I tested Meta's Muse new AI agent powered by Muse Spark - hands on review of Zuck's free AI agent

Vídeo em Inglês, publicado pelo canal Alex Volkov from ThursdAI.

Video transcript

Transcript from available captions; it may contain errors.

Hey, Meta. Yes. Meta just launched a personal AI assistant called Muse hours. I think this may be the most polished version of the AI employee 24 hour agentic AI idea that we've seen so far. This is not a paid promotion. Meta did not sponsor this video and I was generally curious, would people trust Meta enough to give it their Gmail information, contact information? After all, a helpful AI, assistant is an assistant that has access to all of your stuff. So I got access and I started throwing real work at it. And then this happened. I gave Muse and another AI assistant you can see here Instinct the same task. Get me tickets to the Rosh Hashanah dinner this Friday. Muse found the event, filled everything out in the forms, asked me to approve the payment via stripe link, bought the tickets, and finished the task before Instinct even answered. This piqued my interest, so I decided to tell you all about this. In this video I'm going to show you what Musi actually is, how fast it is, how to set it up, and we're going to talk about the security and privacy that Meta is trying to bake into this, including recruiting the founder of Signal to work on the end to end encrypted, vm. I'm going to also show you how it's different than open client Hermes that you may heard about and Grogbot, the latest one from Xai and Instinct, obviously, as I showed you already. And three ways to turn it into a generally useful personal assistant today and stay until the end because there's an iPhone connector inside this that nobody else is talking about yet, including connecting to your health, your home and somehow Bluetooth. And if you are new to this channel, my name is Alex Volkov. I'm an AI evangelist with weights and biases from coreweave and I run weekly AI shows called Thursday I and I've tested every possible agentic AI system out there and this might be the more exciting one. Also, it is free. Let's get into it. All right, so the setup is. It could not be simpler. You go to Muse AI and you log in with your meta account or Instagram account. That's it. And then you set up connectors. I wish I could show you, but I'm already logged in here. So let me just walk you through the UI and show you what you're going to see. So you have this main chat, which, by the way, here you see a document that it prepared for me with the script for this video that I'm completely off the script right now. I'm just showing you the ui. There's. I don't need a script for that. Okay. On the right side, you can see your little assistant. Yours is going to look a little bit different than mine. You're going to say Muse. but you can ask it to change the avatar to a, you know, glowing blob of sparkles. I don't know why anybody will do that, but I will do that just for you just to see. And you can see that something is happening. You see this little video, this video it generated for me after it generated the avatar. So you can see it doing different things, which is beautiful. and also you can set up the Persona for your assistant as well. you click Identity, you can see mine is Wilfred sharp, warm, a little dapper monocle, etc. Here's our examples of the Glowing blob. Do I want the glowing blob? Sure, let's. Let's click this one and let's see what it's going to do. Right. So my assistant now is Wilford the Glowing Blob. And I find it very interesting that they will generate a video for you when stuff works. Also, you have access to the Soul MD and memory of Data system. I'm not going to show you my memory because a lot of personal information, identifiable information about me is in there. I'll show you the SOL system. thinks. You can add a bunch of stuff in there. Steinberger. So if you want your system to be a little nicer, a little stricter, you want it to be concise. You can just add it here and then save this file. And I find that that's pretty cool. Right? So once you've set up your avatar, basically the next thing that I recommend you to do is connect WhatsApp. I don't know if you use WhatsApp, but you definitely should for this. Once you have WhatsApp, you'll have this like, specific chat where you can text with your agent from WhatsApp and you can ask it to do stuff. So you have the main chat, WhatsApp, and then you have something called a side chat. You just open a new side chat and say hello. You can see how this is very, very fast. pull up Thursday. I tell me what the last episode was about. You can see that the browsing that it does behind the scenes. And here it will say searching, searching sources, is happening fairly quickly. In addition to just having a browser and just going to URLs, this agent actually has a full virtual environment. And here you can see that It looked up, our episode very quickly. this used the browser, but also it has a real machine in there. So you can can ask it to install stuff. And I'll tell you all about the interesting things that I've installed in mine. Other connectors you should go and take a look at settings. you have connectors like obviously Facebook comes pre baked, but a connector like Gmail, Google Calendar. According to whether or not you will trust meta AI with this. And we can talk about trust in a bit. having an AI assistant look through your email and sort your inboxes is the very basic thing. The other things that they have are specific connectors that nobody else has. For example, Instagram messages. Here's