Fable 5 is taking over, AI isn't getting cheaper, AI backup plan and WTF Loops? | Ep 17

In this episode, we explore the latest developments in AI, including cost dynamics, looping strategies, and the impressive capabilities of models like Fable 5. We  discuss operational challenges, the future of AI-driven software factories, and how to navigate this rapidly evolving landscape.


Links:
Ray: https://www.youtube.com/@RayFernando1337
Adam: https://www.youtube.com/@GosuCoder
Eric: https://www.youtube.com/@pvncher

Chapters

00:00 Introduction to Rate Limited Podcast
01:35 The Cost of AI: A Nuanced Discussion
07:40 Token Maxing and ROI in AI Usage
16:04 The Evolution of AI: From Prompts to Loops
19:09 The Value of Proactive AI Engagement
22:48 Exploring Fable Five: A New Frontier in AI
26:51 Diverse Perspectives on AI Models
30:54 Future of Work in an AI-Driven World

Transcript

SPEAKER_00 0:00

Ladies and gentlemen, you are tuned in to Rate Limited Podcast with your host Ray Fernando, Adam Larson, and Nathan Snell. And together we are now the new Rate Limited Podcast. Eric Provencea has graduated and has now been adopted by the OpenAI family. And uh we huge congratulations to Rebo Prompt and Eric Provencea. And uh it's really amazing that he's been able to be a part of our show. And uh we're gonna try to see if we can sneak him out from time to time or maybe see if we can get uh permission from some of those folks because he's just really deep in the bowels of the AI movement and uh is been able to provide lots of value. So please drop any comments below. Uh and uh, you know, wish him really well, and and we're really excited about the stuff that's going on there as well. Uh, but joining us today, we have uh Nathan Snell and uh we also have you know Adam Larson, who uh I'm just gonna hand the baton to because uh Adam actually knows him very, very well. And it's gonna be a really exciting episode from we're gonna talk about Fable Five, obviously, uh what's gonna happen with all of our jobs when it's gone. No, I'm just kidding. This is not a Doomer podcast. This is here to like really talk about building. If you're here, because uh someone probably sent you this and they care a lot about you, because we're here to really build and kind of figure out what this next moment of things are. Uh our backgrounds, we have probably decades of experience in the industry with building software and working on so many intricate systems. Uh, I think combined, all of us has probably touched our software, has probably touched a billion people's devices or something like that. So it's really exciting. Uh, you'll get to hang out with us and kind of have this like live conversation. Uh so without further ado, Adam, please go ahead and take it away.

SPEAKER_01 1:35

All right. Nathan, thank you so much for joining. And Bray, it's always a pleasure. Nathan, why don't we do a quick 30-second background on you? So, Nathan and I have been working together for about eight years. Uh, he was a co-founder with me for the company Raelion that was acquired. So, Nathan has been very AI forward from the very beginning. He will tell you he's not a coder, but he is actually a coder. Uh, back in the early days of Rayleon, him and I were both writing code together to get that company off the ground. And with AI, like it's really unlocked a lot of the stuff that he does. He leads a some pretty big initiatives on the AI side today, has a lot of insights into how to design and build products. So, Nathan, anything you want to add there?

SPEAKER_02 2:13

Yeah, I think that's uh yeah, that describes it well. Definitely AI pills or AI Maxi, depending on how you want to describe it. So it's I mean, you can't you kind of can't see around, but I've got like my local LLM machine, I've got another machine. It's like my goal is I want as many agents running as possible for as much as possible.

SPEAKER_01 2:29

So it's good to be here. This is why it segues into an amazing first topic. So early on, or late in 2024, early 2025, Nathan, myself, and a couple other uh folks that worked with us at Rayleon, we had some major debates. One of the debates was about the cost of AI. There was a big firm belief that the cost of AI was going to go to zero. And I remember arguing very adamantly against that. So Sam Altman just came out about a week ago and said for the very first time, AI is actually AI costs are actually a huge factor, and he's never seen this before. It never came up until now. I think companies more than ever are trying to write size based on the ROI they're getting from it. Uh so Nathan, what are your thoughts on that? And it's okay if you want to say that I was right and you were wrong. I'm okay with that as well.

