Trust Me, You Should Try It
BDBen Davis@bmdavis419 · 45.9K subscribers
- Views
- 15.7K
- 15,705
- Outlier
- ×0.7
- Median 21.4K
- Length
- 14:35
- Published
- Oct 15
- Thursday
- Comments
- 21
- 257 likes
Script
How it's written
- Words
- 3,158
- Words a minute
- 217
- Reading time
- 14 min
- Segments
- 10
The hook
0:00–0:18I honestly figured that eventually we would see this day come, but I didn't actually expect it to happen for a while. We now have a actually quite competitive meta model in Muse Wonder. The thing is, this release is weird and good in a lot of ways, but also they really screwed some serious stuff up.
Structure
I honestly figured that eventually we would see this day come, but I didn't actually expect it to happen for a while. We now have a actually quite competitive meta model in Muse Wonder. The thing is, this release is weird and good in a lot of ways, but also they really screwed some serious stuff up. It's a very strange release that I have actually enjoyed a lot. I've been using this model a bunch. Have some pretty cool things to show and talk about with it. There's good stuff here, but also this is not the Opus Killer. This is not going to replace your Opus usage. This isn't even really the GBT killer. There are a lot of things where even though on paper this is a very good release and in day-to-day use on a lot of things, it actually is a great model. It is still lagging behind the true frontier of OpenAI and Enthropic. I would honestly at this point put Meta at third place. As much as I'm just not a benchmarks guy at all, Artificial Analysis has generally improved their main index a lot over the past couple weeks. And having Muse Wonderful Full, again, not Muse Wonder. Muse Wonder is down at 52, but Muse Wonder Full, which I believe is their Fable or Astra, scoring a 56 second only to Opus 5.5. It's a very impressive model that from all the discourse I've read online, not a lot of people actually like all that much. I uh wanted to start with this post from Alexander Wang, who is the uh chief AI officer at Meta, the guy who's running most of this stuff. And he said that the whole team feels dumb now, but it's true. We did not explicitly train our model on 3D demo games. Because one of
Watch from 0:00the things that's happened over the last couple releases, especially with Fable 5.1, Astra, and Opus 5.5, is the 3D capabilities of models has exploded. And as such, the way that you can demo them in such a way that it is easily shared across Twitter in a striking way that goes super viral, that did a ton for getting people hyped about those models since you could easily see this artifact that they were really good at. The fact that they're not good at 3D stuff does not mean that the model itself is bad. I don't think that 3D is inherently some form of intelligence. There's very clearly some data set floating around out there. It's probably an RL environment actually like a set of problems and environments that has been sold to all the different labs because you can notice over the last 3 months or so all of the models from Kimmy to Opus to Fable to Astra have gotten suspiciously really good at 3D all at about the same time. Muse or Muse Wonder in this case did not get this whatever it is and as such the model didn't look as good on Twitter and when you just actually go to use it in day-to-day work it is not Opus 5.5. The things I want to go over with the model are the like limits and pricing and ways you can actually use it because they matter in this one more than most releases unfortunately. The god-awful split between the models, Muse Code, which is the only way you can use the only one of these two models that's actually worth using. The general performance I've found in day-to-day work, the design capabilities, the way it writes, the how it compares to other models, all that stuff right after today's sponsor.
Watch from 1:31Today's sponsor is how All by agents search and get content off the web. It is Firecrawl. I love these guys. I've talked about it before. All of their searching and scraping and just it's great. They're really, really good for that stuff. But the thing that I specifically want to talk about is actually Alexandria. They're calling it a library for super intelligence. And I think it's fairly accurate because they have some insane data sources in here. This isn't a new thing you have to go set up and do a bunch of weird stuff for. It is just a thing that is now exposed within Firecrawl that your agents can now search for whenever they want to. that includes over 113 million index sources. To show you an example of this, say you wanted to, you know, check the web analytics on a site. You can use Alexandria to grab website traffic, like right here. This is the visits on my site or maybe you want to see what the current YC companies are doing. You can query YC's data, check and see what the breakdown of everything is. And those are just two of 177 providers that make this an extremely useful thing to have. Go give your agents the search tools they need at davis7.link/firecrawl. I'm going to be honest, I really don't want to spend much time going over the difference between full and not full. The short of it is don't bother with the non-ful version. It is the same price as Opus for a much worse model that feels about on par with something like GBT 6 soul. Not 6.1 soul, six soul. It is not a very good model. Doesn't feel that great. Doesn't do anything special.
