Hey, AI Enthusiast!

Welcome back to the #1 AI newsletter on the planet!

Here’s a glimpse into what we have today:

Smartwatches

Your Arteries Have an Age

Your arteries have an age, and it doesn’t always match your birthday.

Stiff, older blood vessels change the shape of every heartbeat, and that shape shows up in the pulse reading on your wrist.

Researchers at Peking University and OPPO, the Chinese phone and smartwatch maker, taught an AI to read it. They fed it millions of recordings from more than half a million people wearing OPPO watches.

Now it looks at a pulse and estimates how old your arteries are. The interesting part is when the estimate comes out older than you are.

Those people tended to have stiffer arteries. And in a separate group, the quarter whose arteries read oldest had about 4x the odds of high blood pressure compared to the quarter that read youngest.

If that sounds familiar, Apple Watch already flags possible high blood pressure. This goes a step further. It puts an age on your arteries.

Here’s why the gap matters.

In an earlier study of more than 200,000 people, the same Peking University team found arteries reading more than 9 years too old came with more than double the risk of heart attack, stroke, heart failure, or death.

Checking more than once matters as well. Averaging several readings made the estimate steadier, and in that same study, people whose arteries read too old at both of their checkups faced more than 5x that risk.

Our take: One reading that runs old is a hint. Two is a pattern. A checkup gives you one reading a year. Your watch checks your pulse every day.

The catch: The smartwatch study is a preprint that hasn’t been peer reviewed yet, and OPPO sells the watches. The earlier study used a clinic device, not a watch.

Together with MongoDB

One Server, All Deployments

Your coding agent has read every file in your project. It’s never seen your database, and it writes queries against collection and field names it invented.

The MongoDB MCP Server hands the agent your real schema.

Collection schemas, configs, and sample data land in the tools you already use, and the code that comes back matches what’s in there.

Claude Code, Copilot, Cursor, Codex, and Gemini CLI all connect.

Local, Atlas, and Enterprise sit behind one config. Run it yourself, or let Atlas host it.

You stay in one window, and the answer arrives with your own data behind it.

Agents

Muse Hands Calls to Humans

Meta’s hottest AI agent has a new trick. Sometimes it passes the phone to a person.

Muse, Meta’s personal AI agent, has topped U.S. download charts since it launched two weeks ago. It sends emails, shops, books travel, and now phones U.S. businesses for you, booking a haircut or checking whether a store has something in stock.

Here’s the twist.

Reuters reportedly saw internal posts showing Meta has been testing a “human concierge” behind those calls. Muse hands some requests to a trained contractor, who places the call and handles it from there.

Half of Meta’s employees have it switched on, with an opt-out for anyone who’d rather not. Some staff worry about personal details landing with call-center workers.

Meta says the feedback has been overwhelmingly positive, and it won’t roll this out publicly without proper disclosures.

Facebook has been here before, as well. A decade ago it tried an assistant called M, and reports estimated humans handled about 70% of its tasks.

Muse had a second confession this week. Meta’s Nat Friedman said the app was “heavily inspired” by OpenClaw, the open-source AI agent, though he says Meta built it from scratch.

Our take: Filling gaps with people while the AI learns isn’t the problem. Keeping the person on the other end of the line in the dark would be.

Notable AIs

Notable AI Tools

🧠 CXL’s AI Native Marketer is a six-week live cohort that teaches marketers how to build 6 AI workflows from their own work and keep them running. Get 50% off until Sep 28 when enrollment closes.

🧩 Attio puts agents inside your CRM, watching buying signals and moving deals forward while you’re away from your desk.

💬 Pounce watches X and Reddit around the clock, surfaces the posts worth answering, and drafts replies that sound like you so that you can reply first.

Math

Mathematicians Get a Say at OpenAI

OpenAI’s newest internal model has been busy.

Since training started on August 28, OpenAI says it’s cracked more than 100 long-standing open problems across most areas of math, fast enough to surprise the company’s own mathematicians.

Not everyone’s cheering.

Earlier this month, 25 winners of the Fields Medal, math’s top prize, signed an open letter pushing back on AI labs treating famous open problems like a scoreboard for their models.

OpenAI’s answer is a group of nine mathematicians who’ll advise it independently, hosted at the Institute for Advanced Study in Princeton.

Names include Timothy Gowers, Edward Witten, Martin Hairer, and Melanie Matchett Wood.

They’ll help judge which results matter, when and how those results get shared, and whether the work meets the field’s standards.

They’ll also weigh in on how OpenAI’s tools fit into math research and learning.

And they get real room to push back. OpenAI isn’t paying them, they pick their own members, and they’re free to go public with advice nobody asked for, including on what OpenAI is doing to math.

Our take: This is the right template for any field AI starts racing through. Experts shouldn’t learn about breakthroughs in their own field from a press release.

The catch: The group advises. It has no say over how fast OpenAI pushes its math research.

Coming in Hot

AI Tools of the Day

Open Models

China’s Open Models May Be a Phase

If you build on open AI models, the kind you download and run yourself, odds are some of them came from China.

a16z estimates Chinese models make up about 80% of the open models startups use.

David Cheng, who spent a decade investing at DCM, went to Beijing and Shanghai for a week to find out why.

The word he heard in every room was “involution,” a Chinese term for effort that keeps rising while nobody gets ahead.

Hundreds of labs undercut each other, poach each other’s researchers, and ship every month.

In that fight, giving a model away is how a lab that’s behind gets noticed. One university researcher told him openness pays while you’re behind, and the moment you lead, you close.

ByteDance, which already has a huge audience in its Doubao app, keeps its top model closed.

But the shift has started.

Zhipu delayed releasing its latest model, GLM 5.3, and Kimi and MiniMax have started adding strings to their licenses for bigger companies.

Our take: Treat an open model as a snapshot, not a promise. Download the ones you depend on and own your setup.

Interesting AI

Claude Paints Pixel by Pixel

Every painting here started as a blank grid of pixels and a lot of Python.

Jake Eaton at Anthropic has spent months asking Claude to paint with code alone.

Opus 5.5 writes programs that place each pixel itself, about 7,500 lines of standard Python built to imitate different brush styles.

The agents never see a reference picture. They work only from what they know about each painter.

Every brushstroke you see is math. It’s painting by description, one pixel at a time.

Offline AI

A Survival Computer With AI Inside

The internet goes down, and every answer you’d normally look up goes with it. Project NOMAD is built for that moment.

It’s a “survival computer” you set up on your own hardware.

Inside sit offline Wikipedia, medical references, thousands of books, Khan Academy courses, and downloadable regional maps.

An optional AI assistant runs on the machine itself, and it searches any documents you give it.

A one-click app shelf adds extras like a PDF toolkit, an e-book library, and a password manager, and none of it needs a connection.

It’s caught on fast. The project has passed 38,000 stars on GitHub, where developers bookmark projects they like, with more than 500 added in a single day this week.

Our take: Local AI usually gets pitched on privacy. NOMAD makes a better case, an assistant that still answers when nothing else does.

The catch: Setup means typing commands on a Linux machine, so this one’s for tinkerers first.

Prompt of the Day

Expert in Fifteen

Important Note: You’ll need to click the button to get the complete prompt.

That’s all for this Wednesday!

Thanks for reading,

— There’s An AI For That