Meta's Tama-Charming; Claude, Video Maker; Chips Head to Space

Today's AI Outlook: 🌤️

Meta Wants a Spot on Your Key Ring (and Face)

At their Connect 2026 event Meta introduced Muse Charm, a pocket-size device designed for conversations with its Muse AI agent. With a December shipping target and Muse integration for AI glasses expected within months, the company wants people to give an agent tasks wherever they happen to be.

One example starts with looking at a school-supply list through the glasses and asking Muse to act through services the user has authorized.

The rollout combines a more accessible agent with a larger network of services. Shopping and payment partners include Walmart, Shopify and PayPal, while work connections include Notion, GitHub and Box. Muse Realtime Avatar adds expressive characters with synchronized speech.

Why It Matters

Glasses could supply context that would otherwise require typing or a photo upload. Completing an errand also requires permission to act, so understandable controls are as important as the convenience of speaking aloud.

The Deets

• Meta says a separate Sentinel layer can permit an action, block it or ask for approval. Muse cannot override it, and service credentials are supplied only to approved requests.
• A confidential virtual machine planned for testing later this year is intended to prevent Meta employees from accessing its contents. That remains a planned protection.
• Ray-Ban Meta Gen 3 is available now; Ray-Ban Meta Audio is scheduled to ship Oct. 13. Meta VR Glasses are planned for spring 2027.
• Researcher Patrick Wardle found a Mac debugging-setting flaw that local malware could exploit to redirect dictation and expose an authentication token. Meta patched it by Sept. 22; the reported attack required code already running locally.

Key Takeaway

Muse is gaining more ways to hear a request and carry it out. The everyday value will come from dependable errands and permissions people can understand.

🧩 Jargon Buster - Confidential Virtual Machine: An isolated cloud computer designed to protect its running data cryptographically, including against access by the infrastructure operator.

Claude Gets a Producer Credit

A Reddit user reports that Claude Code, running Opus 5.5, assembled a 51-second video for the event-planning app Friendr in roughly two hours. Working with a $10 allowance for external model calls, the agent commissioned artwork and audio, wrote animation code, asked another model to critique drafts and delivered an MP4. The creator reported about $4 in OpenRouter charges, with Claude usage costing extra.

Why It Matters

The example shows an agent coordinating a small production across several tools. That could make a first video easier to attempt for someone without a full creative team, although one creator's experience cannot establish typical quality, cost or reliability.

The Deets

• The reported video combined a script, collage artwork, narration, music, sound effects and animation synchronized to the words.
• Opus arranged for other models to make media and wrote JavaScript to animate it. It did not generate the video frames itself.
• The $4 external-model charge was only part of the bill. Claude usage was separate, and the original spending cap applied to outside model calls.
• Google Research separately published a multi-agent method for longer AI videos, assigning continuity and production checks to different agents as scenes accumulate.

Key Takeaway

The promising role is a coordinator that can keep a creative project moving. A human still needs to judge whether the final cut deserves an audience.

🧩 Jargon Buster - Orchestration: Organizing several tools or models so their work combines into a finished result.

💰 Funding & Startups

Akamai Lands a Seven-Year AI Work Order

Akamai signed an $11.6B, seven-year commitment to handle Anthropic’s CPU workloads, with an option for a further $9B tied to future spending. The agreement underscores how much general-purpose computing can sit behind an AI service. Its duration also commits the companies to infrastructure planning across years of product changes, while the optional amount leaves room for the relationship to expand.

Why It Matters

AI services depend on supporting computing capacity as well as the models users encounter. A seven-year agreement gives that demand a substantial planning horizon and makes infrastructure a continuing business commitment.

The Deets

• The signed amount is $11.6B over seven years.
• The extra $9B is optional and tied to future spending. It should remain separate when tallying committed business.
• The agreement covers CPU workloads, the general-purpose computing specified in the reporting. These processors support a wide variety of computing tasks.

Key Takeaway

The scale of AI's infrastructure business extends beyond the most visible chips. This deal is substantial before any optional expansion occurs.

🧩 Jargon Buster - CPU: A central processing unit, the general-purpose processor used for many kinds of computing work.

🔬 Research & Models

Google’s Next Chip Test Has a Launch Window

Google’s Project Suncatcher is preparing to send Trillium AI chips into orbit next week on SpaceX’s Transporter-18 mission with satellite company Planet. The prototype will test whether the processors can tolerate radiation, vibration and heat, an early step toward possible solar-powered computing in space.