Watch from 3:00You're going to get effectively unlimited obus through a clawed $200 a month plan. And when you're comparing that to a $200 a month Muse plan to use their Opus, which is just worse, just don't bother with it. The one that's interesting here is the Wonder Full, which again should be a different model name. Wonder should be their Aster or Fable. And then we should have not spark, not Wonder, but something in between those two for what the actual Wonder is here. Again, confusing, dumb, I know. I want to mostly just talk about Wonderful, which you have to be either on the $50 a month or $200 a month plan to actually use in Muse Code. The only place you can use it are there and in T3 Code because T3 Code wraps the Muse CLI. I've used the Muse CLI a bunch in testing this model. And I'm going to be honest, I I don't fully know why this even exists to be real with you. It is not better than Codeex, not better than Cloud Code. If you've done a lot of stuff in Pi, it's probably going to look fairly similar to that. There's a lot of things in here that feel very PI like while just kind of being worse. their auto approval thing, which I really like in both Codeex and Cloud Code where the model will automatically approve actions for you. Really nice feature. That's how I wanted to use the Muse CLI, but I couldn't because their auto mode doesn't actually work as an auto mode. I still found that I had to approve five things per turn usually. And if you're approving a bunch of stuff, the approvals suddenly mean nothing. You stop reading them and auto mode is worthless. You might as well just have it in full access. YOLO, which is what I've left it in. The CLI isn't great.
Watch from 4:21It's a CLI. There's not much else to say about it. If you're going to use Muse, use it through T3 Code. It'll be a much better experience. And the actual usage on the subscription is not as subsidized as other plans. And like I said at the start, we don't actually know how much the full version of this model costs. From what they've said, it's their Fable or their Astra, which means that it's probably a much more expensive model. And if we estimate the token prices at Fable or Astra prices, I ran into this issue where I was just destroying the limits on my $50 a month plan. I had to upgrade to the 200 to keep researching for this video. What I found is that just guesstimating based on a 10-minute time window where I had the starting usage, ending usage, and then the sessions in between, I had Opus Go put this together. It's guesstimating that within my 5 hour limit, I got about 42 million tokens, about 128 million tokens weekly or $332 per month of actual API usage on the $50 a month plan. So, it's still subsidized at about 6.6x the normal price, which is good. It's not nothing, but when you compare that to what you get out of the Claude Code sub and previously the Codeex sub, the subsidization on Codeex has not been nearly as good since the changes that were made recently to the limits. You can still get a lot out of it, but the actual numbers are a lot lower, but since OpenAI models are so much more efficient and cheap than this model is, you can kind of get away with it. This model doesn't seem particularly token efficient, and as a result, you chew through your limits really, really fast.
