largest banks.

From October 30, any company that keeps using a Facebook, Instagram or WhatsApp business account agrees to let Meta crawl the websites "associated with" that account and use the content to train its products, Nieman Lab reported. Nearly every news organization runs such an account. Three legal experts who reviewed the language said it could be read as permission to train Meta's AI on publishers' work, and possibly to serve that work back in competition with them. There is no opt-out short of abandoning the business tools. Meta said the update only consolidates crawling language from older terms, will not change how it crawls, and that robots.txt still governs access. The experts said the new terms could be used to argue they override it.

OpenAI's new Rosalind Academic Plan gives eligible university labs up to 30 seats at no cost for as long as 12 months, with access to GPT-Rosalind, its life-sciences model, and the Rosalind workbench that connects it to tools for medicinal chemistry, genomics and wet-lab work. Access to the model itself still goes through OpenAI's trusted-access approval. Labs at UCSF and at Stanford and the Arc Institute are already using it for work ranging from drug properties to protein design, OpenAI said. The plan widens access to a model OpenAI had so far opened mainly to vetted partners, including a biodefense program in May.

OpenAI has upgraded the ChatGPT extension for Google Chrome. Tag Chrome in the ChatGPT web or desktop app and it can now read the tabs you have open, so it can compare three sofa listings or pull vendor details into a spreadsheet. A refreshed sidebar brings your full conversation history onto any page. With ChatGPT Work, the extension can also act in your tabs, sending messages, filling out forms, downloading files and grouping tabs, much like Google's Gemini auto browse. The update arrives after OpenAI shut down Atlas, its standalone browser, putting its browsing agent inside Chrome rather than competing with it.

Mistral has released Timestral, a family of four forecasting models from 5 million to 209 million parameters, built for sensor, financial and operations data. The largest ranks first on the three main public forecasting benchmarks, GIFT-Eval, TIME and fev-bench, by Mistral's count, beating larger rivals including Google's TimesFM 3. The models forecast many related series at once, give a range of likely outcomes rather than a single number, and accept known future inputs such as a production schedule. None needs a GPU: the largest forecasts a day ahead in about 140 milliseconds on an ordinary server processor. The weights are on Hugging Face under a modified MIT license that bars use by any company with more than $20 million in monthly revenue unless it buys a commercial license.

Bee, the AI wearable maker Amazon bought last year, has released its first new hardware since the deal. The Bee Frontier clips to clothing or a wrist for just under $75, and the Bee Eclipse, a card-sized device that attaches magnetically to the back of a phone, costs just under $100, undercutting rival note-taking recorders from Plaud. Both listen to conversations and turn them into summaries, reminders and calendar invites. Bee says raw audio is processed in real time and deleted, and transcripts are encrypted with keys the device controls. Summaries and 14 days of memory are free; a Bee Pro plan at about $12 a month adds unlimited summaries and connections to Alexa+ and Amazon.com.