I Used to Hate Meta: Muse Wonder Changed My Mind

Ejaaz I don't know how he did it, but Mark Zuckerberg and Meta might have just pulled off the biggest comeback in AI ever. They just released their new model called Muse Wonder. It's basically the largest AI model that they've ever released, and it competes directly with Claude Opus 5.5, GPT-6 Astra, and Fable 5.1 from Anthropic, basically the most powerful AI models that exist today.

Ejaaz And this is crazy because 18 months ago, Meta's models would come in last place pretty much every single time for video generation, image generation, general LLM context. It just sucked overall. They also torched billions of dollars on fire, spent $25 billion to hire one 29-year-old individual to head up their new meta superintelligence labs, and it just seemed like Zuck had kind of lost the plot.

Ejaaz But with this new model, codenamed Watermelon, now officially called Wonder, they have a real competitor to compete against Anthropic in open air. You see, the AI race up until yesterday was pretty much a two-horsed one and now there's a third major player in the works. And the last thing I'll say is Meta's entire AI strategy seems to be very different from that of Anthropic and OpenAI.

Ejaaz They seem to be focusing just on personal agents and I think that's a much larger advantage than I actually gave it credit for and this is coming from a former hater of Meta products now actually someone who enjoys using them. So I'm going to get into all of this and more the brand new models on today's episode of in the loop brought to you by qualcomm.

Ejaaz Okay, so let's start with the model itself. Meta actually released two models, but it's both called Muse Wander. But there's two varying degrees of effort across both of these models. So the standard one is called Muse Wander. And this is basically their Claude Opus 5.5 competitor. And it actually comes in at exactly the same price.

Ejaaz It's $4 per million tokens input in, $10 per million tokens output. So it's very clear that they're trying to compete with Opus 5.5. And it's actually quite good. But before I get into the benchmarks, the second model is their max model, their Fable 5.1 model, their bigger competitor, which is actually kind of behind closed doors right now, but you can access it if you get a $50 subscription.

Ejaaz But the idea here is, can Meta's second model compete directly with Fable 5.1? Well, the answer is a bit of a mixed bag. If you look at this tweet from Aaron Levy, he kind of shows the progression of Meta models over the last year and a half or so. When they started, Meta's models weren't very performative.

Ejaaz We're looking at the Artificial Analysis Intelligence Index, which is basically the best way to assess the general intelligence of a model. Think of it as like the ultimate human exam, but for AI models, they're tested through a mixture of maths, reasoning, coding, literature, creative writing, basically the stuff that humans are tested on when they're kids growing up through high school, college, blah, blah, blah, but for AI models.

Ejaaz Now, back in the day, they used to score pretty low scores. You're seeing on the right-hand side of the screen around 31 to 34. And now the massive leap over a year and a half, leapfrogging often some very prominent names and models, have now put them in second place, just within striking distance of the flagship model from Anthropic Claude Opus 5.5.

Ejaaz And when we look at the benchmarks, at least from the numbers that Meta's showing us, this model seems to be quite good. It's really good on personal superintelligence index. It's pretty good at long-term coding, especially for agents, because this is a model designed for AI agents specifically. It's really good at automation, general work.

Ejaaz But it's also really good at legal stuff. I was looking at this Harvey Lab AA, which basically tests the model's capability to replacing or automating a bunch of lawyering work. And it's more than double the second place, which is Claude Fable 5.1 Max. It's called 13.1. And I'm mentioning this because Zuck specifically said that the models that Metis is putting out Muse Wander, Muse Wander Full, which is their Max version, the entire focus around these models is that they actually are useful to you.

Ejaaz And I mean that specifically, just you. It's not meant to be this kind of generally. Widely available LLM that is weirdly applicable to anyone and everyone. It's trying to optimize for the types of work that the majority of the people face. Now, that's not to say that anyone and everyone in this particular example is a lawyer, but anyone and everyone typically gets involved with some kind of lawyer or contracting or reviewal thing, whether you're signing a new employee contract to join a new company.

Ejaaz Or if you're dealing with a contract with an insurance company or whatever that might be. And this is just one snippet of what's going on inside Zuck's mind and Meta's entire AI strategy. Something that's also net new for them is DeepSwe 1.1 and coding in general. This model is a major leap up from its predecessor. Now, is it as good at coding as Claude Code, Fable 5.1, as GPT-6 Astra Codex? Would you or should you drop your subscriptions there and start using Meta?