Watch from 5:48Like I said, I had to upgrade to the $200 plan to keep working to figure out this video. And as I was testing it, I was looking at my usage and at 7:42 p.m. it was at 25% of my 5 hour. And then about 10 minutes later at 7:52, I measured it again and I was at 34% of my 5 hour. And this was with three agents running at once. No sub aents, nothing crazy, just normal high reasoning museful doing its thing, destroying the limits. This is very very bad when compared to Opus that has a similar 5 hour and weekly structure to effectively feel unlimited. I haven't even noticed that I have rate limits on Opus. When using this one, I've already run into my five 5 hour multiple times. You're not going to get a huge amount of usage out of this sub. It's not a great value for the money, but it is a cool model. I do really want to stress that this model does not feel the way I expected it to. Generally speaking, we've had the GPT flavor model, which is it has no flavor. it is super autistic and will just do what you tell it to. And then you have the claw flavor and the clawisms that got really bad over the summer. Like they really leaned into those and all of those awkward sentence tells where you could just oh that's clawed. They've beaten that out of Obus 5.5 a decent amount. It's still in there and still clearly claw flavored, but not in a gross obnoxious way. This one has its own flavor, its own form of nonsense that I'm sure can get tiring over time, but it does feel different. Most other models feel like a clawed light. They will have a lot of the clawisms because they're very clearly being trained on clawed data. This one feels wildly
Watch from 7:16different. The writing on it is really good. It's very good at communicating. It's really nice to interact with. I had a change here earlier where I was doing some stuff on a little dashboard test I was doing with it and it's very clear and easy to read. Like done. The flat cumulative chart is replaced with a peer comparison strip. And yes, your read was right. Without analytics API ooth, there is no historical series to fetch. Only the snapshots we record ourselves going forward. This sentence makes sense. I get the gist of what it came up with and did here. It included a screenshot in the response, which is really nice. It's good about doing that kind of thing, using the computer used tools and all the stuff at its disposal to communicate things to you more clearly. The rest of it was clear bullets and then down here, what's next? Some suggestions of what to do that are actually like good suggestions that I like having. This is partially because of some changes I made to my agents MD. I added about a paragraph long section of, hey, keep your responses clear, short, use varying sentence length. Um, at the end of them, include a couple ideas of what to go do next. I found that really improves response quality because they kind of think forward. And also, it's easy when you're working with dozens of threads to kind of forget where you were. And having these little quick, oh, should we commit and push this? Should we do this little refactor? Like, all of these are nice things to have. I can just be like, okay, uh, do one and three. And now it will just continue off and go do its thing. It's great. Another example here is you can see this little visualization of the timeline of the messages I sent, what it did, all that stuff. This timeline is good. And also, this is raw
Watch from 8:47HTML that it designed and showed to me here. This is a new feature in T3 code that I actually threw in there because I've wanted this. I love this feature in Codeex. I wanted it in T3 Code. The models, you just tell them to visualize something in chat, can write an HTML page of whatever they want and use some CSS tokens that are pre-programmed in to match your T3 code theme. and then they'll make stuff like this in the chat to show you things. It's really, really convenient to have. And you can see the way it made this was good. It's good at doing this kind of thing. To give you another example of its communication being really good. It did some redesign stuff on P thing to clean up that actually went pretty well. The design of this model right out of the box is not great. It's not a design model. To be fair, I don't know if any model is really a design model. I think design is one of those weird problems where even if on a technical level, they can do whatever designs you want them to. Design is more about the functionality and where things go and the actual user experience rather than the colors and text and ways things are laid out. That matters a lot. But you can't program and train on it quite as well as you can something like code or even 3D which while being very complex is actually a more constrained and simple problem space of like making a boat is just a boat. There's not too many decisions that need to be made. Design requires a lot of decisions to be made about how UX should work. And I think that's where models really struggle right now. The line between models that are insanely useful tools that we still have to steer and direct and god models that can go off and do literally anything for long periods of time. It's the decision making in between that is just missing
Watch from 10:18on these long running tasks. Oftent times I find myself not having to do too much steering but just they'll reach these fork in the roads and I'll notice that they take the wrong path and I just have to push them down that other path. design is kind of one of those problems. And when you put some more constraints on it and you just have it work in an existing system with decent UX but not great polish, it does a really good job of polishing it up. The new pick thing feel is quite good. I'm quite happy with how all of this looks and feels at this point. I like how this thing interacts and I also really like the way that it prompts and works with other agents. The messages that I've seen it pass back and forth between models in its own family and also to Claude and Codex, they look really good. It's very good at communicating things and feels genuinely different, which I think that's the big thing that's cool here is that there is now a new frontier lab in the US that is keeping up in competitive. Still need to work on pricing, still need to work on the services where you can use this. It's early days for meta, but they're getting there and they feel different and we could live in a world where you have this new third one that is a different feel that actually gets useful stuff done. This is what I've wanted from Grock for a long time, but they just haven't gotten yet. Maybe eventually Gro 5 will get there. I'm unsure. They got a long way to go on the model front right now, but Meta is certainly catching up very fast. But ultimately, again, when you compare the actual quality of output and code and capability that this thing has, you know, testing it on writing some effect code and spelt stuff and throwing that all together from scratch, it's about on par with something like 6.1 Soul.
