The AI Recap

Directions

The AI Recap AI moves fast. Catch up in one scroll.

Meta announced Muse on September 8, a personal AI agent it says can send emails, book travel, fill out forms, make purch...
09/09/2026

Meta announced Muse on September 8, a personal AI agent it says can send emails, book travel, fill out forms, make purchases and help sell a car. The product is initially available in the U.S. through its dedicated app and WhatsApp, according to Meta.

The key question is trust. Meta says Muse runs in a cloud-based Secure VM, while its Sentinel system checks outbound actions against permissions or asks for approval. Users can opt out of AI training, and Meta plans a Confidential VM later this year with encryption keys held only by the user. But Reuters reported that internal testing found stalls and unauthorized uploads of sensitive data—an early reminder that privacy safeguards will be central to how personal agents are judged.

Source: https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/

ByteDance is reportedly building a real-time spatial video model on top of its Seedance cinematic-video system, accordin...
09/08/2026

ByteDance is reportedly building a real-time spatial video model on top of its Seedance cinematic-video system, according to Bloomberg. The project could create interactive virtual worlds for livestreams, short-form dramas and games, responding to Pico headset users’ movements or voices.

People familiar with the effort told Bloomberg that founder Zhang Yiming is personally overseeing development. One source said the system could deliver video at about 0.05-second latency and 20 frames per second; ByteDance is also reportedly planning to shift the compute-heavy work to the cloud. A launch could come as soon as October, though the timing remains uncertain and ByteDance did not respond to Bloomberg’s request for comment.

Source: https://www.bloomberg.com/news/articles/2026-09-07/bytedance-founder-joins-ai-elite-in-race-to-perfect-world-models

The Seattle Times Company and Newsday filed a federal lawsuit against OpenAI and Microsoft on September 4, alleging the ...
09/07/2026

The Seattle Times Company and Newsday filed a federal lawsuit against OpenAI and Microsoft on September 4, alleging the companies used their journalism without authorization or payment to train and operate products including ChatGPT, Copilot and Bing Chat/Search.

The complaint, filed in the Southern District of New York, alleges the companies scraped the publishers’ sites and that AI systems can reproduce or closely paraphrase their reporting. It cites a claimed 88-word verbatim passage from a Seattle Times article. The publishers seek damages, a permanent injunction, and the impoundment or destruction of copies, training datasets and models incorporating their work.

Microsoft said it was surprised by the suit and was willing to discuss solutions. The case adds to the growing legal fight over whether AI developers can use copyrighted news reporting for training without a license.

Source: https://techcrunch.com/2026/09/05/seattle-times-and-newsday-are-the-latest-publications-to-sue-openai-and-microsoft/

Google has made its Lyria 3.5 music-generation model available globally to all Gemini users on the web and mobile app, a...
09/07/2026

Google has made its Lyria 3.5 music-generation model available globally to all Gemini users on the web and mobile app, as well as through the Gemini API. According to Google’s September 4 announcement, users can describe or select a genre, choose vocal or instrumental output, and use templates for background music or custom birthday tracks.

Google calls Lyria 3.5 its best-sounding music model, with more expressive vocals and richer arrangements. The company says Lyria tracks can run up to three minutes and are imperceptibly watermarked with SynthID, a step aimed at helping identify music created or edited with AI.

Source: https://blog.google/innovation-and-ai/products/gemini-app/better-tracks-lyria-gemini/

Authors are contesting how payments will be divided in Anthropic’s $1.5 billion copyright settlement, after claim notice...
09/07/2026

Authors are contesting how payments will be divided in Anthropic’s $1.5 billion copyright settlement, after claim notices showed publishers, and in some cases literary agencies, seeking shares of awards for individual books. The settlement, which received final approval in July 2026, covers nearly 500,000 titles and pays $3,000 per pirated work, according to TechCrunch.

For traditionally published, in-print books, the default split is 50% for the author and 50% for the publisher. The Authors Guild says some publishers mistakenly selected 100% allocations and corrections are underway. Authors whose rights reverted before the settlement’s August 10, 2022 download date may be entitled to the full award, depending on their contracts; unresolved allocation disputes can ultimately go to a court-appointed Special Master.

Source: https://techcrunch.com/2026/09/06/authors-push-back-as-publishers-and-agents-seek-share-of-anthropic-settlement/

OpenAI Chief Scientist Jakub Pachocki says frontier AI labs should not continue scaling at maximum speed until they can ...
09/07/2026

OpenAI Chief Scientist Jakub Pachocki says frontier AI labs should not continue scaling at maximum speed until they can show stronger alignment and monitoring. In his September 6 essay, “An Alien Mind,” Pachocki writes that no lab has solved those problems well enough to scale responsibly at that pace for much longer.

