How Meta pulled off AI's biggest comeback with Muse Wonder

Host Ejaaz Ahamadeen breaks down Meta's Muse Wonder, from benchmarks and price tag to Active Focus and Active Memory, and lays out what Zuckerberg's personal-AI-agent strategy means for Anthropic and OpenAI IPOs.

I don't know how he did it, but Mark Zuckerberg and Mehta might have just pulled off the biggest comeback in AI ever. They just released their new model called Muse Wanda. 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. 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 1 29-year-old individual to head up their new meta superintelligence labs, and it just seemed like Zuck had kind of lost the plot. But with this new model, codenamed Watermelon, now officially called Wanda, 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 3rd 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. 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 gonna get into all of this and more, the brand new models on today's episode of In the Loop, brought to you by Qualcomm.OK, so let's start with the model itself. Meta actually released two models, but it's both called Muse Wanda. But there's two varying degrees of effort across both of these models. So the standard one is called Muse Wanda, and this is basically their clawed Opus 5.5 competitor, and it actually comes in at exactly the same price. 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. 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 Metamodel.Over the last year and a half or so. When they started, metas, models weren't very performative. 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. 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. And when we look at the benchmarks, at least from the numbers that Mets is showing us, this model seems to be quite good.It's really good on personal super intelligence 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, 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 scored more than double the second place, which is Claude Fable 5.1 max. It's scored 13.1.And I'm mentioning this because Zo specifically said that the models that Met is putting out, Muse Wanda, Muse Wonder Full, which is their Mac version, the entire focus around these models is that they actually are useful to you. 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.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 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 Metta's entire AI strategy.Something that's also net new for them is Deep SW 1.1 and coding in general. This model is a major leap up from its predecessor. Now, is it as good at coding as Cloud code, Fable 5.1, as GPT 6 Astra Codex? Would you or should you drop your subscriptions there and start using Meta? Some people, and this is, this is genuinely a mixed camp, would say, yes, I saw a review from.Theo Gi, who I, 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. 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 Cloud, I tried GPT and it wasn't able to figure out my problems, but Meta, Muse, Wanda figured it out for me. So that's enough of an excuse to at least incorporate it into my work. 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. 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 GBT who basically were the ones that pioneered this entire personal memory AI assistant type thing, you would be familiar.With, OK, cool, my AI knows more about me. I'm more likely to use it. But what Met 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 uh 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. It's also about making sure that the models actually know what right bits of information to pull about you. And OpenAI has hadAll 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 and feels loyal to me, if that makes sense. And Metas come along and they put out their personal agent Muse, a few weeks ago, literally I think 2.5 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. But theFeedback 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 a 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. So, I'm gonna play around with Muse going forwards, but my guess is it's gonna 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 kind of approach this in a different way. It just does it itself. Now you might have.He versions of this from Chat GBT from Claude. They were only coding specific. That's the difference here. Typically, if you use chat GBT or Cloud, and it's doing the relentless thing, you know, it's like trying to fix and solve a problem. It's usually on coding projects. When you look at Muse Wona specifically, it's for any and every task. If it's like, OK, I think he's trying to figure out a gift for his wife or trying to figure out how to plan a specific trip andIt comes up against an obstacle like, hey, I can't get access to this website, there's a cloud flare 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, OK, 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 Ijaz 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 episod