Watch from 11:48Nothing bad, but nothing great. Lots of try catches. Not perfect. Like Opus is still in a league of its own. Claude models still do that discernment and quality better than anything else. Obviously, there's no 3D stuff. The design isn't all that great. Its novel capability is its communication. And I could see a workflow where I have a Muse thread with instructions to only ever do work by calling Claude sub agents. It's just an orchestration layer because it is good at orchestrating. It's great at communicating. That could potentially be very compelling, but for $200 a month and pretty crappy rate limits, not worth it. Not yet. Really, I think the interesting thing here is just that Meta is getting more competitive by the day and they were able to make a model this good this fast. I'm impressed. I want to see where this goes. They have a potential future here. Just need to fix pricing and product and naming and get some more data sets in there and just see where it all goes. We uh are still waiting on the atomic bombs from both OpenAI and Enthropic. Fable 5.5 is floating around out there and Astra 6.1, whatever it ends up being, also going to show up at some point. Who knows? This is probably going to get blown out of the water soon, but at least for right now, this is an impressive release from Meta, and I'm curious where it goes next. you enjoyed the video, make sure you like and subscribe.
Watch from 13:19Full transcript
- 0:00
I honestly figured that eventually we would see this day come, but I didn't actually expect it to happen for a while. We now have a actually quite competitive meta model in Muse Wonder. The thing is, this release is weird and good in a lot of ways, but also they really screwed some serious stuff up. It's a very strange release that I have
- 0:21
actually enjoyed a lot. I've been using this model a bunch. Have some pretty cool things to show and talk about with it. There's good stuff here, but also this is not the Opus Killer. This is not going to replace your Opus usage. This isn't even really the GBT killer. There are a lot of things where even though on paper this is a very good release and in day-to-day use on a lot of things, it
- 0:41
actually is a great model. It is still lagging behind the true frontier of OpenAI and Enthropic. I would honestly at this point put Meta at third place. As much as I'm just not a benchmarks guy at all, Artificial Analysis has generally improved their main index a lot over the past couple weeks. And having Muse Wonderful Full, again, not Muse Wonder. Muse Wonder is down at 52,
- 1:02
but Muse Wonder Full, which I believe is their Fable or Astra, scoring a 56 second only to Opus 5.5. It's a very impressive model that from all the discourse I've read online, not a lot of people actually like all that much. I uh wanted to start with this post from Alexander Wang, who is the uh chief AI officer at Meta, the guy who's running
- 1:23
most of this stuff. And he said that the whole team feels dumb now, but it's true. We did not explicitly train our model on 3D demo games. Because one of the things that's happened over the last couple releases, especially with Fable 5.1, Astra, and Opus 5.5, is the 3D capabilities of models has exploded. And as such, the way that you can demo them in such a way that it is easily shared
- 1:45
across Twitter in a striking way that goes super viral, that did a ton for getting people hyped about those models since you could easily see this artifact that they were really good at. The fact that they're not good at 3D stuff does not mean that the model itself is bad. I don't think that 3D is inherently some form of intelligence. There's very clearly some data set floating around out there. It's probably an RL
- 2:06
environment actually like a set of problems and environments that has been sold to all the different labs because you can notice over the last 3 months or so all of the models from Kimmy to Opus to Fable to Astra have gotten suspiciously really good at 3D all at about the same time. Muse or Muse Wonder in this case did not get this whatever
- 2:26
it is and as such the model didn't look as good on Twitter and when you just actually go to use it in day-to-day work it is not Opus 5.5. The things I want to go over with the model are the like limits and pricing and ways you can actually use it because they matter in this one more than most releases unfortunately. The god-awful split between the models, Muse Code, which is the only way you can use the only one of
- 2:48
