Meta Muse Just Got a MASSIVE Upgrade (Muse Wonder)

0:00Meta just released Muse Wonder and according to their own tests, it can finally keep up with chat GPT and Claude. So, to test it out, I ran Muse Wonder through five advanced tests and by the end of this video, you will know exactly what it can do and whether or not it is actually worth your time. Muse is built to compete with the top [music] LLMs from OpenAI and Anthropic and it runs inside the Muse app with paid subscribers getting access first. For the first test, I'm giving it a picture of a random bedroom and asking it to rebuild the whole room from scratch in 3D. [music]

0:33It also has to be interactive so I can walk around inside the room and use the things in it. So, I'll upload the image, paste in this prompt and hit enter and it starts working right away. Oh, [music] and while it's working on that, if you want to go from casual using AI to actually working with it like an expert, that's [music] exactly what my school community AI Fluency is for. It's 12 weeks with a new video every day and over a thousand ready-made resources and the link to join is down in the description. Anyway, once it's done, it gives me the whole thing as an artifact

1:04so I can open it in my browser. Compared to our image, it honestly looks pretty simplistic and there isn't a whole lot going on in the room, but it is interactive so I can walk around, click the lamps beside the bed to turn them off and do the same with the one behind the sofa. There's also a slider for the time of day but moving it doesn't change much and there's a dollhouse mode that shows the whole layout from both [music] plus a compare mode that puts the real photo and 3D build side by side. So, it does everything I asked for just without much detail. Now, for reference, [music]

1:36I gave the exact same task to Anthropic's Opus 5.5 and OpenAI's GPT-6 Astra. So, let's see how they did. Starting with Opus 5.5, the difference is pretty big right away [music] because it looks almost identical to our reference photo and surprisingly, even the pictures on the wall are accurate. [music] The time of day slider is here too but this time it actually changes the lighting across the whole room. I'm especially [music] impressed with the wooden floor and the rug because the textures are really close to the original picture. The lamps turn on and off just like they did in Muse, and the dollhouse view is very smooth and shows

2:11exactly what you'd expect. Then there's GPT-6 Astra and its version looks a lot like Opus one. The pictures on the wall are accurate here, too. The rug and [music] floor textures are pretty good, and I can interact with the room the same way. But I have to say one [music] thing, Muse looks a little under rumming here for a reason because Meta's own model card admits it's weak at visual and spatial reasoning, especially when it comes to 3D. So, when it comes to building in 3D, Muse is definitely a little behind Opus and Astra. But most people aren't using AI to build 3D rooms, so let's try something a lot more

2:45practical. To do that, I'll test it on a hospital bill because [music] Meta says that's the exact kind of thing Muse Wonder can now help you with. So, I put together three PDFs for a made-up patient who went [music] to the ER with a sprained ankle. There's the hospital bill, the statement from their insurance company showing what it paid, and a summary of what their plan covers. I've hidden six mistakes across those three files and kept an answer key [music] so

3:11I know exactly what Muse should find. Next, I'll attach all three PDFs and paste in this prompt, which asks it to [music] go through everything and tell me what's wrong. Once it's done, it gives me the answer right in the chat, and it says the hospital is asking for about $4,600 when the patient should really only owe around $880. [music]

3:31That's honestly a pretty big difference. Then it breaks down every mistake it found and how much each one is worth. The biggest one is a $2,200 discount the hospital was supposed to take off the bill but never did, and it also caught smaller ones like the same X-ray [music] being charged twice. The page itself is pretty clean, too, with every single detail laid out right here from the full bill to the letters I can send to fight the charges [music] against me. Answer key it found all six mistakes, but it played one of them a little safe. It [music] noticed the hospital charge for two nights when the patient only stayed

4:05one, but it wasn't sure enough to count it, so it left that $210 out of its total. It's not perfect, but considering how confusing these bills are, I'm [music] pretty impressed. So, from three confusing PDFs you get every mistake and what you should really Oh, all in one place. Another thing a lot of people want from these chatbots is for them to remember stuff about you. Say you mention your sister's birthday in one

4:28chat and your food allergy in another. Then a [music] few weeks later, you just want it to plan something without explaining everything all over again. So, I set up that exact [music] situation. In the first chat, I asked for a few normal things like a text to my landlord and a quick home workout. But along the way, I casually mentioned that my sister Maya turns 30 on November 14th. She's the third young. I've only got $150 for fun stuff right now and I work night shifts as a nurse. Then I open a second chat and do the same thing, slipping in that I'm lactose intolerant, my dog's name is Biscuit,