Ejaaz Some people, and this is genuinely a mixed camp, would say, yes, I saw a review from Theo GG, who I love his videos, I love his content in general, and he got early access, and he's been playing around with this for a few weeks. He's strictly extremely disciplined when it comes to just giving his very honest review. He doesn't hold back. And he's historically hated on Meta.

Ejaaz And his comeback to this entire thing was, I hate to say it, but it's actually quite good. Now, it's not quite good in terms of like creating new apps and websites when it comes to coding, but it's good at detecting flaws and security bugs that other models may not have picked out. And he actually showed a few examples where he was like, listen, I tried Claude, I tried.

Ejaaz GPT and it wasn't able to figure out my problems but Meta, Muse, Wonder figured it out for me. So that's enough of an excuse to at least incorporate it into my workflow. I don't need to replace any of my other subscriptions but if this is free and or cheap I'm going to use it. But putting the benchmarks aside there is actually another good reason why you should be using this model.

Ejaaz The first is called active memory. This is basically a reference to the model's capability of pulling the most context about you and carrying it forwards in every single conversation and interaction that you have with the model. Memory and AI models isn't something new. If you look at Chad GPT, who basically were the ones that pioneered this entire personal memory AI assistant type thing, you would be familiar with, okay, cool, my AI knows more about me, I'm more likely to use it.

Ejaaz But what Meta has been able to achieve here with this model is something a little bit more, It's something a little bit more proactive, but also more intelligent in terms of pulling the right bits of memory. You see, it's not as simple as telling an AI model to understand everything it can about you.

Ejaaz It's also about making sure that the models actually know what right bits of information to pull about you. And OpenAI has had all the time in the world to try and figure this out. And yet they haven't been able to put out an agent or a model that I think truly sticks Xen feels loyal to me, if that makes sense.

Ejaaz And Meta's come along and they put out their personal agent Muse a few weeks ago, literally two and a half weeks ago. And they have 6 million downloads so far. And I think around 3 million weekly active users, which of course is only a drop in the ocean when you compare it to the likes of Anthropic and OpenAI.

Ejaaz But the feedback is where the distinction is. Meta has been able to create a very compelling product. As I said before, I have typically hated Meta products and I'm the first to put my hands up here, but it's actually quite a good feature. So the fact that they're bringing this and pairing it with a Fable 5.1 worthy model can only yield, in my opinion, what I could expect to be a much better personal agent experience.

Ejaaz So I'm going to play around with Muse going forwards, but my guess is it's going to get a hell of a lot better. The second thing, though, is the one that I'm most excited for and something that I haven't seen in any other AI lab. Something called active focus, which basically means that if an AI model or if their AI model comes across a challenge or an obstacle, one which it can't typically surmount and it needs to usually come back to you and say, hey, I can't figure this out. Let's.

Ejaaz Kind of approach this in a different way. It just does it itself. Now, you might have heard versions of this from ChatGPT, from Claude. They were only coding specific. That's the difference here. Typically, if you use ChatGPT or Claude and it's doing the relentless thing, you know, it's like trying to fix and solve a problem.

Ejaaz It's usually on coding projects. When you look at Muse Wander specifically, it's for any and every task. If it's like, okay, I think he's trying to figure out a gift for his wife or try to figure out how to plan a specific trip and it comes up against an obstacle like hey i can't get access to this website there's a cloudflare issue or i can't get access to certain bits of information because it's behind a paywall it'll start to intuitively try to figure out okay maybe i'll go to a secondary resource what secondary resource should i go to what's reputable that i know i can rely on and not have to ask.

Ejaaz To kind of verify or check and it starts to become this weird but very useful personal assistant that applies to, and this is the most important part, 99% of the rest of the world. I've explained this in a previous episode, but Meta's models are starting to feel like models that your grandma would love. And I mean that in the most complimentary way ever, because it's just easy and simple to use.

Ejaaz I speak about AI every day, and I think my mom watches the majority of my episodes, and I kind of gauge how good an episode is by the reaction that she gives me, or whether she reaches out and says, wow, I had no idea that blah, blah does blah, blah, blah. Or I didn't know this model did this. I'm going to go use this.