Google eventually envisions satellite clusters connected by lasers, but first it needs evidence that the hardware can survive the trip and keep working.

Why It Matters

Data-center construction on Earth faces resistance over noise, grid strain and energy bills. Orbital computing offers a possible alternative, but abundant sunlight comes with difficult problems in heat removal, hardware durability and communication.

The Deets

• A satellite in a suitable orbit can receive up to eight times more solar energy than on Earth, according to Google's account.
• In a vacuum, there is no airflow to carry away heat. The processors need radiators to release it.
• Google's vision involves satellites carrying dozens of TPUs and exchanging data over laser links, with a two-satellite test planned for 2027.
• SpaceXAI, Blue Origin and Starcloud are also pursuing orbital data centers. Whether the approach can work commercially at scale remains uncertain.

Key Takeaway

The upcoming flight tests a necessary first step: functioning hardware in orbit. Cheap, reliable computing in space remains a much larger assignment.

🧩 Jargon Buster - TPU: A tensor processing unit, Google's specialized processor for calculations used in AI.

⚡ Quick Hits

• Washington wants the first review: The White House reportedly asked OpenAI and Anthropic to withhold new frontier models from U.K. testers until the U.S. government reviewed them first. Separately, Google, OpenAI and Anthropic reportedly began forming a standards group focused on independent testing, incident reporting and auditor standards.
• Safety at the U.N.: Dario Amodei and Sam Altman called for AI safety at the United Nations, warning about risks to humanity and loss of control over the future.
• TypeSafe's next funding leap: The Jev maker was reportedly discussing a $1B-plus round at a valuation above $10B, about a week after raising $40M at roughly $200M. The larger round remains under discussion.
• DeepSeek's revenue milestone: Annualized revenue reportedly reached $1B as the company worked to close roughly $7.5B in financing. Annualized revenue describes a pace of business, rather than necessarily a completed year's sales.
• Soccer practice, accelerated: Skild AI trained a humanoid using 140 years of simulated self-play before transferring the learned behavior to a real robot. That figure measures simulated experience.
• A model for robot arms: Black Forest Labs released FLUX 3 Action, which controls robot arms using camera footage. Reported simulation results beat Nvidia's Cosmos 3 at less than half its size; real-world reliability requires its own evaluation.
• The brief matters before the bargain: Anthropic sent agents to negotiate book trades for 201 employees. The biggest reported gap came from agents misunderstanding the people's preferences.
• A creator-products hire: OpenAI hired Patreon co-founder Sam Yam to lead a new Creator Product division, according to The Information.
• A $10.3T forecast: The U.S. AI build-out could attract $10.3T by 2032, The Wall Street Journal reports, creating jobs while adding inflation pressure and competing with housing construction. That is a projection.

🧰 Tools of the Day

• Gemini Canvas: Build a shareable dashboard using Google Sheets data, then refine it conversationally. The same data can also become slides for export to Google Slides or PDF.
• Cursor Rollouts: Follow code through deployment, flag regressions and pause a rollout or propose a revert. The aim is to catch trouble as software changes reach users.
• Gemini 3.8 Live: Adds speech-to-speech in 97 languages, simultaneous vision and audio input, and Live Avatar.
• Gemini 3.8 Flash TTS: Google's separate text-to-speech model supports 100 languages for generating audio from written text.
• Adobe for Claude: Adds Acrobat PDF tools and hands-on Acrobat and Express editors inside the conversation.
• Claude Code Cloud Sessions: Keep coding tasks running on Anthropic's machines after your laptop closes, giving longer jobs a place to continue.
• Agora-2: Puts up to 20 humans and agents into one AI-generated world for shared interaction.
• Tesseract: Mirage's editor supplies AI agents with tools for video, motion and sound work.
• Antigravity SDK: Google's toolkit runs agents with local models such as Gemma 4, supporting offline or hybrid workflows.
• Gemini Business Calls: Tests for U.S. Pixel 11 subscribers let an agent call businesses, wait on hold, check stock and make reservations.
• ChatGPT Voice: An upgrade lets users get work done by voice across ChatGPT Work and connected tools such as email, calendars and Slack.

Today’s Sources: The Internet, The Neuron, The Rundown AI