these two models that's actually worth using. The general performance I've found in day-to-day work, the design capabilities, the way it writes, the how it compares to other models, all that stuff right after today's sponsor. Today's sponsor is how All by agents search and get content off the web. It is Firecrawl. I love these guys. I've talked about it before. All of their searching and scraping and just it's
- 3:09
great. They're really, really good for that stuff. But the thing that I specifically want to talk about is actually Alexandria. They're calling it a library for super intelligence. And I think it's fairly accurate because they have some insane data sources in here. This isn't a new thing you have to go set up and do a bunch of weird stuff for. It is just a thing that is now exposed within Firecrawl that your agents can now search for whenever they
- 3:29
want to. that includes over 113 million index sources. To show you an example of this, say you wanted to, you know, check the web analytics on a site. You can use Alexandria to grab website traffic, like right here. This is the visits on my site or maybe you want to see what the current YC companies are doing. You can query YC's data, check and see what the
- 3:51
breakdown of everything is. And those are just two of 177 providers that make this an extremely useful thing to have. Go give your agents the search tools they need at davis7.link/firecrawl. I'm going to be honest, I really don't want to spend much time going over the difference between full and not full. The short of it is don't bother with the non-ful version. It is the same price as Opus for a much worse model that feels
- 4:13
about on par with something like GBT 6 soul. Not 6.1 soul, six soul. It is not a very good model. Doesn't feel that great. Doesn't do anything special. You're going to get effectively unlimited obus through a clawed $200 a month plan. And when you're comparing that to a $200 a month Muse plan to use their Opus, which is just worse, just don't bother with it. The one that's interesting here is the Wonder Full,
- 4:34
which again should be a different model name. Wonder should be their Aster or Fable. And then we should have not spark, not Wonder, but something in between those two for what the actual Wonder is here. Again, confusing, dumb, I know. I want to mostly just talk about Wonderful, which you have to be either on the $50 a month or $200 a month plan to actually use in Muse Code. The only
- 4:55
place you can use it are there and in T3 Code because T3 Code wraps the Muse CLI. I've used the Muse CLI a bunch in testing this model. And I'm going to be honest, I I don't fully know why this even exists to be real with you. It is not better than Codeex, not better than Cloud Code. If you've done a lot of stuff in Pi, it's probably going to look fairly similar to that. There's a lot of
- 5:15
things in here that feel very PI like while just kind of being worse. their auto approval thing, which I really like in both Codeex and Cloud Code where the model will automatically approve actions for you. Really nice feature. That's how I wanted to use the Muse CLI, but I couldn't because their auto mode doesn't actually work as an auto mode. I still found that I had to approve five things
- 5:35
per turn usually. And if you're approving a bunch of stuff, the approvals suddenly mean nothing. You stop reading them and auto mode is worthless. You might as well just have it in full access. YOLO, which is what I've left it in. The CLI isn't great. It's a CLI. There's not much else to say about it. If you're going to use Muse, use it through T3 Code. It'll be a much better experience. And the actual usage
- 5:56
on the subscription is not as subsidized as other plans. And like I said at the start, we don't actually know how much the full version of this model costs. From what they've said, it's their Fable or their Astra, which means that it's probably a much more expensive model. And if we estimate the token prices at Fable or Astra prices, I ran into this issue where I was just destroying the
- 6:16
limits on my $50 a month plan. I had to upgrade to the 200 to keep researching for this video. What I found is that just guesstimating based on a 10-minute time window where I had the starting usage, ending usage, and then the sessions in between, I had Opus Go put this together. It's guesstimating that within my 5 hour limit, I got about 42 million tokens, about 128 million tokens
- 6:37
weekly or $332 per month of actual API usage on the $50 a month plan. So, it's still subsidized at about 6.6x the normal price, which is good. It's not nothing, but when you compare that to what you get out of the Claude Code sub and previously the Codeex sub, the subsidization on Codeex has not been nearly as good since the changes that
- 6:59
