Meta's Muse reached 5 million downloads in 22 days. This puts Muse ahead of ChatGPT, Grok, and Claude at the same point after launch.
I'm one of those downloads. I've been tinkering with it for about a week, and I am kind of impressed. Not because of the tech. It's cool, but any chat tool is just as capable. It's the modality and how proactive it is. That's the novel part.
I started small because there are some horror stories if you allow too much too soon. We'll get to that shortly. That said, I gave it read-only access to my email and my calendar. It also natively has its own browser for research.
With these connections, it started notifying me of things it could see in my email and calendar, without me asking:
- It told me an item I'd been waiting on, which had been on backorder, was available.
- It picked out two warm leads in my inbox and told me to reply first thing in the morning.
- It pulled together the background for an upcoming meeting, so I skipped my notes.
- It started a feed of things I follow: the F1 weekend, my deliveries, bills coming due.
I also tested its research. I asked which Delta Platinum Choice benefit gives me the best value. It read travel forums to see what most people pick and why. Then I gave it a question I'd already answered myself: a camera lens for an upcoming trip. It reached the same conclusions and recommendations I had. So, its research works!
None of this is new. The AI tools I already use will research a question or prep me for a meeting if I ask. The difference is that I didn't ask (well, for everything but my two research tasks).
Now the horror stories. Matt Robb, a Toronto tech reviewer, let Muse handle his Facebook Marketplace listing for a keyboard. According to reporting, it accepted a CA$10 offer on a CA$15 listing, gave the buyer his home address, and replied "Yep, I'm here!" while he was out and the buyer was waiting outside. Robb had picked "Allow Always," expecting Muse to still ask before the bigger decisions. Meta says no privacy control was breached and plans a clearer prompt. Robb disagrees.
The New York Times reports that testers watched Muse change a test user's password without permission. The sources are anonymous insiders, and I haven't seen Meta respond as of the time of this writing.
My prediction: we'll see more and more of these personal agents. Muse already has company. OpenAI launched Dots on September 29. Each Dot gets its own cloud computer and keeps working in the background. Instinct, a startup you reach by text or phone call, raised $1 billion at a $10 billion valuation on September 28. It's still invite-only, although I do have my invite. It's next on my list to tinker with.
I think the value is real, and I plan to keep pushing. I'm not ready to give Muse my credit card numbers. But why not point it at tasks that are tedious, cheap if they go wrong, and involve waiting on someone else? Customer service is the obvious one.
On September 21, Ejaaz (@cryptopunk7213) posted that a seven-hour Delta delay turned into a $250 credit in about five minutes. He says Muse filed the complaint, found him a new flight, and answered the airline's support email. That's one person's account, and I can't verify the credit. His screenshots do show Muse stopping to ask for his mailing address before it submitted Delta's form. I want that pause from an agent.
If you're curious, start where I did: read-only access to email and calendar. Note: training is ON by default. Be sure to switch it off under Settings, then Data controls. See what it surfaces before you give it anything else. Or stay tuned. I will continue to tinker with Instinct, Dots, and others. This newsletter will chronicle the journey should other interesting, delightful, or horrific things occur.
So what are you testing right now? What's working, and what isn't? I would love to hear the tasks that you gave a personal AI agent and how it turned out.
Until next time—keep tinkering, keep prompting, keep innovating.
📩 Not subscribed yet? Hit the button at the top. You won't want to miss what's next.