my Instagram. Feel free to follow me. and you can connect a bunch more. The cool thing is if you are not finding the right connector, you can basically just build one yourself. So for example, I have this connector called Brave Search here. Speaking of connectors, the payment flow is one of the strongest parts of Muse because Muse has a native integration with Stripe link on lunch day. This means that your assistant does not need your raw credit card number sitting in a prompt or pasted in the browser. It can prepare a scoped purchase and then ask you to review and approve it. I've tried other versions of this agentic payment. I've also tried specifically Stripe Link with Hermes and openclaw and Grogbot. I always had issues. This was easily the smoothest one that I ever used. I actually want to show you how this works. Alex Wang also announced an agentic payment protection, including a refund guarantee if Muse makes a mistake. I would read the actual coverage terms before thinking of this is an unlimited insurance. I've not triggered a claim yet, but the direction is right. If companies want agents to transact, they also need the answer who carries the risk when agents get it wrong? This is the first company I ever heard of, of taking that risk on themselves. So let's actually show how this looks. Agent suggested that I can do a Wikipedia donation. So let's do it. All right. I want to donate to the wikimedia.org $5. Let's do this via Stripe link. I, as I recommended before, I suggest narrating, everything that you do is going to open the browser, open the Wikipedia domination, and then bring me the final review card. You can see it's going to be really quick. The most important thing about that card is that they don't get access to my actual card. It's going to fill out the details and I will be able to just approve and $5 to Wikipedia. And to show you how the agentic process works, the cool thing about stripe specifically is that you on your phone will get a notification, which I think just came through on Stripe link. So you can approve this from far away. All right, the, final review Wikimedia Foundation. And you can see this very nice UI. I asked it for $5 and for some reason the estimated total is $10. It places temporary hold of total plus up to 5, but only $5 is actually charged. Okay, let's see. I approved the request and now we've donated, $5 to Wikimedia Foundation. One of the differences that I saw that Muse has on top of every other agent bot is the iPhone connector. So do not miss on those. If you only set up connectors to the web, you won't get access to this if you have an iPhone. Not sure about the Android, but they do have an Android version. Go, to connectors. In there. You'll see native iOS connectors, stuff like Calendar, which you can connect through Gmail, but this one is your iPhone. Calendar, contacts is very important so that your, agent will have access to everybody that you know. I think if you're using WhatsApp, they already have access to your contacts, so that's okay. Reminders, health and home are. Those are very interesting ones. Home I haven't seen yet. So you can control your smart home via your agent. And there's another one called Bluetooth, which apparently it can literally just use your phone's Bluetooth To do some stuff around your home. That's completely up to you. But I just thought that this was novel and I wanted to bring this, and to show you is, this is possible. I want to be very precise. I haven't tested every connector that's listed here, and know, seeing a connector is not a proof that every imaginable automation works. But the list shows you where meta is going about this. Your agent needs to know you it needs to know, you know, your health. You connect the Apple health. It needs to know your reminders. It needs to be able to set your reminders as well. It it's a digital assistant that's focused around you. This can become extremely useful. It also raises the stakes on privacy. So let's talk about that uncomfortable part. So should you trust Meta with more of your life? That's the biggest question, right? I immediately saw a bunch of texts Facebook saying if there's one company that you don't want an agentic AI of, that's Meta. And, from the other side, I saw a bunch of evidence that they're taking this very, very, very seriously. So let's take a look. Stripes, David Singleton, who leads product at Meta Superintelligence Labs and was previously the CTO of Stripe, laid out the security architecture on the launch Day thread. And they posted a very deep dive into the security aspect of this. Zuck. Even Mark Zuckerberg even went further on his interview with Alex Heath and called privacy a major differentiator for Muse. So they're not only saying, we're like the other ones, we collect your data, they're saying, this is differentiate. We're doing that better. First of all, their prompt injection defense is layered. What does this mean? This means that if somebody is trying to influence your agent via sending you an email or a website that it would visit. Meta is trained to recognize and resist malicious instructions, including Including a separate process they called Sentinel, this. Okay, what does this mean? This means essentially that the LLM brain that they probably using new Spark 1.3 or whatever, does not have access to your information and only when the information leaves. When it's about to do an external request, only then they inject the tokens. It's very smart, actually. Payments and logins require human approval, and email connectors start as read only. Zuckerberg described the principle as least privilege. that