SPEAKER_02 3:20

So uh I'd so I don't know. I'd say um I still contend that I think that uh certain AI prices are gonna go down. I mean, I I think the the original part of that argument, right, was uh which I which I think still holds true. So I'm I'm not I'm not ashamed to say that you're right. Should it happen, we'll see over the next few episodes. Um but I do think, right, the continued frontier models, I think we'll see those prices go up. But I but at the same time, right, if you look at it, uh Google just announced, or I think they're rumored to announce the actual price drop on their plus plan, right? So it's like at the same time that you know that Altman is talking about how prices are going to be going up, you also have Google and others dropping prices, right? And sort of and taking, I'd say, certain models, right? Probably not quite the same frontier models, but certain models down in terms of the price uh and what it costs or you know, to I guess get access to them and that sort of thing. So, you know, I think if we're talking again on kind of a the token cost overall, I do think we'll continue to see elements of those prices go down. But also as we look at things like Fable or Mythos or others, I mean, I think as models become more kind of nichely focused, um, you know, I suspect we'll see some of those prices, you know, go up. Um, but I don't know. I still tend to think that we're gonna see overall uh overall access and cost uh going down.

SPEAKER_01 4:32

Interesting. Ray, what's your take on this?

SPEAKER_00 4:35

Um I look at it in two different perspectives. One is like the talking point that he's throwing out that they want us to anchor to, right? To actually discuss the point about costs. And you know, I think like you said, the the costs are just gonna go up for front-year knowledge. But the the the back side of it is like as the people who actually are the controllers of the market, they set the price. So, you know, I don't think they're crying uh to their investors saying, Oh my god, we're not profitable. I they I think they can see the writing on the wall, they're extremely profitable. Um this is just here say, and then this is I'm not speaking for anyone, I'm just kind of a spectator in this whole thing. Uh, I don't have any skin in the game, I don't have any like back-end investments in any of these companies. Uh, you know, this is just us talking as as like a for me, it's just talking personally. And um I I'm just really curious um like w at what point that things kind of start to become embedded into an operating system. Apple just announced the iOS, you know, versions of 27, and they're doing their own custom infra, partnering with Google, doing some really cool stuff, making much smaller models that do really cool things that are embedded into the OS, which to me is like, oh, I didn't see that really coming from Apple. Uh and you know, for them the costs are like sort of narrowly scoped and um uh sort of integrated and taking advantage of the existing systems. So I I think there are different architecture plays here, and I'm really curious. It sounds like it's a much more nuanced discussion in terms of costs for things. So I think like one-to-one infra you know intelligence stuff. I'm just gonna send a blast of documents for a law-specific task may cost a lot because you're gonna expect higher um, you know, degree of like um you know technicalities that you need to actually get right for case law versus something that's just like I'm vibe coding uh GTA 6 or something like that, you know, uh in one HTML file. You know, maybe not that much of a cost over time, you know. So I think that's probably where people are gonna start these businesses. I'm curious for you guys, like how are you guys scoping work for tasks to be done?

SPEAKER_01 6:39

Yeah, I think Gray, like you make some good points. I think I I joke that I agree. I think there's a lot more nuance to this entire thing. Like there are gonna be smaller models that but from a coding perspective, it does feel like it's just up at this point uh from a coding perspective, just to be very fair. And one of the things that I think we missed early on, I know I missed, I I had kind of saw that the companies are highly subsidized and coders are always gonna want the best model. You can see that with Fable, which we'll talk about a bit here in a minute. But what we missed is like the length of these agent loops that are actually starting to be had. And I think this is a very interesting segue into our next topic, um, which is around and maybe we'll close this out for a second before we jump into it, but which really is around the craziness that people are going into around token maxing, writing loops that run. I've I was talking to somebody and they've been running one for 35 hours, which is just mind-blowing to me. Um so anyway, when you start when you're a company and you have people, employees doing this, and you don't see the ROI on the output of all the spin that's actually happening, how do you look at that, Nathan?

SPEAKER_02 7:48

I mean, uh I so I mean, I'd say for me, I mean, I think there's there's two parts of it, right? It's like one on a very personal level, like you know, with employees, you know, where I have like where I see their token budget and everything else. Um, I mean, like honestly, like I internalize it as like the leaderboard where I'm like, I actually want to see the usage um because like I do see really strong ROI. Um, you know, now like if I didn't, you know, then then I think to your point, like if I was not seeing the ROI, then like then I do think that it begins to become more of like a budget management thing, right? Where it's like, hey, like maybe it doesn't make sense to increase your limit to where you can spend $5,000 a month on tokens or you know, or $10 or whatever it might be, and kind of having it uncapped almost in a way. Um, you know, I do think there's a point, right, where like where productivity versus cost becomes uh you know becomes a factor. But at least today, like I haven't seen that challenge. Um I think when it hits, you know, then it then kind of like anything else, it just becomes a factor of like, okay, it probably doesn't make sense, you know, to have you know a token allowance larger than X until sort of that ROI returns and that sort of thing.

unknown 8:45

Yeah.