Ejaaz I've never heard so much feedback from her for one single topic. Than MetaMuse. And Meta has this product DNA, something that I haven't seen in any other AI company when it comes to AI, at least since ChatGPT and the rise of Chatbot LLM. So there's something unique here. It's not just about the model, it's about who they're designing it for.

Ejaaz And ActiveFocus is just something that helps them get there. Now, when it comes to pricing and accessibility, they gave API pricing for the standard model. So I think it's $4 in and $20 out which is exact same competitive pricing to opus 5.5 but for the larger fuller model the max version the model that is meant to compete directly with gpt 6.1 astra and fable 5.1 they haven't really given a general price point they've hidden it actually behind their coding subscription which costs 50 bucks a month so compare that to the pro plans of anthropic and opening eye way you have to pay top dollar like.

Ejaaz $200 to $300 a month, this is a pretty good deal. But I also do wonder whether Meta is doing this, restricting the model usage, because they may not have enough compute to actually serve the models at large scale. At the end of the day, Meta has a crazy distribution of, I think, 3.3 billion users across all of the different platforms.

Ejaaz And it's no secret, they're trying to plug their AI models into every single platform. You can use Meers on WhatsApp, you can use it on Facebook Messenger, or Facebook Marketplace or Facebook in general. And you can use it on Instagram. If they wanted to roll it out to anyone and everyone, they would have done it already, but they haven't.

Ejaaz They've only released it to the US. And then again, in a very limited fashion. So none of my friends in the UK can even use this. So I have a feeling that there's a bit of a restriction happening here because of a strain on resources. But what I would say is don't ditch your ChatGPT subscription, certainly don't ditch Claude, but try out Meta's new model.

Ejaaz I think you're able to access it through Metamuse, their personal agent, which is free for now, 100 million tokens a week. From when I last checked, I'm using it every single day and I'm burning quite a few tokens, so it's useful. I sense that that's where the most distribution or most people will get access to it, but you can also access it directly if you just sign up for the average subscription with Meta, which I believe is $20 a month, so pretty competitive.

Ejaaz Before I made this episode, I took a step back and I was thinking a year and a half ago, 18 months. Meta had the worst AI model on the market, but they'd also just torched $25 billion to hire this guy called Alexander Wang. He was the head or CEO, founder of Scale AI, 29 years old, to head up their new Meta superintelligence lab, which they'd just created because they'd fired their ex-AI team.

Ejaaz So it was a very tumultuous occasion and Zuck seemed like he was just going in complete war mode but having no idea what he was fighting for or a strategy to win and then Meta started to release a bunch of various different AI type. Products if you want to call it that I remember Meta Vibe which was their Sora or AI video app competitor AI TikTok style competitor that sucked completely flopped I remember when they launched a public news feed of their AI model this was like, one of the craziest features ever where people would have private conversations with Meta's chatbot and it would be publicly posted on a feed, much like the Facebook feed or Twitter feed, for anyone and everyone to see.

Ejaaz They thought it would be really useful. So again, many blunders have been made from Meta since back then, but then they had started to turn things around. I remember when they had announced that they were acquiring Manus, who is an AI agent company founded in China, but then they moved and domiciled in Singapore.

Ejaaz They acquired this company for $2 billion. Now, back then, AI agents wasn't really a thing, but Zuck made it very clear when he put up the money for this company. He said he believes the future is everyone and anyone should get their own personal agent. And his whole emphasis around personalization, I guess, is the entire moat and strategy behind Facebook and all of the meta social media products like Instagram, WhatsApp, or whatever that might be, is that it appeals directly to you and.

Ejaaz I, or rather you individually. It's trained on your own individual data, and it's optimized around figuring out what your preferences are. And him combining this with a Fable 5.1 type AI model, I guess, makes a lot of sense. Now, of course, that Manus acquisition was reversed two months ago when China came in and said, you can't do this.

Ejaaz We're going to reverse the entire transaction. But by that time, Meta had already got enough information and data that they required to build their own personal agent. And hey, presto, two or three weeks ago, you have their first super intelligent personal agent called Metamuse launch. And it's been a massive success for now.