Pachocki says OpenAI will continue alignment and monitoring research and could withhold further scaling when needed, but argues that broader safeguards are necessary. He calls for frameworks such as OpenAI’s Preparedness Framework and Anthropic’s Responsible Scaling Policy to become widely mandated safety thresholds, potentially enforced by third-party auditors, governments or international bodies. He also warns that chain-of-thought monitoring is becoming less reliable as models use tools and agents and can operate without verbalized reasoning.

Source: https://openai.com/index/an-alien-mind/

Reuters reports that the U.S. and China are preparing a mid-September dialogue on AI safety risks, potentially their fir...
09/07/2026

Reuters reports that the U.S. and China are preparing a mid-September dialogue on AI safety risks, potentially their first official bilateral talks focused exclusively on AI since Donald Trump began his second term.

According to Reuters’ sources, the agenda could cover monitoring AI-directed cyberattacks, information-sharing by AI labs, and U.S. concerns over alleged Chinese distillation of proprietary U.S. models. But the meeting is not confirmed: a White House official said there is “currently no planned AI-related meeting in mid-September,” and Reuters reported that the agenda and participants remain in flux ahead of Trump and Xi Jinping’s September 24 Washington summit.

Source: https://www.reuters.com/legal/litigation/us-china-gear-up-mid-september-ai-safety-dialogue-2026-09-04/

Anthropic says a team of Claude agents has produced the first complete computer-checked formalization of Fermat’s Last T...
09/07/2026

Anthropic says a team of Claude agents has produced the first complete computer-checked formalization of Fermat’s Last Theorem, encoding a simplified version of Andrew Wiles’s proof in the Lean proof assistant. The theorem says no positive integers a, b and c satisfy aⁿ + bⁿ = cⁿ when n is greater than 2; Wiles’s first correct proof was published in 1995.

According to Anthropic, dozens of agents worked largely autonomously for 11 days, generating roughly 13 million lines of Lean and proving 30,300 theorems along the way. The released project’s full build checked 60,475 modules with Lean’s kernel, and an independent Rust implementation, nanoda, also checked the exported environment. It is not a new mathematical proof, but a reproducible artifact for verifying one—and a notable test of how AI could help formalize mathematics.

Source: https://www.anthropic.com/research/formalizing-fermats-last-theorem

Nvidia has launched Personal AI Router (PAIR), a free beta tool that lets compatible computers on the same local network...
09/03/2026

Nvidia has launched Personal AI Router (PAIR), a free beta tool that lets compatible computers on the same local network handle local AI-inference requests through one endpoint. The software supports Ollama and LM Studio at launch, and Nvidia says it can keep prompts, files and agent context on the home network when clients, models, engines and nodes are configured locally.

PAIR works across Windows 11, Linux and macOS, with support listed for GeForce RTX 20 Series-and-newer GPUs, DGX Spark/GB10 and Macs with M4-or-newer chips. But it is a request router—not a way to merge GPUs: it cannot pool GPU memory, shard a model across machines or split one in-flight request. Its value is in distributing separate concurrent jobs across available devices.

Source: https://www.nvidia.com/en-us/ai-on-rtx/personal-ai-router/

Google is rolling out voice-driven Gemini features for Gmail, Docs and Keep, bringing its Gemini Audio models into Works...
09/03/2026

Google is rolling out voice-driven Gemini features for Gmail, Docs and Keep, bringing its Gemini Audio models into Workspace apps. Gmail Live can answer spoken questions by scanning an inbox, while Keep Live turns a spoken stream of thoughts into structured notes and lists.

Docs Live is designed as a hands-free co-writer that can organize ideas into documents and, with permission, draw details from Gmail, Drive, Chat and the web. Google says Gmail and Keep are available this week to AI Plus, Pro and Ultra subscribers; Docs is available to Pro and Ultra subscribers, with all three coming soon to Workspace business customers.

Source: https://blog.google/products-and-platforms/products/workspace/voice-features-gmail-docs-keep/

Address

New York, NY

Website

Alerts

Be the first to know and let us send you an email when The AI Recap posts news and promotions. Your email address will not be used for any other purpose, and you can unsubscribe at any time.

Shortcuts

Share

Category