SPEAKER_01 8:46

Ray, do you have any any other thoughts on that?

SPEAKER_00 8:48

Um I'm curious about you guys and like harness engineering, because I know that like part of the reason why some of the limits started to get actually set were from these Korean kids that I actually met um almost like uh six months ago plus. Like they were literally spending two plus billion tokens a day, but they automated the whole suite workflow. And one of the guys is a quant trader, so he basically ported over all of his methodologies and like made this crazy system. And I don't think he realized how smart he was, but like made a whole agentec operating system and you know just slid in whatever provider they could do work for. That seems to evolved when I started talking to the folks at Factory AI, uh, because they made the whole droid missions thing. So that thing can actually run for several days and it creates tests and it kind of loops on itself. Uh and it's all like, you know, basically done via prompting. But what's crazy about it is like the longer that it runs, the more that it kind of starts to keep its own guardrails because their agentic harness is so well kind of built out from the ground, you know, going up and stuff like that. And then I'm seeing the latest with the latest versions of clawed code and stuff, where it's like starting to do that, launch sub agents to keep context really small. And these different companies are already sort of optimizing that, maybe it's because they understand that uh a little bit more of these workflows. But uh I'm just curious about your guys' take on that. Like, you know, uh are they gonna be the bigger factor here for getting costs, or how are you guys thinking about those things too?

SPEAKER_01 10:12

So so what's interesting to me is I I look at all the companies, like um, you know, Uber came out recently and said that they blew through their token budget in four months, but they didn't see an output, um they didn't see an output relative to how much was spent. I talked to folks that I've worked with in the past, their company went from all in on AI to like, hey, we need to be a little bit more conscious about what we're spending here. So the so what is ROI? Like, is ROI us saving time as individuals, or is ROI actually getting things out to the customer that equate to revenue? I think it's the latter. I think like if we're not able to actually drive more revenue for the company, there is no ROI. And I can't really think of a company that has actually been able to crack that to a enormous scale yet, other than like the big players like Anthropic and and OpenAI. I don't think we're seeing we're you know, companies are spending many millions of dollars on tokens, and we still are limited by the same factors, which is deciding what to build, how to build, making sure customers are happy, they're getting what they need on it. And I think there's a lot more to to to figure out here on the ROI side. And when it comes to the particular case you're talking about, Ray, I think it what's interesting about that is like how are they affording that? Like, I I don't understand, like, you know, it's it's wild to me. Like, I just don't understand.

SPEAKER_00 11:31

Like you don't just have a trillion dollars in your bank. I mean, bro, you're pro. Like, come on. Like what about you, Nathan?

SPEAKER_02 11:41

Yeah, I mean, I think you're right on the ROI side. And I mean, to me, there's um, and I'm again again, I'm probably somewhat biased, just like, you know, as you and IRC are even like actively building agentic, you know, uh, agentic, you know, products is there's an element too where it's like, I think to build a really great agenc product, you have to be a really, really active user of the tooling, right? So like I also think that there's that element where it's like, you know, if you're a company and you're not building agentic products as a whole, then I think that's sort of, you know, that ROI value question that you brought up, Adam, is spot on, which is like, okay, like are we actually delivering you know more, you know, more impact to a customer that's driving more revenue in ROI for them and for the company? And if the answer to that is no, then like something needs to be looked at, right? And and like I don't, I don't know, maybe I'm too A, you know, maybe I'm too bullish on AI, but like I would even say if the answer to that is no, it's probably not a harness token problem, right? It's probably an operational system problem, right? Because like at the end of the day, it's like working in an AI native way, right? Not even building an AI native product, but working as an AI native organization, I think done properly, ends up moving the bottleneck, kind of as you're describing, right? So like if the bottleneck is still in engineering, you know, then like it's you know, then it may mean, you know, and oftentimes means that like the that um like that it hasn't been sort of adopted really well yet, or like the right systems haven't been put in place such that pressure can then be put on the design team, right? Or then onto the product team and so on. So I think to me, it's like, you know, I would almost liken it to it's probably not like an AI token challenge as much as it is an operational challenge that sort of is manifesting itself in like an overusage or over-reliance on AI in such a way to where it's not actually delivering on the results, right? Because at the end of the day, like an agent isn't responsible for delivering the results. The people are today still, as much as, you know, as much as we all might like to be like, oh yeah, an agent's just gonna do everything. Like, I'm still on the hook for delivering particular results and impact. And like if I don't do my job, it's not the agent's fault, right? It's it's my inability to properly manage the team or drive the right processes and that sort of thing. So I don't know. That's sort of that's kind of my thinking.

