Fable Is Back: Claude Code and the Harness Difference

I also wanted to talk a little bit about the Fable launch. So, there's been a massive set of changes in the AI world since the last time I came on here and discussed anything AI related. um some discussion points around things that I've uh been talking to people online about and uh like LLMs versus tools and some people not even knowing the difference between the things uh the different the different like what is an LLM versus what is a tool that is called and not understanding that difference I think is uh you know dangerous like in terms of uh your capabilities and also my approach to how I learn things. I think that might be meaningful and impactful if you're newer to starting in development, which this person clearly was. And then instead of whenever I pointed out that they didn't know the difference between a a large language model and a tool that is called by the large language model, instead of saying, "Hey, I I should learn these things." Oh, no, no, no. They just said you're playing semantics and you know basically calling me names and stuff like that which uh I was shocked to be honest with you. I mean I guess I'm not that shocked that it was like an online conversation uh specifically in one of the Hermes groups where I'm trying to learn a little bit more about Hermes. And this person responded with here's some large language models that edit videos and it's like they don't edit videos. That's not what these things do. uh you could make the case that some of them have the capability to do some of what we're talking about by stitching things together by stitching uh clips together right some of them can do that WAN can do that some of them and that's what I thought they were referring to no they were referring to the direct editing which is being done by ffmpeg and other tool calls so I wanted to go into tool calls versus LLMs and I wanted to go over in general where everything's at and start with fable so Let's start with the the the elephant in the room, right? So, we had Fable for a few days. We lost Fable and now Fable is back and I have not noticed any degradation in Fable whereas a lot of other people online are reporting massive degradation. So, I am using uh and this might be the reason by the way uh one thing that there's of the terminal right. So, I'm always using Claude Code uh command line. And what's by the way, what's this is another thing that I got into a little bit of an argument online about. What's confusing about Claude Code is that all of these things have a distinct name. So, this is the Claude Code command line agent. Then they have the Claude code desktop agent and they have all of these things named Claude, right? So all of their all of their models are also Claude, Opus, this Claude, but the company name is Enthropic, right? And they don't name any of their models or any of their developer studios anthropic anything which is just kind of bizarre that the company names all of their stuff like all of their stuff with this Claude nomenclature but Claude isn't really a it's a brand and not really a a individual piece of software or denotes whether it's an LLM or you know like for example this thing is you know the command line app and I've still found the command line app to be really really good. Uh they were trying to use fable inside of cursor. So it could either be hey they're Asian descent and it's being dumbed down because of that or it could be something to do with the cursor routing of their requests and what they're trying to do because their experience has not been mine. Uh post the return of Fable I can tell you I have had uh the same kind of success that I had with Fable in the first place. So Fable for sure of any model I've ever seen, this is another argument that I got into that I just thought was asinine. People are claiming that these open- source models are anywhere close to Fable in terms of their capabilities. And I can tell you it's night and day. Like I loved Miniax M3, but if Fable was sticking around and claude code command line and it was going to and you could use 100% of your usage for Fable, I would be paying the $200 and dropping all other subscriptions and ditch project Arya that I was working on, which has routing to multiple different models. Like all of that would go away and it would just be fable because it's it's been that good in my experience. So this model is gamechangingly good in my opinion. It's the best thing I have ever seen in terms of you prompt it once and then it delivers a full-blown feature set based upon that. Tested with quality, very few mistakes. Um it's honestly been the most impressive thing I've ever seen. And and before it went away and I actually got a little bit I I didn't make any videos in its absence. um on it coming out soon and I'm pretty excited to test that versus Fable. That's where I pay for a larger subscription. You can see already since Fable's return I have I've gotten to 52% of the usage that I have available for that which is 50% of your total model usage. If you didn't know, um the way the way they're doing it with the Fable return is we only have I believe actually until next Wednesday. We don't have much time with Fable before it goes only to API pricing and and I'm on I'm on the pro plan, right? I paid yearly for it and that's that's a decent value. I would have switched it had Fable been around and there wasn't GPT56 on the corner, which sounds like it's also mythos class. Um, so it doesn't make sense for me to especially when they're going to remove this. We know that when GBT56 rules out to everybody, they're probably not going to gate it the way that Anthropic gates things. Uh that so my personal opinion is yes, this is the best model. I'm pretty excited to see GBT56. I'm finally excited about a model out there. My personal opinion is everyone should go out and try Fable. I see online where people say local LLMs are just as good as Fable. That's wrong. They're not like Mini M. Like this person was saying Miniax M3 is just as good as Fable. That's not true. Like I'm a huge Miniax M3 fan. You saw in my last video where I was talking about Miniax M3 and how incredible it was and what it's capable of. I still have the Miniax. So the MiniAX application which is actually pretty good, surprisingly good. So Miniax code, their application is actually surprisingly good. Uh if I pull it up here. So, um you can see I haven't used it as much in a while because I haven't been building quite as much uh with that system, but yes, this Miniax uh this Miniax code is actually a pretty good and then Miniax is also inside my Hermes agents. I was using it, but I switched over to GPT for both of my Hermes agents. I'll talk about them in a little bit, but to say Minia Maxes and they're using the thing is they were also comparing apples to oranges. They were using Miniax inside Hermes. Um, and then they were using Fable inside cursor. One, Fable inside cursor makes no sense. Like that's not the logical choice there, especially because I think you're losing its recursive loop capabilities that it has inside of cloud code. So the thing you notice inside cloud code is you put a prompt into it and then it will sit there and like it seems like what it's doing is it's actually gathering all of the information it can locally and running a whole bunch of tool calls locally. Then it's going out with all of the information it surfaced back to anthropic using a ton of tokens coming back with with a result and then checking the result again which I think inside cursor the fable model is not going to do as well as inside their own harness. So, I think it was a a very unfair comparison to to go inside of Hermes with Miniax M3 where uh Hermes has a a recursive loop of learning. So, as he's been making requests, Hermes is learning and learning and learning on top of what the LLM itself can do. And so, Hermes is so his Hermes is actually farther along in its development. And cursor doesn't have any I don't believe cursor has any of these capabilities. It was built as a it was built as a coding tool, right? As a as a basic coding tool. So you're looking at two, you know, it's an AB comparison. And the person is also, you know, of Asian descent, which I think may have been, and he was also speculating that it may actually be giving him worse results than it should be. But in terms of in terms of capabilities, this Fable, first of all, running Mini Max M3 on I don't even think the weights are available yet to run it locally. But even when they do become available, you're going to have to have some serious hardware that even I don't have. And speaking of hardware, like I do have decent hardware. So I have an M3 Ultra with 96 gigabytes. I have uh the GX10 which is actually the the ASOS G GX10 which is actually um which is actually a uh that's a DJX Spark and this is the exact model that that is the DGX Spark. So that's and I finally have that loaded into my Exo configuration. So all three of these uh together are running qu3 coder and it's actually incredibly fast to communicate back and forth with quen coder next inside of this exo configuration. I've been really impressed with it so far, but realize this is not the primary daily driver of anything. It's uh it's, you know, GPT uh 5.5 and and Fable, right? Those are the primary drivers. Um, so that that's where I'm at with that. So, let's actually go over LLM versus tools, right?