Your Muse AI Agent Warning [The AmberMac Show Ep089]

Vídeo em Inglês, publicado pelo canal Amber Mac.

About this content: The summary, key points and topics below are generated automatically from the video transcript. They reflect the content, opinions, claims and predictions presented in the video and do not represent a position, validation or fact-check by IA Dots.

Video summary

The video discusses the emergence of AI agents, focusing on Meta's Muse Assistant, its rapid adoption, core differences from chatbots, and the various concerns surrounding its use.

Key points

• Meta launched its Muse Assistant, which quickly surpassed 10 million downloads in the US and Canada.
• AI agents differ from chatbots by being proactive and capable of executing goals independently across multiple tools.
• Users report issues with Muse, including hallucinations, cascading errors, and automated tasks running without permission.
• Concerns about AI agents include lack of empathy, potential liability risks, privacy, and increased internet traffic and energy consumption.

Topics

AI AgentsMeta Muse AssistantChatbots vs AgentsHallucinationsLiabilityPrivacy and SurveillanceInternet TrafficEthical Concerns

Video transcript

Transcript from available captions; it may contain errors.

Well, it seems we've shifted gears in terms of the availability of AI agents with Metal launching its Muse Assistant in September of 2026, followed by OpenAI's Dots agent a week later and most recently Anthropic's more work focused Claude Duet, which it pointedly labels a partner you can trust. Well, I haven't had a chance to use all of these just yet. I have been playing around with Meta's Muse enough that uh we are going to focus for an entire segment on that topic, especially since Muse is now being used by more than 10 million people in the US and Canada. And unbelievably, it got to that level of downloads three times as fast as Chat GPT did when it launched. So, pretty staggering in terms of the number of people using Muse already. >> Absolutely. And before we get too far into that, I I think it's worth starting off with the definition of what's meant by an AI agent. And frankly, Meta actually provides a pretty concise explanation of this. And they say as follows. An AI agent is an artificial intelligence system designed to process goals, plan steps, and execute actions across tools on a user's behalf. So there are a couple big differences really between agents and the chat bots or other type of AI assistants that have been with us for for a few years now. And the biggest difference is that AI agents are proactive. They can make their decisions independent of further user input uh in order to accomplish the goals that we set out for them. So that's different than a chatbot which while it may be able to complete some tasks does so directly in response to user requests so very reactively and without the power to make further decisions or wander as far a field as a full autonomous AI agent might. >> Yes, exactly. So unlike a chatbot which I think most of us are used to at this point, you actually can give an AI agent a goal and the agent will make a plan for how to achieve it by breaking it down into tasks and subtasks as well. some of which it may assign to its own sub agents and then executing on those tasks so it can reach the goal that you've given it. Now it may come back uh to its user to get some additional information or run through some options for feedback. But if you leave it alone and this is where it differs from a chatbot. It will take whatever action it needs to complete what you've asked of it. And maybe that's the the problem. [laughter] >> Maybe it is. And so some examples, this this could be a person asking their AI agent to find them the cheapest way to see a concert by a band that they really love or or plan and make the purchases necessary for their next vacation. Or or it could be a a company that's asking its agent to design and execute, say, an online marketing campaign or or to review and document computer code on an ongoing basis and implement some improvements as it goes along. And these agents keep learning as they go. They've got a defined persona that you interact with, persistent memory in order to have more relevant information at hand when they're making decisions and a set of tools that they can use to interact with other systems. And all of this is built upon a large language model to allow them to understand, reason, and take action. Now, as as I've already said, I've been trying out Meta's Muse uh for a few things, and I can see what is compelling about the idea, but there are also some serious limitations and concerns that people should be aware of as well. >> Oh, absolutely. And and full disclosure, I can't speak to using Muse directly from experience, although I've done a lot of reading about it. Uh as I wouldn't even consider installing a product like that from Meta on my computer because I I barely trust them with a social media post, so I'm certainly not going to see what they get into on my laptop, but uh uh reasonable people can differ. Um so, but why don't you let people know some of the things you've been asking Muse for help with, Amber, and what kind of results you've been getting? Yeah. So, I I decided just