Why I’m Keeping Meta’s Muse on a Short Leash
After returning from a mostly internet-free vacation a little over a week ago, I couldn’t believe the online buzz about Meta Platform’s new agent app, Muse. I immediately downloaded it and began assigning tasks to it like it was an intern in an era without labor laws.
I created an agent to monitor the internet for reviews of a new album by the artist Beck. I ordered Muse to keep an eye on Facebook Marketplace for obscure stereo equipment. And I used it to create alerts that flagged new SEC filings by companies I want to follow closely. Muse accomplished all of those jobs effortlessly, and using the app was even—dare I say it?—delightful.
These were all fairly banal assignments that I could have accomplished pretty easily without Muse. If I wanted more out of the app, I would need to give Muse access to things like my email inbox—which it promised to scour to build agendas for the day—and my credit card account, which it could use to flag recurring charges I might have forgotten about.
That’s when I began to wobble. Giving Muse access to some of my most sensitive personal accounts is the kind of thing that would make people like my wife—who is whatever the opposite of “AI-pilled” is—question my sanity. I had already read my colleague Abram Brown’s account of how Muse had double-booked hotel rooms for him. And yesterday another colleague, Jyoti Mann, reported on the discovery of a security vulnerability in Muse that could let a hacker snag personal information.
This is the trust purgatory I predict a lot of people are going to get stuck in when it comes to giving agents the digital keys to run their lives. Overall, I am at the more trustful end of the spectrum when it comes to tech companies. I have a doorbell camera made by Google. I use Amazon for prescription medicines. I figure that the reputational risks for them of violating that trust are great enough so it’s really not in their interest to do so. But the drumbeat of cybersecurity incidents involving AI models, including the OpenAI–Hugging Face intrusion, makes the leap of faith with AI feel especially perilous at this moment.
In the meantime, I continue to conduct what I view as safe experiments with Muse. In fact, I may have discovered what I believe is one of its best use cases.
There are at least a dozen apps on my phone with atrocious user interfaces that haven’t been updated in years. One of them is for a Wi-Fi–enabled thermostat that controls the heat in a section of my house. Could Muse make it easier to use the appalling software for this thermostat?
I gave Muse my login for the web version of the app and ordered it to set a fall and winter heating schedule. It took Muse more than 15 minutes, but ultimately it accomplished the task. I cringe thinking about the electricity that chore required.
Still, I will respect Muse if it can act as my go-between with the ugliest apps on my phone—as long as the risks of doing so are minimal. For the time being, I’ll be keeping Muse on a short leash.