5:01[music] and Maya hates loud places and is allergic to nuts. On top of that, I mention that I'm in Chicago and I hate flying. [music] But I never told it to remember any of this. So, if it knows these things later, that's all down to its memory. [music] So, in a fresh chat, I'll just ask what it knows about me and it remembers pretty much everything from Maya's birthday and her nut allergy to my night shifts and even Biscuit. And it even picked up that the $150 was just for this month. Then in the same chat, I'll ask it to plan Maya's birthday dinner without giving it any details. It

5:34sets the dinner for the evening of November 14th, so it works around my night shifts and picks somewhere quiet since [music] Maya hates loud places. Plus, the food is all vegetarian, there's a note for the restaurant about her nut allergy, it adds dairy-free options for me, and it keeps everything under $150. The only thing it missed was [music] my city, but instead of guessing, it just asked me. So,

5:56honestly, it works the way it should. [music] So, out of 10 little details spread across two checks, it remembered nine of them without me having to repeat myself. And since it can keep track of personal details like this, I wanted to see how [music] it handles something even more personal and detailed, which is someone's money. So, I put together three months of bank and credit card statements for a made-up person who

6:18makes decent money but is always broke. [music] And I've hidden a few traps in there, like a paycheck that got smaller for no clear reason and a refund never arrived. I already have an answer key, too, so I can check if Muse's math is actually accurate or not. Next, I'll paste all of this in with this prompt, which asks it to go through everything like a friend [music] who's good with money. Once it's done, it tells me right in the chat that this person spent almost $2,000 more than they earned over 3 months. It also explains why the paycheck got smaller, which is because they signed up to save more for

6:50retirement and added dental insurance. [music] So, less money lands in their account. Then, there's the dashboard it gives me, which is pretty clean, and even though it's a little more cluttered than I'd like, [music] it has every detail in it. I can click into leaks, which shows everything they're losing money on, like a gym they haven't been to since June, and an app trial that never actually got canceled. Money back shows what they're owed, like $67 from a friend and [music] that missing Amazon refund. The credit card tab shows that the card adds more in interest every month than they're [music] paying, so the balance just

7:23keeps going up. And my plan says to pay off the credit card debt before [music] saving anything because the card is charging them more than their savings would ever earn. There's also a checklist in the due to date app with specific [music] things to do. And when I checked the results against my answer

7:38key, every number matched perfectly. [music] So, it did exactly what it was supposed to do without any errors. My only complaint is that the plan is a bit slow since it takes almost 3 years to pay off the card. So, that's [music] 3 months of massive finances sorted out and you can see exactly where it's all going and what to fix first. And now that the money side is sorted, it's time for the fun part because the last test is [music] planning an entire trip from scratch, which is honestly one of the

8:04most useful things you can do with AI. So, for this test, I asked it to plan a 5-day trip to Tokyo for me and a friend in late [music] March with a budget of $1,800 each, not counting the flights. My friend is vegetarian and into photography and neither of us is a morning person. We also want one day trip outside the city and I want the whole thing on one page with a map and a budget tracker. And what I got back is a full dashboard for the trip. In the budget tab, I can switch between yen and dollars, which makes the numbers a lot easier to make sense of. Then, there's

8:36the day-by-day plan where I can click [music] through each day and see exactly what it has us doing, including a day trip out to Kamakura. Every meal works for my vegetarian friend, nothing starts too early, and the whole trip stays well under budget. And at the bottom, there's a packing list [music] and a booking checklist I can tick off, so nothing gets forgotten. Since this is a pretty big job, I also gave the exact same prompt to GPT-6 Astra and Claude Opus 5.5 to see how they compare. Let's start with Opus. Its plan looks a lot like

9:04Muse's except the map doesn't load. Everything else is just as detailed and it actually comes out the cheapest of the three. Next up is Astra. [music] It definitely looks the prettiest with nicer fonts and it's just as detailed as the other two. It even has a rain mode that gives you a different plan for each day if the weather turns bad. [music] If I had to pick one, I'd go with Astra's because it's the easiest to read and feels the most thought-through. But honestly, any of these three would get you a good trip. So, after all five tests, here's where Muse Wonder really stands. On the stuff that needs careful

9:37thinking, like the hospital bill, the money, and remembering details about you, it got pretty much everything right. But, when it has to build something you can see and click around in, like the 3D room, Opus and GPT are still a little ahead. And [music] if you want any of these models to work even better for you, the next step is turning your best prompts into agent skills. And I made a full video on exactly how to do that. It's right here on the screen, so go check it out. Thanks for watching,

10:03and I'll see you in the next one.