were made recently to the limits. You can still get a lot out of it, but the actual numbers are a lot lower, but since OpenAI models are so much more efficient and cheap than this model is, you can kind of get away with it. This model doesn't seem particularly token efficient, and as a result, you chew through your limits really, really fast. Like I said, I had to upgrade to the $200 plan to keep working to figure out
- 7:20
this video. And as I was testing it, I was looking at my usage and at 7:42 p.m. it was at 25% of my 5 hour. And then about 10 minutes later at 7:52, I measured it again and I was at 34% of my 5 hour. And this was with three agents running at once. No sub aents, nothing crazy, just normal high reasoning
- 7:40
museful doing its thing, destroying the limits. This is very very bad when compared to Opus that has a similar 5 hour and weekly structure to effectively feel unlimited. I haven't even noticed that I have rate limits on Opus. When using this one, I've already run into my five 5 hour multiple times. You're not going to get a huge amount of usage out of this sub. It's not a great value for
- 8:02
the money, but it is a cool model. I do really want to stress that this model does not feel the way I expected it to. Generally speaking, we've had the GPT flavor model, which is it has no flavor. it is super autistic and will just do what you tell it to. And then you have the claw flavor and the clawisms that got really bad over the summer. Like they really leaned into those and all of those awkward sentence tells where you
- 8:24
could just oh that's clawed. They've beaten that out of Obus 5.5 a decent amount. It's still in there and still clearly claw flavored, but not in a gross obnoxious way. This one has its own flavor, its own form of nonsense that I'm sure can get tiring over time, but it does feel different. Most other models feel like a clawed light. They will have a lot of the clawisms because they're very clearly being trained on
- 8:45
clawed data. This one feels wildly different. The writing on it is really good. It's very good at communicating. It's really nice to interact with. I had a change here earlier where I was doing some stuff on a little dashboard test I was doing with it and it's very clear and easy to read. Like done. The flat cumulative chart is replaced with a peer comparison strip. And yes, your read was right. Without analytics API ooth, there
- 9:07
is no historical series to fetch. Only the snapshots we record ourselves going forward. This sentence makes sense. I get the gist of what it came up with and did here. It included a screenshot in the response, which is really nice. It's good about doing that kind of thing, using the computer used tools and all the stuff at its disposal to communicate things to you more clearly. The rest of it was clear bullets and then down here,
- 9:29
what's next? Some suggestions of what to do that are actually like good suggestions that I like having. This is partially because of some changes I made to my agents MD. I added about a paragraph long section of, hey, keep your responses clear, short, use varying sentence length. Um, at the end of them, include a couple ideas of what to go do next. I found that really improves
- 9:49
response quality because they kind of think forward. And also, it's easy when you're working with dozens of threads to kind of forget where you were. And having these little quick, oh, should we commit and push this? Should we do this little refactor? Like, all of these are nice things to have. I can just be like, okay, uh, do one and three. And now it will just continue off and go do its thing. It's great. Another example here is you can see this little visualization
- 10:11
of the timeline of the messages I sent, what it did, all that stuff. This timeline is good. And also, this is raw HTML that it designed and showed to me here. This is a new feature in T3 code that I actually threw in there because I've wanted this. I love this feature in Codeex. I wanted it in T3 Code. The models, you just tell them to visualize something in chat, can write an HTML page of whatever they want and use some
- 10:32
CSS tokens that are pre-programmed in to match your T3 code theme. and then they'll make stuff like this in the chat to show you things. It's really, really convenient to have. And you can see the way it made this was good. It's good at doing this kind of thing. To give you another example of its communication being really good. It did some redesign stuff on P thing to clean up that actually went pretty well. The design of
- 10:52