SPEAKER_01 13:41

Yeah, I think my my closing thought um on this entire thing is relatively simple. I think like at the end of the day, everything's gonna be right-sized and we're gonna end up start spending the the money that needs to be spent there. And it's just gonna be interesting to watch like how all this ends up playing out long term. And in my in my mind, there's a bit more hype and and when we think about like the pro so I was working with one of the big AI companies, and we there was a bug that needed to be fixed. Very simple bug. And like a relatively simple bug. And their their comment was, well, that's not on the roadmap, we have it on the engineering backlog. Okay, but if you have all the AI in the world, how is there a backlog? Like, why is there a bug that can't be fixed right now? So until that goes away, whatever is stopping that from being the case, I don't see like the ROI being at the level that you know it's being touted by like that company that did a hundred X organization. You know, they wanted to do a hundred X output. So it's interesting. But it should be a goal we strive for. But we also need to be realistic that the bottleneck isn't on the code right now, it's on the operational side. And we need to figure that out. And I don't know if AI fixes that uh very easily right now.

SPEAKER_02 14:57

Yeah. I mean, that's that's one of the things that I've uh you know, even internally with like you know, at you know, add into it, you know, and you know, and even with sort of friends that are running companies, one of the big sort of distinctions I've tried to drive is that bifurcation of like AI helps with the mechanics, right? So like it makes the mechanics of doing work faster. It doesn't stop the operational inefficiencies, right? So it's like, you know, you you can, you know, and even like, for example, on the product management side, right? It's like we can have PMs that do tons of stuff really quickly and they vibe prototypes and they do research and they do all this stuff, and they can be ready to go. But then if operationally they have 17 report ups and they have to do, you know, program review, like and all these other things, like like AI can't solve the operational inefficiencies. Um, I mean, it can it can help sort of make it can drive efficiency, but at the end of the day, it still requires somebody saying, hey, we want to operate as an organization in a more lean way or more efficient in these ways. And like it's a person that has to make that decision and then try to use AI to like to solve maybe whatever gap they were trying to solve before. But I but I think you're exactly right of like mechanically, AI is great. Operationally, like there's change that has to happen to really drive results there.

SPEAKER_01 16:04

That's perfect. All right, let's segue into the next topic. The creator of Cloud Code. The creator of Cloud Code said, I don't write prompts anymore. I write loops. So Ray, I feel like this is a perfect topic for you, so I'm gonna let you take this one first.

SPEAKER_00 16:23

I was like, where you been, bro? Like these Korean kids have been doing it for like a year plus. I think uh I think the writing was on the wall a little bit, and I think that's kind of why people were starting to see like slash commands, and I remember Cursor talked about like grind mode back in the day, and people were trying to get long-running stuff with you know, Ralph Wigum and everything like that. I think a lot of people's gut reactions initially were like, of course, I I can barely get output that's reliable from my agent. How come people are just saying this out loud? And with the introduction of the latest version of Claude Code with Fable 5, it's you know, I've sent it on a really complicated task to do tons of research and come back and and look through data that would just literally take me several days even with AI because I have to keep all of this in my head, you know, beyond a million token context window. And to see it do that and launch like 12 sub agents and and like really do and write code on the fly for this and parse data and analyze things and keep those things and think through all those processes, it's just been very, very impressive. And to me, my um it's like wow, I think everything. Uh if you come get a smart model, if you have a really great harness that can um basically keep track of memory, keep track of the context, do automatic compactions and and do these techniques where you can either write its own files to kind of you know keep itself towards a goal, you know, it it it can do a lot of great work, right? And I think for me as a person with a lot of experience in software engineering, I'm trying to think about how can I make either a code base or things that I used to do in different stages of software development. Uh how can I sort of codify them so that these agents will just naturally pick them up and just start to adopt them into things. And so, like, that's kind of the way a little bit back to Nathan's point about operational speaking, like there are operations that usually happen. The engineers work on code, they hand it off to the QA team, the QA team does their job, they make reports, and then somebody has to review it. And then, you know, there's like this discussion in terms of impact for the release and then what timelines you have left. You know, if you only have two weeks left, you have a month left, how do you then fix all the bugs that you have left, and which ones should you actually focus on and regression impacts? So those are the things that like literally there's organizations that are built for this stuff to handle this, you know, and and can I make my own mini version of that with these tegentic loops so that they kind of start to do that themselves. And do they have enough taste or do I need to like what parts of me do I need to instruct to give them that vision? Uh and that's where I think loops are perfect, because then they can do those things. I get the feedback and keep improving. And I think um this is a uh I I love I love that it's kind of going in this direction. It's it's it's it's gonna cost a lot of money wise, but uh we're just curious to hear what you guys' uh experiments are or kind of what you guys have been thinking about, how you've been thinking about loops and stuff. Yeah.