is a much better design than giving an autonomous model, broad access on day one and hoping that it behaves. You still need to choose connectors one at a time, as I showed you. And you can limit permissions and can disconnect them. Also a very important thing, you can opt out for having your data to train future meta models. That's an important thing to do on day one. Let me show you how to do this. You go to settings, you go to data controls and you uncheck this button that says help improve REI models. Very important to do so. Okay. David Singleton says that the new Muse bug bounty pays up to 300,000 for valid reports, including up to 130,000 for a successful prompt injection attack. So if you think that you actually this kind of insurance in a way as well, if you got prompt injected, you can prove it. You'll get $130,000 maybe. the company is separately announced 1Password integration. So Muse will eventually be able to work with logins you already have without pasting passwords into chat, which is super useful. That integration is coming soon. I haven't seen it yet. it's not live when I checked. And then there's Muse Confidential vm. I really want to talk about this one almost as a new chapter. All right, so in this very long blog research Meta AI on the security and safety for AI agents. They go very deep into the least privileged principle, into defense in depth, etc, and how they use the browser and what they use inside the browser, whether or not the browser is yours. But I think the highlight there is this Meta Confidential VM before VM standing for virtual machine. They're saying that we believe people should also be able to choose a fully private mode for personal agents where even Meta or any other service provider cannot see or grant access to your information. That's why we're investing significantly in the Muse Confidential VM and plan to deliver this capability later this year. In an interview with Alex Heath, Zuckerberg specifically said that they've recruited Moxie Marlinspike, the co founder of Signal and the guy who they previously got help from to create the end to end encryption inside WhatsApp in order to build this very secure system. I find it very incredibly exciting that they're working towards this. The one thing that we should say to this meta and they will make money and this is a product and when they say the product is free, they say that you are the product. So they are claiming that some of the stuff that they're used are going to get used towards ad. So, for example, they're saying here, here we go. They're saying that Meta doesn't share a conversation or the data in your virtual machine. With Meta ad systems, however, there are some legitimate scenarios. Legitimate scenarios. And that's up to you to decide, dear reader. if they're legitimate with. When Muse browses the Internet, it will appear as your activity, so that if you ask Muse to buy a shirt from a clothing designer, that designer will be able to retarget you and send you ads. Okay. Instagram, similarly, if Muse makes a restaurant reservations for you or help you find great product on Facebook Marketplace, that may also indirectly influence the ads that you see. Obviously, Meta needs to make money here, but hey, I think the trade off is almost worth it, especially if they launch the this very specific secure VM that I'm very excited about. What about you? Leave me a comment, please, and tell me whether or not you'd be able to trust somebody like Zuckerberg after the privacy issues from before. If you're going to be able to trust and, connect your Gmail account, for example, for an agent like this. According to some reporting Meta says it delayed the release to improve the safety and of this product. But no serious agent company can honestly promise and fully autonomous system that never makes a mistake. That's just not where we are right now. So my rule is simple. Connect the minimum connectors that you need, grant the Navis permissions and keep approval gates for sending and spending until you get the feel of what it can and cannot do for you. security architecture matters for the geeks of us, but your operating behavior also matters. Now here's how to make the most of your muse, Agent. genuinely useful. This applies to other agent, but since, a lot of you will start your agent journey with Muse, I think it's very important. Use case number one. Give Muse a window into X research. Right? X, formerly known as Twitter, is the world's most immediate source of information. there's not a native X connector on the list that I saw, but there's a clean workaround for research. It installs the Grox CLI and authenticates via X auth. If you have a X premium that this just works for you, you can paste this URL and say, hey, read this and set this up. That's it. I'm going to give you a QR code. And instead of handling, Muse, my X password. Once connected, Muse can now use Grok's search capabilities to find posts and research what people are saying in real time. This is X search. So not automatic access to your private DMs or notifications, but if your work depends on tracking launches, reactions or, or if you're like me, you just want to know what's going on. This is a very basic setup. Again, install the Grox cli, start the Xai oauth flow and let me handle the browser approval. It will then just give you a URL and a link and that's it. Use case number two, which kind of connects to use case number one. Your agent can become the most powerful research assistant in the world. And to do that There's an amazing skill by Matt van hornkel. Last 30 days. time. All you need to do