Latest / Now Shipping: A Mind the Product podcast / Meta’s Muse: How much control would you hand over?

Transcript

Meta’s Muse: How much control would you hand over? · Now Shipping: A Mind the Product podcast

• 0:01
  SPEAKER_00: Today's story is all about Muse, Meta's new personal AI agent, the thing that Mark Zuckerberg has been building for over a year now under a name he's been calling personal superintelligence.
• 0:15
  SPEAKER_00: On September 8th, Meta launched Muse, a personal AI agent that is available right now in the US on the web, on iOS, on Android, and inside WhatsApp.
• 0:28
  SPEAKER_00: Now, here's Meta's pitch for Muse.
• 0:30
  SPEAKER_00: Muse isn't a chatbot that you talk to, it's an agent that you actually hand things off to, whether it's drafting an email, booking a flight, buying concert tickets before they sell out, negotiating your car insurance bill, anything really, the list goes on and on.
• 0:47
  SPEAKER_00: And yeah, it's kind of interesting that this social platform releases an agent just a couple weeks after another social platform like X from SpaceX AI.
• 0:58
  SPEAKER_00: They just released Grokbot a couple weeks ago.
• 1:00
  SPEAKER_00: And Meta is releasing a similar product, but it's a little bit different.
• 1:04
  SPEAKER_00: And we'll get into that later.
• 1:05
  SPEAKER_00: But as far as Muse goes, Zuckerberg wants Muse to understand your goals, whether it be your health, your career, your finances, your relationships, and work on them 24-7 in the background.
• 1:17
  SPEAKER_00: The way a good, you know, like chief of household would if you could afford to hire a chief of household for your house.
• 1:25
  SPEAKER_00: I can't.
• 1:26
  SPEAKER_00: And I do have a very busy home life with a 14-year-old and a 10-year-old, and they're in all sorts of sports and activities.
• 1:33
  SPEAKER_00: My wife's super busy with work stuff.
• 1:35
  SPEAKER_00: So maybe Muse is designed for somebody like me.
• 1:38
  SPEAKER_00: Anyway, how does Muse actually work?
• 1:41
  SPEAKER_00: Well, every single user essentially gets their own dedicated virtual machine, and that's all running in Meta's cloud.
• 1:48
  SPEAKER_00: Now, so this is your agent, your data, it's walled off from everybody else's.
• 1:53
  SPEAKER_00: And inside that virtual machine, there's a separate system that Meta calls Sentinel.
• 1:57
  SPEAKER_00: And its whole job is deciding what Muse is allowed to do.
• 2:01
  SPEAKER_00: You could see the agent's browser window while it works.
• 2:04
  SPEAKER_00: You control which apps and which accounts it can touch, and you could cut it off at any time.
• 2:10
  SPEAKER_00: And before it does something like spend your money or send a message as you, it has to come back and ask for permission first.
• 2:18
  SPEAKER_00: At least that's the way it's supposed to work.
• 2:20
  SPEAKER_00: Now, Meta is also promising a quote confidential computing version later this year.
• 2:26
  SPEAKER_00: And this is one where they claim that even Meta itself can't see what's happening inside your virtual machine.
• 2:31
  SPEAKER_00: Now, Zuckerberg personally recruited Moxie Marlin Spike.
• 2:36
  SPEAKER_00: First of all, what a name.
• 2:37
  SPEAKER_00: But that is a real person.
• 2:39
  SPEAKER_00: That's not some character.
• 2:40
  SPEAKER_00: It's actually the guy that built Signal, the encrypted messaging app that a lot of privacy people trust more than any other app on Earth.
• 2:48
  SPEAKER_00: Anyway, he recruited Marlon Spike to work on Muse's privacy.
• 2:52
  SPEAKER_00: And I'd say it's a pretty big statement move.
• 2:54
  SPEAKER_00: I mean, Meta's own credibility on privacy is let's just say it's mixed, but Marlon Spike is a big, credible name here.
• 3:03
  SPEAKER_00: So I'm sure that was a move that Zuckerberg made very intentionally.
• 3:08
  SPEAKER_00: Anyway, underneath all of this is a model that Meta calls Muse Spark, which actually came out in April and it was built by a group inside Meta called Superintelligence Labs.
• 3:20
  SPEAKER_00: Now, in August, just a couple weeks ago, Meta shipped something called Glimmer, Muse Glimmer.
• 3:26
  SPEAKER_00: This was a smaller open weight version of basically the same technology, but the kind that you could run on your own laptop.
• 3:33
  SPEAKER_00: And you might remember we reported on this just a few weeks back.
• 3:36
  SPEAKER_00: So it's interesting to think about the rollout order here.
• 3:38
  SPEAKER_00: I think it tells you something.