SPEAKER_01 19:14

Uh yeah, Nathan, let's hand it over to you before I give my probably countertake to all of this.

SPEAKER_02 19:18

Yeah, I I think there's a few thoughts to me there. I mean, one, if you go back to like the cost side that you brought up at the very beginning, you know, there is sort of an interesting element of like even if the token cost goes down, if the overall consumption goes up, like it doesn't really matter, right? I mean, it's like it can be cheaper, but like it can still cost more. Um, I don't want to drail us back into that, but it's just one of the things that sort of strikes me there. Um, you know, one of the things I was actually surprised about on the loop piece, you know, and and like and I think conceptually in everything that you were saying, Ray, I mean, I I agree with him. Like, I've, you know, I've ran my own loops and continue to run my own loops on different things. Um, you know, I was sort of surprised that this is getting the attention in a way that it is, just because um, you know, it's not like and I don't know, and maybe I'm too sort of far into it, but it's like it's not sort of novel to me in a way. And like to me, the like, I guess where I've been anticipating more time being spent is on like the proactive side, right? Because it's like from a loop standpoint, like, okay, I still have to do the mental load, right? I still have to get out of reviewer mode and be like, okay, do the thing and kind of drive the AI into something versus like I think the value, not that it's not in a loop and there's not value there because there certainly is, but it's like I think there's a big difference in that versus the value of like, hey, like I'm proactively looking at something, I'm then kicking off that loop, you know, however long or short it may be. And like, and I'm kind of actively beginning to do that work. And I don't know. So to me, it's like, I feel like the value isn't necessarily the loop per se, as much as it is like the proactive, you know, actions that can begin to happen. Because even on like the process side we were talking about of the example that you gave, Ray, of like, okay, like, you know, we only have so much time, like, you know, what capacity can we actually fit into this time? Like, that's something that like that an agent could very easily proactively go into JIRA, for example, estimate sort of the average velocity, whether by tickets or points or what have you, and go, hey, we have this challenge going on. Here's what I suggest, right? And like, you know, I don't know if you need a loop for that per se, but I feel like that to me is where I've I've anticipated things to go from a value standpoint, like the ROI, you know, comment that Adam brought up earlier, versus like, you know, I don't know, the idea of like, oh, let's like let's maximize this as much as possible, getting these loops running and all that sort of stuff. So I don't know.

SPEAKER_01 21:23

Yeah. My take is very similar to I'd say both of yours. Um, but I I I just feel like we sound so dumb talking about loops. Like I it's it's been something that's been we've been doing this for years already now. And I remember I remember back in like early 2025, I would be like, hey, keep going until you're past all the unit tests. That's a loop. Like what the heck are we talking about here? My post on um X was I feel like collectively we as engineers sound dumber and dumber every day. This nonsense this nonsense about writing loops is really just nonsense. Why are we Acting like it's the greatest thing ever. And I say, learn to code, folks. Learn to use AI efficiently. Don't burn tokens for no reason. Look at your code. And please always look at the incentives for anyone giving advice. And I said, this is coming from someone that is very pro-AI. Like I agree. I think loops are great, but it's not anything new. It's not anything crazy. The people that are running these 35-hour, 50-hour loops and burning millions of dollars in tokens, that is not normal. And it is something that like the output of that is really how do you even measure that you got a million dollars worth of value out of that? I'm not seeing these new products pop up that are showing me that someone did a loop that launched this brand new thing out of nowhere. Just not seeing that that side of it. Go ahead, Ray. I saw you unmute there.