to play around with it a little bit. I I do share your concerns in terms of anything from uh Meta. I think uh I don't trust it uh entirely or maybe at all, but I wanted to get a sense of how it works, especially since I'm keynoting every single week and people are asking me these questions and and to our point of how many people are already using it. I think it's critical to understand it. So, the first task that I I gave Muse, I thought was relatively straightforward. So, I was flying back from a keynote. I was booked to come back uh two days later, but I wanted to come back the next day after the keynote, but the flight was full. So, I assigned Muse the task of keeping an eye on that flight on Air Canada to see if any seats became available. So, very straightforward. I just needed one seat. I did not care where that seat was. It was very straightforward, a direct flight. Uh and I said, "Let me know, even if it's every single hour, the status of what's going on." And so, you know, of course, like all uh agents or chat bots, it came back and reassured me that it was, you know, set on the task. It would do a great job and thank you for the information. And then hours and hours went on to the point that uh later on that night, it was almost 12 hours later. I I finished my work for the day. I checked in with Muse and and I said, "Hey, I haven't heard from you. What's going on with that flight? I'm leaving tomorrow." And it said, "Oh, there there is actually one seat that became available at 10:23 a.m. this morning." And that was kind of the end of it. And I said, 'Well, I said, why? Why didn't you tell me about that? That was the whole point. I said, I'm kind of disappointed. And I think it's important for me just to read the quick reply. So, this is what Muse said after I said I was disappointed. You have every right to be. You handed me one job, watched that flight, and tell you the moment a seat opened. And when the moment came, I didn't tell you. Then I made it worse by confidently telling you the opposite twice when you checked in. You shouldn't have had to chase your own assistant for the thing you'd already asked for. I'm sorry, Amber. No qualifiers on that. >> And this is one thing that I think is just ludicrous about these AI agents are these useless fake apologies. Don't don't anthropomorphize yourself or whatever, pretending you're, oh my gosh, I'm disappointed in myself. It's like you you you're not disappointed. You don't care about any of this stuff. you you obviously didn't do what I asked you to do, but there's not much point in conjuring up some sort of empathetic excuse because you don't you don't have any empathy. You don't you don't feel this way. And frankly, it's just these things use a lot of energy. It's just a waste of energy to go through all the different ways I disappointed you. So, you can find these kinds of apologies all over the place online. And it's it's like this this is we get hallucinations with the chat bots where they're inexplicably conjure up information that doesn't exist or is totally inaccurate. It's that kind of thing. They can have these cascading errors where they'll they'll acting on their own. and they'll make a little mistake and then they accumulate in unexpected ways and they they always come back with this this sincere sounding apology that is is pretty pointless as far as I they should just go error [laughter] because that's about as much useful information that was uh provided in that uh apology to you I think. >> Well, I mean this is the thing. So, I I mentioned I checked in at the end of the day, but like you see from the reply, there was a couple times during the day that I just wrote, "Hey, just checking in." And it was like there's no seats available. And I mean, that was an outright I I guess we would call it a hallucination. I mean, or a lie in terms of the information it had found. And when I actually dug into it a little bit further, Jeff, in terms of asking it like what happened and it explained that it went out um and it got the information and it found out there was a seat and then it considered it closed. And then the second part was the failure that it forgot to actually tell me even though that's the only task. And I was joking at a keynote recently that I think even a six-year-old would have done a better job. Like they would have understood the assignment and been able to complete it. >> Well, and this is this is part of the issue with these AI agents. Don't get me wrong, there's there's lots of very cool aspects to them and all this, but but they they do have some major issues. And uh uh so you you've just seen in terms of that hallucination one, they also don't you know they don't have the ethical or moral standards that we have. So they can exercise some really poor judgment because of a lack of human empathy or understanding and do do [clears throat] things that they shouldn't because they're so focused on what their priority is to realizing your goal. And this is kind of like the uh one of the scenarios for uh sort of AI taking over the world where it's like the uh the paperclip problem where a uh a an AI agent is is told to make as many paper clips as they can and they decide that they just can't do that with all these humans in the way and so they