• 3:40
  SPEAKER_00: It's like closed model first, open model second, and now finally the consumer product that ties it all together.
• 3:47
  SPEAKER_00: Now, this is interesting.
• 3:49
  SPEAKER_00: Let's talk about pricing.
• 3:50
  SPEAKER_00: So Muse is actually free for what most of a regular person would use it for.
• 3:56
  SPEAKER_00: Meta said that's a hundred million tokens a week totally on the house from Meta.
• 4:02
  SPEAKER_00: Now that's a huge deal.
• 4:03
  SPEAKER_00: That's something different than what a lot of these other agentic companies are offering.
• 4:08
  SPEAKER_00: Now, there are paid plans.
• 4:09
  SPEAKER_00: Above that, there's a power tier plan for$20 a month and a maximum tier at$100 a month for people that are really, really running it hard.
• 4:17
  SPEAKER_00: I don't know, maybe they have an entire community of people that they're running it for, not just their own family.
• 4:24
  SPEAKER_00: But Meta's chief AI officer, Alexander Wang, said that the vast majority of users should be able to do what they need to do within that free tier.
• 4:32
  SPEAKER_00: Now, how Meta makes money off of something that's mostly free, Zuckerberg said that Meta will take, in his words, a very small cut of transactions that Muse completes for you.
• 4:44
  SPEAKER_00: So I don't know exactly what that means quite yet.
• 4:47
  SPEAKER_00: I don't know if that means transactions that have to be built into Muse, because Muse should be able to make you know purchases on your behalf.
• 4:56
  SPEAKER_00: I guess we'll find out about that.
• 4:58
  SPEAKER_00: But there's no ads.
• 4:59
  SPEAKER_00: That's not something that Meta's making money off of Muse with.
• 5:02
  SPEAKER_00: This is essentially Meta explicitly trying to build something that pays for itself by helping you, not by selling your attention, which has pretty much been Meta's game to date.
• 5:13
  SPEAKER_00: Now, this part gets pretty interesting because right alongside the launch coverage came another story related to the launch.
• 5:22
  SPEAKER_00: Now, Reuters reported on internal meta employees posting about testing Muse before it shipped, and these posts were written as recent as launch week, and it was not very clean.
• 5:34
  SPEAKER_00: These weren't very rosy posts.
• 5:36
  SPEAKER_00: In one of these posts, an employee said that Muse was asked to identify toys in some birthday party photos, and it bypassed its own safeguards and exposed private iCloud photos that it had no business touching.
• 5:49
  SPEAKER_00: Now, separately, an employee testing Muse to track fast-selling tickets and inventory for an event ran into what was described as, quote, many failure modes that made it unreliable.
• 6:00
  SPEAKER_00: And even Meta's own CTO, Andrew Bosworth, said he kept getting logged out mid-task.
• 6:05
  SPEAKER_00: Now, Meta's VP of AI products, Vishal Shaw, didn't try to spin it away.
• 6:10
  SPEAKER_00: He said, look, it's impossible to say it's never going to make a mistake.
• 6:14
  SPEAKER_00: But he insisted that the team built Muse to be as safe, as secure, and as private as they possibly can make it.
• 6:21
  SPEAKER_00: Now, here's the thing Muse was actually originally supposed to ship back in April alongside Spark, but Meta pushed the launch back specifically to work through security concerns.
• 6:33
  SPEAKER_00: So that means the version that shipped this week, even with these issues, the version that made it through this big delay, this is the version that Meta thought was essentially good enough, even though it still has these gaps.
• 6:47
  SPEAKER_00: Now, one report even cited internal data showing security incidents across Meta's AI systems up 40% year over year, with the time engineers spend firefighting those incidents up 70%.
• 6:59
  SPEAKER_00: Now, of course, that's not some sort of headline stat that Meta put in their launch materials.
• 7:05
  SPEAKER_00: Reporters had to go and dig for those numbers.
• 7:07
  SPEAKER_00: But to be fair to Meta here, they didn't just wing the security story.
• 7:11
  SPEAKER_00: I mean, back in the summer, Meta published something that they call the agent's rule of two.
• 7:16
  SPEAKER_00: And the idea is pretty simple, even if the engineering behind it is a little more complicated.
• 7:21
  SPEAKER_00: It means an AI agent should never be allowed to do all three of these things at once: process content from an untrusted source, touch your sensitive data or systems, or take an action that changes something or talks to the outside world.
• 7:37
  SPEAKER_00: Combine all three, and you've built the exact recipe for a prompt injection attack where somebody hides instructions inside of whatever might be a web page, an email, and all of a sudden your helpful agent starts working for them against you.
• 7:52
  SPEAKER_00: So it's an interesting framework.