SPEAKER_00 22:48

Fear. Fear, fear, fear. We gotta push the fear. We gotta push the sales. We gotta get more enterprise sales in the back end. How do we do this? Oh, security. Ah, yes, everyone could break into your systems now with these intelligent models. So buy our stuff to protect you.

SPEAKER_01 23:01

Yep. It's all about incentives. Like we always got to look at like the incentives behind all of this. Alright. So in the last topic here, we're gonna talk about Fable Five. Ray, I know you actually I don't want to put words in your mouth, but are blown away by it. I know Nathan, you've got a very interesting take. So we're gonna it's gonna be fun to talk to you guys about that. So Ray, why don't you start talking about your uh hands-on experience with Fable V, or I guess what was known as mythos.

SPEAKER_00 23:30

Yeah. Blown away, I think is um what I've seen and kind of what the tasks that I threw at it made me think that there is more intelligence that's that's happening between thinking and the next set of steps that need to happen than um you know, in terms of like if you were to like tell someone to just go from New York to California and there's no roads or nothing and go figure it out. Like I feel like this is that point where we can actually kind of point someone in that direction. They're like, okay, I know enough about weather, I know enough about how to, you know, f feed myself and know enough about the world to kind of start to navigate in certain areas. Um and if that's my sole goal, then I'm I'm just gonna do it, you know, and just go. And like I don't know how long it's gonna take, but something's gonna happen. And I think to me, uh I I can actually not start not to start to see that. Like that was kind of the wish and the goal from the very beginning. And this is to me, uh I'm seeing some things more specifically in agentic like coding practices that uh that just like just the basic stuff you need. And I think it's just like literally just researching autonomously, being able to um like write code to look at large pieces of data, uh, split that tasks out to different agents so that they can understand the tasks that they need to do, uh, and then come that uh bring that back into an orchestrator, which is a basic agent loop, right? So those are like probably the four key big pieces that it can do uh on its own um so that it can kind of navigate the software world and start to build these software factories. So the biggest problem I threw at it was like I have an open claw with like like over 300 plus megabytes in the database of just text. And I said, I've had all these conversations. Can you just go specifically look at the agent that's working on health? You'll have a lot of like tons of data that you need to work with. Uh I'm trying to, you know, get down to 10% body fat. Go and like, I want to build a whole system around this and analyze all of the different conversations I've had. And it did like an amazing breakdown spinning off these different agents, thinking through things, and and then like the next step I said, okay, let's go ahead and act like put on our senior architect hat. Um, I want to make a data model that I want to store in a central data store because I'm gonna make you know widgets, I'm gonna make you know, uh open like telegram versus little conversations, I'm gonna have real-time voice, and then I'm also gonna have like a full up full-blown iOS app to do different uh uh aspects of this whole journey. Can you start to model the data that we would need to have these different interactions? And it just like started to think through and keep all of the stuff in the context because like it started to organize itself in terms of like, oh keeping a food logs, these things over here should be transactional, these things should be atomic, these things should be. And I was like, Oh, and it just continued that throughout the entire process for like almost like an hour and some change. And that's the intelligence that that really started to blow me away because I started to think about like I don't want to refactor my databases, like I'm already seeing memory problems with open claw that I know what need to be fixed, but I don't want to fix them, you know. So it's like uh but I want to kind of design my own system for this. And it's like I I think we're there at that point where and I'm trying to explain this to my my my parents and other folks. It's like uh McDonald's made uh restaurant factories, and it's not the healthiest of food, but they can produce a burger that tastes the same wherever you are in the world, and they don't need a PhD level person to make this burger. You know, they can hire someone off the street. So the this new models have the capabilities to make these software factories in which you can start to point them at operational problems or different things and have them try to go figure those things out. And then once they build the things, they can start building the smaller things or delegate those tasks. So that's kind of why I feel like this is you know scratching the surface. I haven't had you know weeks or months with this model. I've just public whatever has been publicly available, but I can already see that um from what's in there.

SPEAKER_01 27:39

So that's awesome. Nathan, you had a different take, and so this is I I want to give you a little background on Nathan. So Nathan and I over the years have like always jumped on whatever new model comes out because it's always like exciting to test it out, and we would always talk about it. But Nathan saw this one, and what was your reaction to that, Nathan?