Ejaaz So then the final question is, what is Meta's AI strategy and how does it differ to Anthropic and OpenAir? And the reason why I want to talk about this, well, there's many, but the main reason is there's some pretty big IPOs coming up. Anthropic is rumored to be IPO-ing at a $2.2 trillion valuation. And then OpenAir is coming in at, I think, a $1.3 to $1.5 trillion valuation.

Ejaaz So those two alone combined with SpaceX, I think makes up for 64, no, sorry, 67% of all capital raised from IPOs since the 1980s. That is a heck ton of capital raised by three companies alone. So obviously Meta's looking at this and I think they're at a, what's it, a 1.5, 1.7 trillion valuation. So if I'm Zuck, it's not only good enough for me to compete on the same territory that these guys are doing, generalized LLMs.

Ejaaz It's important for me to leapfrog them, to beat them at a different, but just as important game. And that's what I'm excited to get into on today's episode brought to you by Qualcomm. So my first take is Zuck and Meta are building an AI model for you that is personal to you, that knows you the best. Now you compare that to Anthropic and OpenAI, and they're building the best brain for everyone.

Ejaaz And Meta is building the best brain for you, individually you. It's a very slight nuance, but it is an extremely important and differentiated one. Because when you look at Anthropic and OpenAI, they're going for general artificial intelligence or super intelligence, whichever term you prefer. And the idea here is one model that serves anyone and everyone.

Ejaaz But in my mind, a general model can never eventually be as good at doing specialized work as a model that is just trained on specialized data. Makes sense when you talk about it, but has not been proven as much given that Anthropic and OpenAI has had such a lead. Until we had the likes of models that were trained locally in the open source world, especially with the Chinese LLMs that were really good at doing specific things versus the general LLMs from Anthropic and OpenAI. And then you had MetaMuse come out, which was the personal intelligence.

Ejaaz And that's the hint at where Meta and Zuck are going. They see a new territory that they want to break into and own. And that is the medium or the vector that 99% of the rest of the world who aren't AI nerds, who aren't in the bubble, who aren't crazy software engineers using cloud code to spend on multiple agents to, I don't know, build a billion dollar company.

Ejaaz They just want to focus on the average normal person. Okay, so how do you build the best general intelligence for 99% of the world? Well, you don't. You build the best personal AI model for them. And that's what Zuck figured out. That's what he realized. And so if he could own that conduit to do that, then he will own the best frontier model.

Ejaaz For years, we've been grading these AI models on random arbitrary benchmarks for general intelligence. And Zuck just came along and said, just make a product that people actually like. And it got me thinking about this a bit more, which is kind of like my second thesis. I've never viewed meta in a favorable light, If I'm being completely honest, there's the whole thing around data transparency issues, them taking our data and creating extremely addictive products.

Ejaaz I believe they had to settle like a multi-billion dollar court case recently because it was proven that they were doing exactly that. But since then, Meta has also had a strong pivot, especially when it comes to the AI stuff around privacy and owning your own keys. There's a version of Metamuse that's actually coming out a month from now that is going to be even more private than the one that they have right now.

Ejaaz And that stems from what I'm guessing to be is good product DNA. I mean, it's obvious when you kind of like just look at the numbers, Meta owns some of the most engaging, lucrative, high network effects, social media platforms ever. When you look at WhatsApp, Instagram, Facebook, Threads, which is their most recent launch, and now Meta Muse, they're all viral hit products.

Ejaaz So Meta basically knows how to build a product that can connect to many different people. So they were to try and build the best AI model for 99% of the rest of the world that aren't AI nerds. They probably have the chops to build it. So I'm realizing they have the right product DNA. But who could be the competitor to Meta? Well, I was talking to producer Luke before I started recording this episode, and I realized that one thing that OpenAI has done extremely well is build all the necessary features to have the best AI agent, like what Meta is doing, for themselves.

Ejaaz They have amazing voice capabilities. They have amazing image capabilities. Amazing frontier general intelligence with GPT-6 Astra, really sick at coding. But combining all of them together hasn't been something that they've been able to pull off. OpenAI, in my opinion, is one good product person away from building the best AI agent product, and I'm still holding out that they're able to do it.