• 7:55
  SPEAKER_00: I will say, by meta's own admission, though, it's not a complete solution.
• 7:59
  SPEAKER_00: It's just one layer, and it needs human approval gates stacked on top of it.
• 8:04
  SPEAKER_00: The human approval gate problem isn't hypothetical.
• 8:07
  SPEAKER_00: Back in February, we heard stories about OpenClaw, which of course isn't Meta's, but anyway, open claw went haywire on somebody's personal machine.
• 8:15
  SPEAKER_00: It was told to suggest email deletions pending approval, but instead started deleting messages on its own, ignoring the stop commands until the person had to physically walk over to the computer to kill it.
• 8:26
  SPEAKER_00: The person on the other end of that was Summer U, director of alignment at Meta's own superintelligent safety lab, the lab adjacent to the team that built the safety systems inside Muse.
• 8:39
  SPEAKER_00: Now, of course, there are only these reported incidents with Meta and Muse so far, but I don't know that Meta should exactly get the benefit of the doubt.
• 8:49
  SPEAKER_00: I mean, this is a company that spent the better part of a decade being under fire for concerns about what it does with your data, who it shares it with, and what it knows about you.
• 8:57
  SPEAKER_00: So when that same company turns around and asks to connect your inbox to your calendar to have a credit card of yours put on file so an autonomous agent can help you, I think it's right to have a little bit of pause.
• 9:13
  SPEAKER_00: But anyway, let's step back for a second because Muse isn't launching this into some empty playing field.
• 9:19
  SPEAKER_00: It's really launching into a space that has seen a lot of activity lately, including some products that have already gone down, right?
• 9:27
  SPEAKER_00: OpenAI's standalone AI browser Atlas is effectively dead.
• 9:31
  SPEAKER_00: It's folded back into the regular Chat GPT app in August after failing to get traction.
• 9:36
  SPEAKER_00: Perplexities Comet is the one standalone AI browser that's still actually growing.
• 9:41
  SPEAKER_00: It's backed by Claude under the hood.
• 9:43
  SPEAKER_00: It does cited real-time research, and it just landed a deal to sit inside of Samsung's phone browser, which is real distribution.
• 9:51
  SPEAKER_00: And Google's Gemini agent now lives inside of Chrome, already the most used browser on earth with deep hooks in your Gmail calendar and drive.
• 10:02
  SPEAKER_00: So there's a lot going on in the space.
• 10:04
  SPEAKER_00: Now, we did also talk about Grokbot a couple of weeks ago.
• 10:08
  SPEAKER_00: And you might be wondering, okay, well, how is this different than Grokbot?
• 10:12
  SPEAKER_00: Now, Grokbot from SpaceX AI, it gives you your own persistent, always-on virtual machine in the cloud.
• 10:19
  SPEAKER_00: You can spin up multiple bots that coordinate with each other in essentially a group chat working through tasks 24-7, even when your laptop's closed.
• 10:27
  SPEAKER_00: Now, Grokbot actually lets you screen record yourself doing a workflow once, and then it turns that recording into a repeatable skill.
• 10:36
  SPEAKER_00: It's shipped with 220 pre-built integrations on day one.
• 10:40
  SPEAKER_00: That includes Slack, Microsoft 365, Salesforce, Notion.
• 10:45
  SPEAKER_00: Now, if you notice, all of those connectors, it's pretty much aiming at knowledge work.
• 10:51
  SPEAKER_00: Not necessarily personal errands for you.
• 10:53
  SPEAKER_00: Just last week, Grokbot did expand to Android, landing pretty much in the same moment that Muse did, which is probably why the two are sort of getting lumped together in people's heads right now.
• 11:04
  SPEAKER_00: Now, Muse, again, is free for most people with paid tiers, again, at$20 up to$100 a month.
• 11:12
  SPEAKER_00: Grokbot, there's no free tier.
• 11:14
  SPEAKER_00: I mean, it started kind of expensive though, comparatively, but now you can access Grokbot for as little as$20 a month through Cursor Pro.
• 11:23
  SPEAKER_00: So the pricing has come down.
• 11:24
  SPEAKER_00: So if you did want to use it personally, I think it's priced at a point where you can do that.
• 11:29
  SPEAKER_00: Now, in terms of the focus, Meta seems to be building for mass consumer adoption.
• 11:34
  SPEAKER_00: Grokbot seems to be building for power users and the enterprise, which which does track.
• 11:40
  SPEAKER_00: I mean, SpaceX AI has been posting steep losses, and I might need a business line that isn't necessarily giving away a lot of compute for free.
• 11:49
  SPEAKER_00: So, bottom line, if you're trying to think of all right, well, should I get Muse or Grokbot, might be depending on what you actually want to use it for.
• 11:57
  SPEAKER_00: One seems to be trying to be your personal assistant, the other sort of your always on work computer.
