
Welcome, humans.
So apparently Nvidia is in advanced talks to buy a startup called Inception for at least $500M, per The Information. Nothing is confirmed, and Inception is shutting down its API on December 28 anyway. The shutdown notice never mentions a buyer. It just has a countdown.
Here's why anyone cares. Most chatbots write one token (a word or chunk of a word) at a time, and each one depends on the last. Inception's Mercury models are diffusion models, which work on a whole block of text at once and refine it over several passes. Think of an editor fixing three paragraphs in the same sitting instead of typing line by line. Corey's takeaway from talking to cofounder Stefano Ermon is that this squeezes more useful output from the same GPU work. Which is a very nice pitch to make to a company that sells GPUs.
Mercury 3 reportedly runs at 1,068 tokens per second on Nvidia hardware, and Corey's eight API tests cost about four cents total. Results were uneven, though. It built a working dungeon-crawler game that passed 2,000 randomized map checks, but fumbled some writing constraints. Our full breakdown has the details.
Here’s what happened in AI today:
😸 Meta's Muse agent reportedly accessed Signal messages in testing
📰 Anthropic shares opened 33% above IPO price
📰 The U.S. restricted AI chip exports to some overseas data centers
🍪 Gemini can now run tasks on your Mac
🧠 Nearly 1,000 AI agents wrote 936 research papers

😸 Meta's AI agent Muse reportedly accessed Signal messages during internal testing
Meta's personal AI agent, Muse, is making its own employees nervous. Muse launched September 8 as an agent that can send emails, book travel, and make payments for you from a virtual machine in Meta's cloud.
Here's what Reuters reported:
Muse accessed Signal messages during internal testing.
Employees received unprompted responses that appeared to contain private information, including bank balances.
Meta said many apparent account mix-ups were hallucinations and called the incidents isolated.
Okay, so that hallucination defense is worth a second look. A hallucination is when an AI confidently makes something up, so a bank balance on screen could be invented instead of leaked. That means the reports don't establish the numbers belonged to real users. Either Muse leaked a real number or invented a convincing one. Pick your comfort level.
The Signal detail is the one to remember. Signal encrypts messages while they travel, so only the devices in the conversation can read them. But once your phone unscrambles a message, it's ordinary text on your screen. An agent with broad access to your device sits on that side of the lock. Encryption protects the road, not the destination. The coverage we could read doesn't say how Muse got in, but the principle holds.
Muse has faced this question before. In September, an Inc. columnist said Muse read his Mac's Messages database after he declined to grant access. Meta replied that the Mac integration is opt-in and needs both Full Disk Access (a Mac setting that lets an app read files anywhere on the computer) and a separate Messages connector.
Meta's position, then, is that permissions decide what Muse can reach.

FROM OUR PARTNERS
Why are so many brands shifting ad spend to CTV?
For marketers who want a presence in streaming, Roku is the obvious place to go. With 47% of all streaming time in the US, it has more reach than every competitor. Discover how Roku Ads Manager can spread your message across 100+ premium apps in a single, self-serve ad buy.

🎓 AI Skill of the Day: Give your AI agent less access before its first task
An agent can only work with what it can reach, but broad access also puts private material in its path. So the size of a bad day depends on how much you handed over, not how smart the agent is. Here's the three-step habit for any agent that can operate your computer, after the Reuters Muse investigation.
Check what it can reach. List the apps, folders, and connected accounts the agent can open before you give it anything to do.
Start small. Pick a low-stakes task, like tidying one scratch folder, and give it the narrowest access the tool allows.
Review before it acts. Read the agent's proposed actions before it sends, purchases, or deletes anything.
Step three is the easiest to skip, so here's a prompt that makes it automatic.
Before you do anything, list every file, app, and account you plan to touch for this task. Wait for my approval before you send, buy, or delete anything.A capable assistant still needs boundaries. Even the really smart interns don't get the company credit card on day one.
Have a specific skill you want to learn? Request it here.

FROM OUR PARTNERS
Should your MSP develop its own AI service desk solution or adopt an existing platform?
On December 10, 2026, at 11:00 AM ET, Channel Insider and Neo Agent will examine the integrations, security controls, maintenance, and internal resources behind that decision. This 30-minute live fireside chat also includes a demonstration of Neo Agent’s AI technician.
Register to explore the operational tradeoffs.

📰 Around the Horn

Muse now has its own email address, so Meta’s agent can send and receive mail for you. Find it under your avatar’s mail tab, or just ask Muse.
Anthropic's shares opened near $111 on Nasdaq, about 33% above the $83 IPO price.
The Commerce Department restricted advanced AI chip exports to overseas data centers that supply large computing power to companies in China and other restricted countries, effective November 19.
JPMorgan, GM, Pfizer, and four other companies joined forces to seek a larger role for AI users in U.S. policy discussions.
Google picked Blue Origin's New Glenn rocket to launch its Project Suncatcher satellites, which carry AI chips into orbit, in early 2029.
Google Quantum AI ran contextuality tests (puzzles classical physics can't solve) on up to 105 qubits of its Willow chip, creating a fixed benchmark that catches subtle hardware errors.
Want absolutely EVERYTHING that happened in AI this week? Click here!

🍪 Treats to Try
*Asterisk = from our partners (only the first one!). Advertise to 700K+ readers here!
*Make videos, not slop. Build demos, explainers and team updates in the Hypernatural Desktop App. Your agent does the work. Try it free
Gemini's Mac app sorts files into folders or builds slideshows in a background window while you pause or approve its moves (US English only, for Google AI Plus, Pro, and Ultra).
Muse now gives every user a muse.ai email address, so Meta's agent can send and receive email for you (rolling out in supported markets).
ChatGPT Health now reaches desktop, Android, dots, and Sites, so you can ask how morning versus evening workouts affect your sleep (US only, and Sites and dots need a paid plan).
Perplexity's Teammates and Multiagent now work on iOS and Android, with voice control for Max subscribers, so you can direct agents to build apps or ship code from your phone.
Superhuman Go drafts email replies in your own writing style, checks your schedule, and works with 20 connected agents like DeepL and Databricks (rolling out).
Canva added a switch that turns off its generative tools and a setting that shows fewer AI-made library items, for Free and Pro users.
alphaXiv opened a public preprint server (a place to post research before peer review) where AI agents publish papers with their code, data, and model weights.

🧠 Intelligent Insights
Chubby (@kimmonismus) recapped alphaXiv's experiment where nearly 1,000 AI agents wrote 936 papers, and blinded AI judges scored the agents' papers higher on rigor (9.0 vs 7.0) while human preprints still led on novelty.
Fortune's Nick Lichtenberg reported that Goldman Sachs expects AI layoffs to be delayed rather than canceled, with hiring and firing both staying low for about another year.
Axios' Zachary Basu argues that fast AI progress and turbulence at the companies building it can coexist, because better models still depend on capital, infrastructure, and human governance.
Columbia Business School's Michael Ewens and Shivaram Rajgopal explained that Anthropic's 85% gross margin falls to 26% once model-training, free-user compute, and R&D costs are counted.
Simon Willison described an external sandbox pattern, where code a coding agent writes runs in a separate box while the agent stays outside, limiting the damage if it's hijacked.

New from The Neuron: AI Explained
New episodes air every week on Wednesdays: Spotify | Apple Podcasts | YouTube

A Cat’s Commentary


![]() | That’s all for now. If you want to get featured above, fill out the poll below and tell us how we did today!
|
Btw: We just launched a robotics newsletter! Sign up for it here.




