AI Daily Brief — 2026-10-15

Muse Spark Open, a new natively multimodal model with 816B total parameters (42B active) and a 1M-token context window, is now available. This model brings leading agentic capabilities to desktop users, specifically on Apple M5 Ultra Mac Studios. Designed for coding and agentic work, Muse Spark Open excels at long-horizon tasks, complex software engineering, and multi-step reasoning. Developers are encouraged to share their creations on the Hugging Face Forums.

Modern language models are more interpretable than commonly thought. Research like Anthropic’s “Towards Monosemanticity” (2023) and OpenAI’s “Language Models Can Explain Neurons in Language Models” (2023) show that while some neurons are hard to interpret, a significant portion can be understood. The concept of recursive generalization, where similar things are grouped together, suggests that interpretability is possible by examining how the network learns and organizes information, offering a fractal understanding of its functions.

Granite Switch 4.1 models, introduced as part of the Granite Switch project, offer a modular approach to model development. These models, equipped with a Decision adapter and a fast ASR model, can efficiently handle natural-language responses, general and specialized decisions, and speech-to-text tasks from a single hosted endpoint. In a comparison, Granite-Switch-4.1-30B demonstrated significantly faster completion of a 5-step RAG workflow compared to Jev-27B.

Nvidia has launched Nemotron 3.5 Super VL, a new 120B-parameter open model designed for enterprise agents. According to Nvidia, this model surpasses frontier models on specialized benchmarks. It is capable of assisting with complex tasks such as extracting figures from tables, interpreting charts, and checking related webpages, which are crucial for enterprise applications like reviewing credit agreements. This release marks a significant advancement in AI tools for businesses.

Meta and Apple are collaborating to enhance the performance of Muse AI models on Macs. Meta's new natively multimodal model, with 816 billion parameters, excels at complex tasks like software engineering and multi-step reasoning. Apple has also released a technical post about MLX, an array framework designed for machine learning tasks on Apple Silicon, alongside Muse Spark Open and two other open models.

HiRO-ACEv2 is an advanced AI model designed for simulating extreme weather conditions with high precision. Building upon its predecessor, HiRO-ACE, which was introduced in 2026, HiRO-ACEv2 features a new downscaling model, HiROv2. This updated model can predict multiple high-resolution fields simultaneously, rather than being limited to rainfall, thereby enabling it to capture a broader spectrum of extreme weather events. Furthermore, HiROv2 demonstrates enhanced proficiency in forecasting such events.

Bloomberg is making its ASKB conversational AI interface available to all Terminal users. This allows users to integrate their own content with existing financial data, feeding the system with a new level of context. The generative AI-based ASKB interface will benefit from users adding their own information, enhancing its capabilities for financial analysis and information retrieval.

Google has launched its "Universal Cart" shopping tool in the U.S., allowing users to collect and monitor products found across Google Search and the Gemini app. This tool, first previewed at Google I/O in May, is now available to American users.

Google has introduced a "universal" online shopping cart designed to simplify holiday shopping by allowing users to save products from various retailers into a single digital space. This new feature integrates across different Google platforms, enabling users to add items via the Gemini app and Google Search AI Mode. While future plans include integration with Gmail and YouTube, a timeline for these additions has not yet been announced.

An OpenAI lead engineer secretly built a wiretap using his wife's ChatGPT account. He hid the device in toilet paper behind the executive boardroom TV. The engineer was caught after 48 hours. This incident highlights a significant security breach within OpenAI, involving a high-ranking employee and a sophisticated, albeit short-lived, surveillance operation.

A federal judge denied bail for a former OpenAI engineer accused of planting a recording device in an executive conference room. Prosecutors allege the Saudi national communicated with a Saudi royal about an AI initiative. This development comes as FOX News Channel, a leading cable news network, continues to cover significant legal and technological stories, maintaining its position as a top-cited and trusted news source for many Americans.

U.S. federal agents have arrested an OpenAI engineer for allegedly planting a sophisticated recording device in an executive conference room at the company's San Francisco headquarters. The engineer, who works for the company behind ChatGPT, was taken into custody following reports of the listening device's discovery. This incident has raised concerns about security within the prominent AI firm.

Bank of America (BofA) hired 1,000 contractors in Q2 to assist 3,000 internal staff in patching software flaws. These vulnerabilities, found in both proprietary and vendor software, were exposed by AI models such as Mythos. This strategic hiring aims to bolster BofA's efforts in addressing advanced artificial intelligence-driven security challenges and maintaining robust software integrity across its operations.

Abdulrahman Alfozan, an OpenAI employee, was arrested by the FBI on October 9. This arrest came after he, his wife, and a family member had booked one-way tickets from the US to Saudi Arabia for October 12. He was charged with allegedly using a hidden recording device within OpenAI.

Anthropic has committed to stronger independent, third-party assessment of its models and safety practices, following its pledge to partner with independent external auditors or evaluators as part of the White House Accord on Super Intelligence. The company has granted a team from METR, a nonprofit evaluating frontier AI models, employee-like access to its internal tools and systems. This includes providing METR employees with company hardware and office space, enabling them to assess alignment risks similarly to internal teams. Additionally, Anthropic is releasing a new version of its Usage Policy.

Anthropic has issued a warning that artificial intelligence "worms" could infest home devices within months. Jordan Rae Kelly, senior managing director and head of cybersecurity at FTI Consulting, discussed this concern on "The Takeout." This potential threat highlights the evolving cybersecurity landscape as AI technology becomes more integrated into daily life, raising questions about device security and data protection for consumers.

A judge has ordered OpenAI to provide Apple with expedited access to hiring materials as part of a trade secrets lawsuit filed by Apple in July. While not all of Apple's requests were granted, U.S. District Judge Edward Davila ruled that Apple could access interview materials and private LINE messages between individuals named in the case, including OpenAI employees Chang Liu and Alyssa Peng.

OpenAI has agreed not to use trade secrets that Apple claims were stolen by former employees. This agreement comes as a lawsuit between the two companies plays out. As part of the arrangement, OpenAI will also collect forensic images from the devices of 14 employees, ensuring compliance and addressing Apple's concerns regarding the alleged theft of proprietary information.

The author criticizes Anthropic for insufficient attribution regarding its Scion VCS, alleging it closely resembles Mononoke, a component of Meta's Sapling ecosystem. Despite Anthropic adding attribution after initial complaints, the author argues that crediting Sapling for “hard-won lessons” is inadequate, as Scion's similarities to Mononoke are much deeper. The author views this as another instance of large AI companies appropriating others' work.