• 12:04
  SPEAKER_00: But what does this all mean for you as a product person?
• 12:07
  SPEAKER_00: I think there's a few things to pull from this.
• 12:09
  SPEAKER_00: First of all, saying we shipped it and this product is super safe, it's become very apparent that these are not the same things.
• 12:18
  SPEAKER_00: And it's almost becoming normalized that these types of vagentic products are being shipped with a lot of holes in them.
• 12:26
  SPEAKER_00: And I guess people are okay with them because people are using them right out of the gate.
• 12:30
  SPEAKER_00: I mean, Meta here is a company with basically unlimited resources, and they delayed this launch once already, specifically because of security.
• 12:40
  SPEAKER_00: They published their own security frameworks for agents, and yet they still shipped a version that internal testers found leaking private photos, locking people out mid-task.
• 12:50
  SPEAKER_00: So I don't know if that's the bar at Meta.
• 12:52
  SPEAKER_00: It might be worth asking where your bar is for using products like this, not just from Meta, but from other companies too, because again, companies like to follow each other.
• 13:03
  SPEAKER_00: I think the second thing is it's an interesting business model here, right?
• 13:07
  SPEAKER_00: Now, Muse is free for most users the way that they would use it.
• 13:12
  SPEAKER_00: Now, they do have, again, paid tiers for power users.
• 13:15
  SPEAKER_00: Now, Meta's actual plan is to make money by taking a cut of the transactions that the agent completes for you, not ads.
• 13:24
  SPEAKER_00: So if that works, this becomes a real alternative model for consumer AI products, one where the agent's incentive is supposed to be aligned with getting you a good outcome, not holding your attention.
• 13:36
  SPEAKER_00: Now, if you're building anything that touches consumer AI, this could be a monetization pattern to start studying right now.
• 13:44
  SPEAKER_00: And it might end up being the playbook for the future.
• 13:46
  SPEAKER_00: Now, the third thing, the question of where the agent lives, I think this is another important question that product people need to think of.
• 13:54
  SPEAKER_00: There are three different things that we've seen this year: a browser as an agent from Google and Perplexity, a standalone app that essentially has died from open AI, and now messaging as an agent, essentially from Meta.
• 14:07
  SPEAKER_00: So if you're deciding where your own products, AI features should live, whether it's a chat panel, browser extension inside of an app that people use all day, this is sort of the closest thing that you might get to a live experiment with three very well-funded companies making very different bets in real time.
• 14:25
  SPEAKER_00: So it's important to be paying attention to what's actually working well here in the market.
• 14:30
  SPEAKER_00: Now, I also think we're gonna start to see people make real product comparisons on permissions, which sort of used to just be a footnote.
• 14:39
  SPEAKER_00: Sentinel, the dedicated virtual machine from Meta, it's essentially the approval gate before anything gets sent or money gets spent.
• 14:47
  SPEAKER_00: This is Meta's entire pitch for why you should trust Muse with your inbox and your credit card.
• 14:53
  SPEAKER_00: When you're evaluating any agent for your own team or for your own personal life, ask the same question that Meta's trying to answer here.
• 15:00
  SPEAKER_00: What exactly can this thing touch?
• 15:03
  SPEAKER_00: Who approves it?
• 15:04
  SPEAKER_00: When does it approve it?
• 15:05
  SPEAKER_00: And what happens when it gets it wrong?
• 15:09
  SPEAKER_00: Because let's face it, there is always gonna be a first time where the agent will get it wrong.
• 15:14
  SPEAKER_00: So I guess here's my question for you.
• 15:18
  SPEAKER_00: Are you ready to hand Muse your calendar, your inbox, and maybe most importantly, your credit card?
• 15:25
  SPEAKER_00: Knowing what you know now about how Muse performed with testing, knowing Meta's track record, I'm curious where you land on that.
• 15:33
  SPEAKER_00: Maybe leave it in the comments.
• 15:34
  SPEAKER_00: I want to hear from you.
• 15:36
  SPEAKER_00: For me personally, I want to experiment with everything, so I'm probably gonna do it, but I don't know.
• 15:40
  SPEAKER_00: Maybe I'm gonna regret that in the future.
• 15:43
  SPEAKER_00: I guess we'll have to see.
• 15:44
  SPEAKER_00: But that's gonna do it for this week's story.
• 15:47
  SPEAKER_00: If you found this story valuable, please tell a friend about the show.
• 15:52
  SPEAKER_00: Leave a comment below.
• 15:53
  SPEAKER_00: I will be right here next week with more AI stories that matter to you as a product person.
• 16:00
  SPEAKER_00: Once again, my name is Mike Belsito, and from the team at Mind the Product, this is now shipping.