is connect your agent to this and provide some API keys if you have to. It also works, you know, for free. but if you provide some API keys, it's significantly more, useful. Last 30 days searches, Reddit, X, YouTube, TikTok, Instagram, polymarket and a bunch of other sources for information. And it's really, really useful. week. So shout out to Matt Van Horn. He was on the podcast, by the way. Check it out. And you just basically tell your agent, hey, set up last 30 days skill. And then, like this, and then you give it the URL. You don't even have to. It's going to find it. But okay. And then once Your agent installs it on its own computer, you'd be able to do last 30 days research. One note about last 30 days. You can also restrict it to last 7 days if you absolutely want to know just the news or you can give it a whole year. So it will go to the best topics on Reddit on a specific thing that's very useful for agents to be able to go and find that niche thing that you want. it's really, really good. By the way, there's a tiny thing here where you can ask it to launch this in a. Launch this research in a sub agent in a subagent and then when you click the sidebar you can see the sub agents working here. So you'd be able to actually monitor what they're doing. For example, here subagent verified some stuff and did some things for the X Grok guide. One of the most interesting thing about this don't ask about the topic. Give it the brief. Speak to your agent about what you want to research. The more details you will give it, and the better the agent search. The agentic search will be right. Use case number three. Your, your agent is only as useful as the connectors you give it. As we said. So you can build a personal stops briefing every morning. The cool thing about news agent is that they have a heartbeat call happening every 30 minutes which looks like that you can edit. That's pretty cool. you can just say edit my heartbeat so you can get it to work for you every 15 minutes. For example, free. Harvard basically will wake up and do some stuff for you. You can also schedule things you can ask it for every morning. Give me a briefing over my email in my calendar the best important, most important things that I need to know about. Or you can schedule something like hey, clean my inbox or tell me about the most important things, in my inbox. The folks that matter know that many folks who are new to agents will try to explore this product. And so there's a whole tab called ideas here. Ideas like name a habit you want to break, I'll nudge you before you trigger. Or let me triage your inbox down to only what matters. For example and shopping. The most unique connector that they have is Facebook Marketplace. Which by the way, if you use other agents in Facebook Marketplace, they may block you. But this one is a native connector. So if you want to haggle a price or arrange a pickup or Something you can get your agent to do so or go and find something. let's try to find Mac mini under 500 bucks. Let's see if our friend here is able to go into Facebook marketplace and find me a Mac mini under $500. I gave a little, I got a little thumbs up emoji reaction. So this means that hey, you know the agent is working for me and I'm hunting down Mac mini under 500 bucks. new one starts at 599 so it's looking for refurbished. And I drop you a side by side table plus a sortable page once I verify the prices are live and you can see the actual actions is happening. I think I showed you this before. and it's searching the catalog and you can see the little animation that we have. I brought back the wolf. All right, so while this happens let's explore a few other ideas. Obviously you can come up with anything that you want. If you own a business. What I strongly recommend is you go on the record, just start recording via this button and say hey, I run the business here. Please tell me what you can help me with via automation, via access and tell me what access do you need. And then you just talk about your day to day business, talk about the emails that you get, talk about the meetings that you're having, talk about the suppliers, talk about the processes, talk about the system of record, talk about the HR system, talk about all of these things and then have this assistant suggest what it can actually help you with. And trust me, there's a lot if you haven't used an assistant like this before. I've installed this assistant to many of my friends who run businesses and they're just mind blown about the productivity that can they can achieve when something like this takes care of the grunt of the work. Here's another cool example while we hunt down the Mac Mini is I can get your money back when the price drops. An item, that you bought. This is a silly thing I learned about when I moved to America. If you buy something and then the price drops, you can get some money back. So once you buy some stuff or you can connect to your Gmail or Amazon accounts and have it monitored. If the price drops after you bought it, this will try to get your money back which is very interesting. The important connector about this actually is the plaid connector. There's a finances connector via plaid that you can connect all your bank accounts to in a secure way. ChatGPT also has this by the way. And then your Muse can monitor your transaction history, your incoming and outgoing credit card activity, etc. if you are willing to share this, then something like this will work. And it looks like we're almost there with my Mac Mini. Let's take a look. Let me close the DS panel. here we go. It gave me a nice table. It