Ejaaz Now I look at Anthropic, and I'm like, okay, these are the nerds that are focused on coding, and they are very, very good at it. They're staying in their lane. I don't expect a good AI agent that's intuitive for 99% of the rest of the people, but that's okay. Anthropik is going off to a different market. Then I look at Google and I think.

Ejaaz Google is the only other company that has arguably more or better data and distribution than Meta has. Think about it. Like Google touches search, they touch work productivity across the entire G Suite. They have Google Maps, they have the Android hardware side of things. That data moat is so lucrative and yet they still haven't been able to put out a frontier model, let alone a brand new AI agent, or even come up with the idea of doing that.

Ejaaz They haven't been able to pull it off. And that comes down to product DNA. So the final point I'll make is, when it comes to strategy, I think Meta has been able to tap into a brand new source that none of these other AI companies have found. Even OpenAI, who has all the necessary components, still doesn't have as wide a distribution as Meta.

Ejaaz Meta has business building at scale for a large number of users, OpenAI is still a private startup which requires a lot of cash up front in order to fund compute. And that brings me to my final point, which is compute. Meta is in a world where I guess they've proven that if you pay enough money, you can definitely play at the frontier, which then gives me hope that SpaceX AI or SpaceX SI, I think he renamed the entire company to SpaceX Superintelligence, Elon Musk will still be able to compete because there's a lot of capital.

Ejaaz I think they're raising around $40 billion to do exactly this right now as I'm recording this episode. And so it becomes a wall between, do you have the right product DNA? Do you have the right resources? Are you spending those resources in the right way? And do you have enough cash to fund this entire thing? It becomes a cash game when you actually know what you're doing.

Ejaaz Meta's got all of that going for them right now, as well as going after a brand new territory. And I wonder if Anthropic or mainly OpenAIR will follow suit. But anyway, that is basically my thoughts on this entire thing if i sound a little frantic or if my hair looks a little frazzled or if you're unhappy with the lighting i'm sorry i am in kyoto i'm in japan i'm here on a family holiday but i was chatting to luke every day this week and i was like dude, there's just too much going on and we need to put out episodes i want to put out episodes i actually promised him that this is going to be a short episode and.

Ejaaz I doubt this is sorry luke if you're in the process of editing this right now but the idea here is messes change things around and when i look at their stock from an investment perspective when i look at how people are treating them or at least viewing them in the, I think it's largely asymmetric at this point.

Ejaaz I think people still look upon Meta unfavorably, and obviously there are other reasons than is just AI. But Zux made it pretty clear that he wants to go all in here, and that is the future of Meta. So it makes sense to at least talk about it, at least for a bit on a single episode, whilst I'm jet-lagged as hell. Anyway, that is the end of the episode.

Ejaaz Thank you so much for listening. If any of you have been to Kyoto or Japan before, please give me Rex. I'm going to be here for another four days. So you're going to be getting more jet lagged episodes at odd hours of the night. You don't even want to know what time it is over here. And yeah, let me know about the feedback for this episode and previous episodes.

Ejaaz If you haven't checked out any of the previous episodes that we've started releasing, please let me know. The feedback, by the way, from you guys have been astounding. There's been so much support coming from you guys, a bunch of really helpful bits of feedback. I'm trying to figure this whole solo thing out.

Ejaaz But it's helpful to know that like you know being out of a studio and just kind of like talking to you face to face which is something actually that i personally enjoy a hell of a lot more is good so. I'm just going to start doing more of that but if you enjoyed this episode and you're listening to it on youtube spotify apple or wherever you listen to and you aren't subscribed or you haven't given us a rating or you haven't given us a thumbs up or if you haven't turned off notifications please do so it helps us out massively and yeah if you want to share it with a friend please do and what else oh yeah we're spinning up our newsletter back twice a week hopefully very soon getting the x thing sorted as well and you know trying to iron out

Ejaaz a few other kinks before we kind of get into our stride but i love hearing from you guys please dm me please let me know anything and everything if you disagree with me i would love to hear from you as well and yeah i will see you on the next one thank you so much for listening see you guys on today's episode of in the loop brought to you by qualcomm.

I Used to Hate Meta: Muse Wonder Changed My Mind

Kapitelklicken springt direkt hin

Shownotes

Transkriptvom Podcast bereitgestellt · Zeitmarken springen hin