kill humanity. So there's there's still this despite all of trying to wrap them in this very human sounding language, they they just don't understand things that way. And uh it's a little bit I think it makes it more frustrating for them to try to pretend that they do. >> Yeah. I mean and and that is the the most frustrating thing. And I I think they they can be really good at one-dimensional tasks. And what I mean by that is another example of how I I use Muse just for a quick little test is I said, "Hey, I'm looking for an electric bike. Can you search on marketplace in uh communities nearby to let me know what's available?" And to be fair, because there are some good things that Muse can do. That was a really good use case scenario. I didn't really need it to do much more, but it gave me the the photos from marketplace. Obviously, that's all within the ecosystem, Jeff. So, I feel it does a better job of that. that's within the Meta environment, of course, because Meta is owned by uh or Facebook is owned by Meta. So, it did do a really good job of that type of thing. So, I can understand why people are using it a lot for marketplace, but we've seen as well how that can also go sideways. >> Yeah. Like there was the guy who discovered that his AI agent that he had let take over his Facebook marketplace sales had given out his his personal address to a stranger and told them to come over and then told them that yeah I'll be right down when when it's like you you can't be right down. [laughter] You're you're an AI agent. You're kind of stuck on the internet. Uh and all of this without the guy knowing. So there there are some issues like that too. And there's liability issues. This this came up uh at uh allin when we were there and I was looking at the uh at an agentic AI uh e-commerce kind of uh talk and you know who who's responsible for for when when something goes wrong and it's one thing when something go and I mean we've seen terrible things go wrong with just chat bots but these AI agents may have access to your like credit card and banking and stuff like that and so it it's there there's going to be some real friction when uh one of these guys starts making purchases you don't want to make and when as we found out with David Fraser last week when they discovered that uh the liability is [snorts] actually pretty clearly on the user's side we do have we use agents for very various things like human agents there's legal ramifications around that that's probably what would apply to these AI agents but the the real question is what's the liability of the AI companies who build these and I I don't think they obviously don't want to have any liability for that but that's going to rear its head I'm sure uh that repeatedly so we'll see how that evolves >> it's interesting that you talk about that from uh the that perspective in terms of focusing on liability because I I couldn't agree more and it reminds me of a comment that I saw on LinkedIn when I posted this interaction with Muse where Philip Shane who is a filmmaker who's been on the show in the past. He said something I thought was really funny. He said, um, "Real businesses offer at least a coupon or something for your trouble." And, uh, he he kind of put an emoji after that because I I thought that was a fascinating look in terms of the liability or even more specifically accountability, right? You know, who's responsible for just doing a terrible job when it comes to these agents? Because they they aren't responsible. It's not like all of a sudden MetaMas Muse is saying, "Hey, you know, I'm going to give you a discount on advertising for your business because I screwed up this task." Nobody's accountable. Nobody cares. >> Well, and speaking of of advertising, another question around these is how do they make money? So, so Meta has a combination of uh at a certain level there's monthly fees, uh there's advertising, uh they've talked about taking a cut of the the payments you make, the purchases that are made through it. And you know, all of that sort of stuff is exactly like Door Dash and these other ridiculous things making everything expensive for all of us. But even if I saw some analysis and I can't remember the exact numbers, but even if Meta were going to, you know, maximize the monthly fees and ads and and those purchase cuts, it's going to not come anywhere close to the cost that it's taken Meta to build and run this thing. There's they're supposed to be doing $200 billion of capital spending next year and there's only about 2.5 million daily active users for for Muse now. Now, now Zuckerberg has said he's preparing to have a billion users. That's a big step up from 2 and a half million. I know it's just started, but a billion is a lot. And it's like, you know, Mark, uh maybe you should be a little more humble when you had to close down your metaverse that you thought would be have exactly the same growth potential. and uh you had to shutter it because people weren't interested. So, it it'll be interesting to see. I mean, they're going to find some way to make money off your data, I guess, but uh a lot of this stuff just doesn't make financial sense yet. >> Well, we've talked a little bit about hallucinations, about cascading errors, liability, data privacy, um