00:00Muse is Meta's new, long-touted, personal AI agent.
00:04It can book a trip, it can make purchases, it can schedule appointments, it can make
00:09phone calls, it can post things for you online.
00:13Pretty much any action that you can take on other apps on your phone, it can do for you.
00:19For the past few years, people have been using generative AI apps like ChatGPT, and for the
00:25most part, what you do with those apps so far is that you ask it questions, and then it'll
00:31return responses for you.
00:32So it's kind of a traditional chatbot experience.
00:35What Muse is doing, and the direction that the whole industry is headed in, is toward
00:40agentic behavior, toward agents taking actions.
00:44That means that the bot is not just going to give you an answer to your query, whether
00:50that's text, generating a picture or a video, it's actually going to go over to another
00:55app and take an action on your behalf.
00:59OpenAI with ChatGPT and Anthropic with Claude have really been focused heavily on the enterprise
01:05market, so they also have connectors to other apps for their agents, but those tend to be
01:12more work-focused.
01:14Meta is the first to really own the consumer experience, which makes sense for Meta because
01:19it's a social media company, and that's its heritage.
01:23It is a consumer-facing company.
01:25The potential advantages are enormous.
01:28If you can have a kind of casual chat experience with a personal assistant who can then go and
01:35call your dry cleaner for you or send emails without you needing to draft them, that's hugely
01:40useful.
01:41It takes a lot off your plate, and that's basically the sales pitch.
01:45The company's been very heavily marketing the security of its agent experience.
01:50It put a lot of thought into building a kind of secure virtual machine where the agent could
01:56live and giving consumers assurances that when they share information with this agent, it's
02:01going to be held securely.
02:03Behind the scenes, the agent lives in its own secure virtual machine, which is a dedicated
02:09computer in the cloud that Meta maintains, which is designed to keep people's sessions isolated.
02:16So there's where the main agent sits, and that's the model that figures out how to carry out
02:22actions on behalf of the user.
02:24And then separately, there's another model, another agent called a sentinel that keeps tabs
02:30on that first one, and makes sure that it is doing what the user wants it to do.
02:36And then checks for high-risk actions that that agent might be taking.
02:41Like, it might be for something like a purchase.
02:43So before someone actually completes a transaction, they might receive a pop-up in their app.
02:49Then they're given the opportunity to click Allow and confirm that it looks okay.
02:54There's also a separate box for storage of sensitive data, like passwords and other authorization
03:02credentials.
03:03And the idea is that all of those boxes never mix.
03:07And then on the far side of it, coming out of the virtual machine is the connection to
03:13all those external services that people want to use.
03:16The drawback is that it's really risky, and it means that you're handing over a lot of your
03:22digital life, all of your digital life, really, to an agent that is based on technology that
03:28we know to be kind of unreliable, right?
03:30Generative AI is known for hallucinations, making up information that isn't really true,
03:36or for making mistakes or misunderstanding things.
03:40And if you are now entrusting that technology to actually carry out actions on your behalf,
03:46you run a huge risk of errors cropping up.
03:50And then you are maybe on the hook for purchases as a result, or you've sent an email that you
03:55didn't mean to send or taken some other action that you didn't mean to take.
03:59Whether those people keep on using the app and actually trust Meta enough to consolidate
04:06their entire digital lives in this one app and make all their purchases through it and share
04:13more data with the company through it and not really use any other apps directly, that's
04:20a big open question.