SPEAKER_02 27:56

So yeah, so I was like, I mean, honestly, my reaction was like meh, like not that excited, and and I think a little bit more backstory to that too. So I've been a Claude Maxi for I don't know, eight, twelve, in probably eight or twelve months. And uh, and I think with with ChatGPT's 5.5 drop, and I've been using Claude Code, you know, you know, you know, since the beginning. And uh and Adam actually, you know, you, you know, you were like, look, like I know you're using this all the time. You have to go go try Codex and try 5.5 on it. It is so good. And I'm like, uh, I don't want to do that. Like, I don't like I've already got all my systems set up, they're running on cloud code, it's fine. You're like, just just try it. So like I did, and it is awesome. And I'm like, good. So I moved everything over. I've used the goal features and everything. And I'm like, this is amazing. Like, I'm a like I'm a heavy codex user now, not as much on the cloud side at this point. And like, so when Fable came out, and I was like, like, okay, yeah, I'm like, I've got goals in codex, like I've got I'm like, you know what, like I'm fine. So like I I am now like I am trying it, but as a whole, it's more of like it's it's more to like learn and experience. Um, I don't know, like it'll be interesting to see uh to see sort of like how it plays out. But this is also, I feel like, one of the first instances in a while, just as a kudos to open AI, where like open AI has had sort of like the goal concept and they've actually led here. So to me, it was actually exciting to see like, okay, like open AI, I think they've been sort of maybe a little bit less scattershot and more focused in areas, you know, and now um now Enthropic in a way is actually playing catch-up, whereas I feel like for the last six to eight months, OpenAI has been playing that a bit more.

SPEAKER_01 29:25

So yeah, for me, I feel like Fable, it's a great model. Like I think it does a great, I've done uh quite a few like long-running tasks on it. It's slow and it's expensive. And I don't know if the if that equates to like the value you're getting from it for most things. There are things that I think it totally makes sense to point it at. I do have a funny trivia question for you all. So we have the creator of Claude Code, you know, saying he writes loops now, and we have Fable 5 Mythos, which is the best model ever that we should be scared of. How many open issues do you think there are on Claude Code right now? If you had to guess, like how many 582. Okay. What do you think, Ray?

SPEAKER_00 30:12

I I want to say like 5,000 because it's just all automated.

SPEAKER_01 30:16

8,545. So wow.

SPEAKER_02 30:20

Well, like, I mean, it's 80%, right? I mean, I guess yeah. Is it the fax machine happening now, right?

SPEAKER_01 30:25

But 8,545. So anyway, like it just goes to show you that no matter what we have, and you know, some of those issues probably are not legit. We do not we do not have coding solved, like loops aren't gonna fix things. We've got the company with the best access to every model still having trouble keeping up with their backlog of stuff that's happening. And Cloud Code has always got bugs that that pop up into it constantly. Um, so it is just a funny, it's a funny like narrative that is pushed. And I will say I agree with you. Uh GPT 5.5, still my go-to model. It's fast, it's good, it's got really good design taste. And it it is the first GPT model that I actually feel like I've enjoyed coding with that I can actually remember. I remember using O3 for like my planning and reasoning back in the day. I like that model. Yep. And this model is just like my regular coder. It does good at planning, it does good at coding, it's got good design taste, and it's fast and fairly inexpensive compared to you know some of the Opus and and I guess now Fable 5. Um Alright. I think that's probably a good place to wrap. Uh Ray, do you have any closing thoughts before we call this episode?

SPEAKER_00 31:39

I think if this model with Mythos and like GPT, you know, six comes out and it's just as crazy as we think it is, or whatever, what is your guys' like next phase of life? You know what I mean? Like, I'm just curious. Like, uh where you get you like your jobs are gone, something happens, it's like, what do you do next? Well, I mean, what is it? Like, I mean, what are you thinking, Nathan?