this model right out of the box is not great. It's not a design model. To be fair, I don't know if any model is really a design model. I think design is one of those weird problems where even if on a technical level, they can do whatever designs you want them to. Design is more about the functionality and where things go and the actual user experience rather than the colors and text and ways things are laid out. That
- 11:14
matters a lot. But you can't program and train on it quite as well as you can something like code or even 3D which while being very complex is actually a more constrained and simple problem space of like making a boat is just a boat. There's not too many decisions that need to be made. Design requires a lot of decisions to be made about how UX should work. And I think that's where models really struggle right now. The
- 11:36
line between models that are insanely useful tools that we still have to steer and direct and god models that can go off and do literally anything for long periods of time. It's the decision making in between that is just missing on these long running tasks. Oftent times I find myself not having to do too much steering but just they'll reach these fork in the roads and I'll notice that they take the wrong path and I just
- 11:57
have to push them down that other path. design is kind of one of those problems. And when you put some more constraints on it and you just have it work in an existing system with decent UX but not great polish, it does a really good job of polishing it up. The new pick thing feel is quite good. I'm quite happy with how all of this looks and feels at this point. I like how this thing interacts and I also really like the way that it
- 12:19
prompts and works with other agents. The messages that I've seen it pass back and forth between models in its own family and also to Claude and Codex, they look really good. It's very good at communicating things and feels genuinely different, which I think that's the big thing that's cool here is that there is now a new frontier lab in the US that is keeping up in competitive. Still need to
- 12:39
work on pricing, still need to work on the services where you can use this. It's early days for meta, but they're getting there and they feel different and we could live in a world where you have this new third one that is a different feel that actually gets useful stuff done. This is what I've wanted from Grock for a long time, but they just haven't gotten yet. Maybe eventually Gro 5 will get there. I'm unsure. They got a long way to go on the
- 13:01
model front right now, but Meta is certainly catching up very fast. But ultimately, again, when you compare the actual quality of output and code and capability that this thing has, you know, testing it on writing some effect code and spelt stuff and throwing that all together from scratch, it's about on par with something like 6.1 Soul. Nothing bad, but nothing great. Lots of
- 13:21
try catches. Not perfect. Like Opus is still in a league of its own. Claude models still do that discernment and quality better than anything else. Obviously, there's no 3D stuff. The design isn't all that great. Its novel capability is its communication. And I could see a workflow where I have a Muse thread with instructions to only ever do work by calling Claude sub agents. It's
- 13:42
just an orchestration layer because it is good at orchestrating. It's great at communicating. That could potentially be very compelling, but for $200 a month and pretty crappy rate limits, not worth it. Not yet. Really, I think the interesting thing here is just that Meta is getting more competitive by the day and they were able to make a model this good this fast. I'm impressed. I want to
- 14:02
see where this goes. They have a potential future here. Just need to fix pricing and product and naming and get some more data sets in there and just see where it all goes. We uh are still waiting on the atomic bombs from both OpenAI and Enthropic. Fable 5.5 is floating around out there and Astra 6.1, whatever it ends up being, also going to show up at some point. Who knows? This
- 14:24
is probably going to get blown out of the water soon, but at least for right now, this is an impressive release from Meta, and I'm curious where it goes next. you enjoyed the video, make sure you like and subscribe.
Description
Muse Wonder, the full version is way better than you think. It's not quite Opus, but it's really good and has some unique flavor that I'm really liking (but also some huge issues)
Thank you Firecrawl for sponsoring! Check them out at: https://davis7.link/firecrawl
my limits issues - https://x.com/davis7/status/211019999...
benches - https://slopalytics.com/?models=gpt-6...
My X, Discord, and more: https://davis7.sh/