found me a Mac mini M4 in 495 and put a little star in here. Use like new Apple warranty to 2026 in Denver. that's very interesting. That's actually a very very good deal. a week ago. So that's very interesting. Maybe I'll buy that. look at that. Even this is new folks. I didn't, I didn't see this. It generated a comparison and then put a little Mac Mini in the icon. Let's actually see about this comparison. Okay, so this is a new thing that I haven't planned on recording and I'm reacting to this in the first time as much as you guys are seeing this. Okay, we're going to wait for this to build, but it looks like there's an artifact feature where it's going to build some sort of UI or something and then it's going to actually keep it here in notes. wow. You have to take a look at this. So not only did it went on the research and brought me this table, then it worked. I will admit this worked a little bit. So I skipped forward and it built me this artifact that says your searchable comparison page is ready. Filter by chip, price or warranty. Let's click into this and take a look. Okay, this is actually crazy. crazy. This is a whole website that it built just for that specific thing, with sourceable columns, 16 gigabyte rams. and you can click I'm assuming, and you can see coverage. There's all state coverage for this Mac mini. And then there's a comparison chart with Warranty and under $400. And then the link to the actual, listing. This is really, really cool. There is a way to customize your Muse beyond just the profile picture and give it a name when you go to Settings and you hit data controls. There's a feature here that says import memory here. It will give you a prompt to your other assistants like ChatGPT or Cloud or whatever. You use multiple ones. If you use multiple ones to extract just the information that they know about you, which you can then review and you can paste it here, this is very helpful. For example, if you want this agent to go and fill out forms in your name to know where you live, etc. This is also the place where you turn off the training on your information. this. One word about the LLMs, by the way, Meta Meta is running The Muse Spark 1.3 model behind for this agent and it's very, very capable model, actually very fast, but also very, very capable. And, Meta is very soon coming out with significantly more advanced models that will rival the ones from Anthropic and OpenAI. So if you are feeling that this agent is not yet fully agentic or the intelligence is not there, don't fret. If you are investing in this and you're putting your memories and routines and processes into this, you'll see that this agent can do a lot more coming very, very, very soon. All right. So is Muse the Open Claw killer likely? Open Claw has been quite down for quite a while. Is it better than Hermes? I don't know. Folks who like Hermes like it for the customizability. Did Meta just beat Grogbot? Grogbot is just a recent release from the folks at Cursor and xai and they're doing an incredible job and their approach has multiple bots talking to each other. So that's a completely different thing. It's too early for me to decide honestly about this claim. However, Meta has the distribution. Multiple people will use this just because Meta will shove it into Instagram and WhatsApp and threads and, and on Facebook and not only that, this one is free. It's hard to beat free. They're giving a very generous 100 million tokens budget for free for every user, probably United States only, but rolling out to other countries very, very soon. They, they have the best onboarding and the most interesting connector strategy, they're building this for regular people. It's not a coding assistant. Although it can code, it's fast and the interface makes delegation understandable. Stripe link makes real purchases feel controlled. WhatsApp makes it accessible from remote. And the Meta native connectors give it capability competitors cannot easily reproduce, like Facebook Marketplace. My biggest positive surprise is the speed. My biggest question is whether the safety and privacy architecture holds up once people connect the most sensitive parts of their lives. And also the follow up question is when they're going to release the secure vm. And that's why I would not recommend you to connect everything from day one. Start with one workflow, see what works, fill it out, keep approvals on and expand access only after it runs your trust. But I am generally impressed. I expected a, lunch day demo and what I got was a full working product assistant. That's very surprising coming from Meta that completed a real purchase before a competitor has even answered. This is a meaningful difference and I will use this for other things like this. And the last disclosure, the research for this video, the structure and the first draft are you hearing right now were all produced with Muse. You can see the draft here, you can see the final wording, you can see the fact check sources which I'm gonna add to the show notes. So yes, the assistant reviewed itself and helped me write the script. If you want, I, make a follow up later with a full week of using it and show what broke what I kept, how much the cost was, and whether those iPhone connectors actually as wild as they look. Please subscribe to thursdayi News to see my full recap after a week. And you know, hit that like button if you want more. See you on the next one.

Assistir ao vídeo original no YouTube

Assuntos: Agentes de IA, Assistente pessoal, Muse. Publicado no YouTube em 2026-09-09.

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