SPEAKER_02 32:08

So uh so I've always been into like the red team mentality, right? Where it's like, okay, like anytime I'm working on something, it's like, all right, like how would I like how would I sort of beat myself? Like how do I like how would I beat my own company, that sort of stuff, just because I think from like a product standpoint, it you know it keeps me sharp. So I've done this, especially with like with you know all the generative AI stuff, I've done this on the life side. So um, so I'd say one of the things that I was thinking about is like, okay, what's really hard for AI to replicate? And it's like in physical goods, right? I mean, like very difficult for AI to replicate physical goods. Um, because like, you know, like if it's software related, which has been, you know, my life for two decades, you know, two plus decades, uh, I was like, yeah, it probably needs to not be software related. Um, so anyway, so I'm actually getting ready to launch an e-commerce brand for for like a niche category that I found. And part of actually, even the reason that I'm investing in that, honestly, is because it's like working in the frontier of stuff. There's times where you're like, holy smokes, like this is crazy. Like you change the operating models, you change the story, you see the models, and you're like, this is going to change things. I like and to being honest, I'm like, there's certain things where I feel like I have a minor amount of clarity, but honestly, it changes so fast. I'm like, I don't know. So anyway, so my my sort of uh, you know, I don't even call it a side hustle per se, but like my my sort of counter against myself. It's like if uh if I if I don't win, right? If I win in creating better frontier model stuff and everything else, it can and it ends up putting me out of a job, I'll always have an e-commerce brain, and that'll be great. You know, or like best case is like I don't put myself out of a job, and I have two things that are doing great because of AI. So that's sort of that's my that's my kind of counter to myself.

SPEAKER_01 33:38

So I've always told myself if I have to get out of software, I'm probably going back to some service job. And so back when college, I worked my way through building, like doing a lot of woodworking and stuff like that. So I probably I'd probably fall back to that. Um I I do also have e-commerce brand that that we're building out as well. Uh both so there is that that's happening as well. But it it is like one of those things where you do you really do got to kind of take that seriously. Now, I think there is a 95% chance that engineering is not going to be done away with. I think it's a very high likelihood. I could be wrong though. I've been wrong about several things. For example, I predicted React was not going to be useful and it has ended up being incredibly useful. So that's like the worst take of my life. Um pricing.

SPEAKER_02 34:23

You're wrong about pricing. I'm just kidding.

SPEAKER_01 34:24

Pricing? I think I'm pretty spot on with that. But anyway, yeah, that's what that's probably what I would do. But I I really think engineering is like going to be fun over the next 20 to 30 years, and I think if you're getting into it, the job market is going to suck for you coming out of college, but that's going to turn around. And I remember after the dot-com crash, which is when I came into the industry, I ended up having to take an unpaid internship to get started because it was that hard to actually get a job. No one was hiring, and you had to get in and kind of prove your value. So, Ray, what's your backup plan?

SPEAKER_00 34:57

A couple of things. Like, I spend time between the islands and uh the Bay Area, so I get out of the bubble quite a bit. And there's something so simple just about you know, being in nature and and just kind of chilling. And you know, I do a lot of outrigger canoe paddling, which is just amazing. You know, I could be on the water all day. And like those people who hang out on the water all day have a very different mentality about life, right? Like their whole ecosystem is completely different, you know. They feed from the water, they live by it, they hang out. It's just a whole like different vibe.

SPEAKER_02 35:28

And they don't even have fishing loops that they're just constantly trying to maximize. And sorry, I couldn't.

SPEAKER_00 35:32

Like the fishing is like the thinking, you know, it's like they're just being on the water. It's like you're kind of taking the wave, you're you feel your in your skin, you can change, like I can feel the temperature changes, and I'm like, oh, it's gonna rain the next five minutes, you know, without even looking at a phone. I don't even have my phone with me or even a watch anymore. I don't even wear a watch. And um, I worked on the Apple Watch, which is kind of crazy, right? So it's like oh, I'm getting more in touch with uh myself as a human and what humans already have built in as far as technologies. And so I'm uniquely trying to combine like how can I be more human in this AI world, and then is AI just a tool like I'm using for fishing, for paddling, for being on the water, for getting stuff done? And how is the world coming to me or or how how am I going to shape things around me? And uh, that's a question I really don't know yet. And I feel like the only counter to that is live streaming. And so if I stream live, what I'm working on thinking about, you know, that's AI has to take that knowledge, index it, serve it to a model, and like, you know, make weights, and that takes time and that's a delay. So re being real time is actually kind of a moat that I feel like I want to just keep you know plugging away at and figure out what this means for you know, being human in the age of AI.

SPEAKER_01 36:45

So yeah, Ray, you are the live stream and master. I feel like I always see you live on uh X or or YouTube or something. It's awesome. Very good. Well, guys, this has been an awesome episode. And Nathan, thank you so much for joining. Uh, hopefully we can keep you hopefully you're willing to keep doing this with us and hang out. We didn't scare you away too much. Uh all right, everyone. That's the end. Take care. Thank you.

SPEAKER_00 37:09

Take it easy, y'all. We'll see you on the next rate limited podcast.