and how are they going to make money. Uh I also wanted to just uh get into uh another question which is the ethical or moral standards because one of the things that I have been thinking a lot about is the fact that um there are some things that muse is doing and can do that can have serious implications. So um let's think about it in in two different ways. So the first example in terms of using Muse, I had a chance to ask it to run through Jeff uh all of your your public social media and to find out when you were critical of the um US administration. And [laughter] >> this was fascinating because it found uh more than a hundred posts um and hundreds of replies, specific details, dates, links, and then it said, "I'll keep monitoring this." And as soon as it says something like that, I'm sure we all get chills because you you think about what can happen from the perspective of um surveillance uh with an agent actively watching what someone's doing, both from a government perspective, but even someone who is stalking someone, which most often happens to women, a whole bunch of issues can arise that live in that um that that moral area where we have to ask these really tough questions. >> Yeah. and and beyond that we've talked about uh but in the same sort of spirit we've talked about the environmental damage that AI does due to its higher energy uh use and uh uh there's some sort of rule of thumb where AI agents require about 100 times more computational power than just LLMs alone. But [snorts] another thing that doesn't get considered very directly is what that does to internet traffic. So you've got an agent that's just constantly out there looking at around things at like you know AI speed and uh it is imp over half of internet traffic already and so consider something like if you're looking for a purchase when if a human being goes to shop for a pair of shoes or something they might visit half a dozen websites well when your AI agent goes out to shop for shoes for you they've visit thousands of pages. And so the the the traffic impact of that is is meaningful. It's it and so these are things that you don't directly consider but are sort of going on in the background and the you know to have the perpetual stalker kind of thing like you're talking about. That's that's just a terrifying angle that I hadn't even thought of yet. So uh there's certainly certainly lots of concerns like that. Well, I and and I will say uh I I there is a person online who's been sort of mildly harassing me for a couple of years. So, I do have a dossier on that person that Muse is keeping. So, from that [laughter] perspective, it can be uh useful but also stalkerish as well. The last just quickly I want to mention something too that I've noticed with Muse which is is the most concerning Jeff is that um I've asked it questions just to see how it works and it will go and and want desperately to take action. I'll give you an example. So I was asking I someone I know needs to get a certain type of scan at the hospital and I was saying what is the typical wait time in Toronto for this particular procedure? I got the information. This was three days ago. I woke up this morning. I open Muse. I'm not even talking to Muse whatsoever. And it says, "Hey, I've drafted an email to the hospital." Um, you can either email them with this information or you can call them. Here's the number. Do you want me to help further? So, it it keeps working on tasks that you've not asked for additional help for. This is the second time this has happened to me. And I say back to it, hey, I don't need your help on this. Like, let's let's end that. So, very bizarre behavior as well. >> Yeah. and and behavior that again uses energy and and clogs up the internet and all this sort of stuff. And and you consider things like RBC here in Canada just came out saying that they wouldn't let Muse access bank accounts as it stands now. And I can really understand that. Think of what happens when these AI agents get hacked. If this AI agent has all the information about you, all your contacts, your relatives, your your your your banking information and credit cards, the the hacking potential is is very very disturbing. I mean, it's already disturbing given that what Meta is doing with it, but if you think it's somebody who's actually trying purposely to cause harm to you, uh there are some scary things that we're we're really going to be seeing cropping up, I'm sure, as more and more people use these in the future. >> I I totally agree. So, uh it's going to be fascinating to see how these agents continue to evolve, and we will be right here covering it, uh along the way. We'd also love to hear from you. Are you excited? are you scared about AI agents as well as your experiences with them? So, please let us know on social media. Share some comments with us and uh thank you uh in advance uh for any of your experiences and and we'll share them on the show. Coming up next, we're going to dive into a different topic where we explore how artificial intelligence is changing the way that people find love online.

Assistir ao vídeo original no YouTube

Assuntos: Agentes de IA, Muse. Publicado no YouTube em 2026-10-27.

Seleção automática por assunto e idioma. O vídeo pertence ao canal de origem.