In Today’s Issue:

🤗 NVIDIA buys Hugging Face for $12.93 billion

🎓 New York City bans AI through eighth grade

🛡️ Google ships a Gemini model only defenders can have

🪄 Meta's Muse Spark 1.3 does more by doing less

✨ And more AI goodness…

⚡ The Signal

Two labs shipped an agent model on the same day, with opposite theories of how an agent should spend its time.

Google's Gemini 3.8 Flash is built to work harder, winning by "executing extra reasoning steps, and calling tools iteratively" and spending more tokens on purpose. Meta's Muse Spark 1.3 went the other way: its biggest jump yet came with about 20% fewer tool calls and 25% fewer tokens than the version before it, plus a new habit of stopping to ask when a prompt is ambiguous. Both are chasing the same customer, the developer who leaves an agent running for hours, and that customer pays by the token. Underneath both sits the same distribution layer, which NVIDIA has now agreed to buy for $12.93 billion. And in New York, half a million children up to eighth grade have been taken out of the market entirely.

All the best,

Kim Isenberg

NVIDIA announced today that it will acquire Hugging Face for about $12.93 billion, taking ownership of the platform where most of the world's open models and datasets are published. Jensen Huang promised a light touch: the team keeps "their same iconic 🤗 brand", and "NVIDIA compute will not be required to build on or deploy through Hugging Face." Reports of an agreement first surfaced in late August. NVIDIA already has all the GPU demand it can serve; what it did not own until today is the default download address of the entire open-model ecosystem.

👉 tl;dr: NVIDIA now owns the distribution layer of open AI, and says it will not wire it shut around its own chips.

2 minutes. Your URL. A customer profile worth using.

Most founders can describe their product. They can't describe their customer. Not in a way that actually changes how they sell.

HubSpot for Startups built a free tool to fix that. Paste in your URL, answer a few quick questions, and it generates a structured profile of your best-fit customer. Firmographics, buying triggers, the works.

Takes 2 minutes. No spreadsheet required.

Build Your Free ICP.

🎬 Watch This

– Mark Zuckerberg, founder and CEO, Meta, on X

Meta's chief AI officer spent launch night dunking on Google.

An hour after Zuckerberg's post, Alexandr Wang quote-tweeted Artificial Analysis's line that Muse Spark 1.3 now trails only Claude Fable 5.1 and Claude Opus 5 on its Intelligence Index, and added: "i really hate to say it, but... gemini who?" Gemini 3.8 Flash had shipped that same day.

Meta shipped an agent model yesterday whose headline improvement is that it does less work. Muse Spark 1.3 went live in Muse Code and the Meta Model API, and Mark Zuckerberg called it the biggest single jump his team has made on coding and agentic work. The number Meta leads with is an efficiency one: the model makes about 20% fewer tool calls and burns about 25% fewer tokens than Muse Spark 1.2, because it "takes fewer turns where not needed and is less verbose, while having a cleaner overall coding style".

What actually changed is the model's manners. Meta trained 1.3 to "sustain longer-horizon work by collaborating with users", which in practice means it asks a clarifying question when a prompt is ambiguous, pulls the human in when it is stuck, and checks before doing something it cannot undo. Meta also says it is better calibrated on "what constitutes irreversible actions" and harder to talk out of its instructions. It reads a 1M-token context on the Meta Model API and takes text, images and video.

The scores land exactly where the model was built to work. On DeepSWE v1.1, a test of long-horizon agentic coding, Meta's max reasoning configuration scores 75.4 against Muse Spark 1.2's 55.0, ahead of Opus 5 at 74.0. On MRCR, which measures how much a model still recalls across a very long document, it goes from 66.3 to 98.5, the best numbers anywhere in Meta's table. That is twenty points of agentic coding and thirty-two points of long-context recall inside a single version, and it is precisely the range where a long job either finishes or quietly loses the thread.

Why it matters: Token count has become a product feature. If Meta is right that the win comes from fewer and better-chosen steps rather than more of them, the cost of leaving an agent running all day falls without anyone needing a bigger model.

Sources:
🔗 https://developer.meta.com/ai/models/muse-spark/
🔗 https://research.meta.ai/blog/introducing-muse-spark-1-3
🔗 https://www.unite.ai/meta-launches-muse-spark-1-3-citing-gains-in-coding-and-agentic-tasks/

The experiment in one sentence: take mice that are already old, put them on Ozempic for the rest of their lives, and see how long they last. Danica Chen's lab at UC Berkeley did exactly that with 20-month-old female mice, roughly the equivalent of a woman in her sixties, and published the result in Nature yesterday. The treated animals reached a median lifespan of 834 days against 742 for the ones given saline. That is 92 extra days, about 12 percent.

Living longer is the easy part to measure. The more interesting finding is that the mice aged better. The treated animals kept more muscle strength and coordination, solved mazes faster and held their blood sugar steadier than the controls. When the researchers read their gene expression, the usual signatures of old age, rising inflammation and a fading ability to repair tissue, were muted.

For decades the only reliable way to stretch a mammal's life in a laboratory has been to feed it less. Calorie restriction works, and almost nobody can stick to it. So the obvious suspicion was that semaglutide is simply dieting by another route, since it suppresses appetite anyway. The Berkeley team tested that head to head: the drug matched the diet on strength and coordination, and beat it on memory and blood sugar. Chen's reading is that "These differences point to the possibility that GLP-1 drugs tap into a biological pathway independent of calorie restriction."

The limits are real. These are mice and not people, every animal in the lifespan arm was female, and that arm of the experiment is still running, so nobody yet knows which approach wins over a full life. What is already true is that a compound sitting in millions of medicine cabinets extended a mammal's life when the treatment began in old age, which is the hard version of the problem and the only version that helps anyone who is already sixty.

Dictate acceptance criteria and reproductions inside Cursor or Warp. Wispr Flow auto-tags file names, preserves syntax, and gives you paste-ready text in seconds. 4x faster than typing.

Try Wispr Flow free