Fast AI Development: S1:E1

Jefferson Nunn: Good morning and welcome to our, I guess, first episode of Fast AI Development. You know, this episode is really just a high, you know, getting started, you know, getting going, that type of thing. And, you know, just to get a little bit of an idea as to... what this episode and what these episodes are going to be all about. ⁓ So long story short, I've been in the technology field all my life. It's all that I've ever done. It's all that I know. And so I have a lot of experience in dealing with these things. and dealing with new technology and new ideas. But, you know, all that doesn't matter, you know, without proof and that sort of thing. And so here, ⁓ I'm going to actually, hopefully help all you guys, help everyone to do better, to be better in technology, to avoid the rabbit holes that you might go down to avoid, you know, trying to install something and then you realize you spend all day. trying to install something rather than getting something done. And so these days, it's all about getting things done. It's all about getting technology to work for you, not to have technology be your master. You don't want to be a slave to technology. That's just not a good place to be in. You know, every day you come in and at the drudgery every day you come in and, ⁓ hey, it's something else that doesn't work. ⁓ You know, again, we don't have time for that. We got people that want things yesterday. You want things yesterday. You don't want be held up in this vicious cycle of trying to get something done and not getting it done. And so that's why we're here. That's why I'm talking to you today. And that's why I would hope that you can follow me, you know, like and share or don't. I honestly don't care. Even if only one person, you know, finds a benefit from this stream, I'll be thrilled and ecstatic that I've managed to help one person. ⁓ I'm also doing this for me as a little bit of a diary, if you will. So I can reflect on perhaps what I've learned in the last. bit. And so what I'm going to cover today is just a little bit about what's going on in the industry, you know, as a whole, you know, some of the newest updates and so forth, things that you may not ordinarily come across because I read very quickly. And now with AI, I can even synthesize information even faster. And that'll be one of the things I'll help you guys to do. ⁓ Next, ⁓ one of the things that I've uncovered most recently ⁓ is something I think you'll find very compelling. It's a way or method of integrating several different AI components in such a way that you can execute faster and better. ⁓ I'm actually setting it up right now for myself and so perhaps tomorrow I can give you all some updates on how that is. And last but not least, ⁓ I would invite you all to contact me. I've always been extraordinarily findable, as I like to say, because the internet ⁓ is my life. It's what I like. Jefferson at JeffersonNun.com is my email address. JeffersonNun.com, a very basic website. But I'm doing all this under the fast AI development kind of a brand and kind of a title of this show and series. And we'll see how all that works out. So, industry news. There's a couple of very interesting things that I noticed. ⁓ One, surething.io. If you're not on it, well... Y'all should be on it. They are in beta. ⁓ And it's actually truly a phenomenal kind of a platform, a kind of a system. And perhaps by tomorrow, ⁓ I'll be able to have a, what do call it? Like a screen share or whatever that I can show you. ⁓ But I will at least cover the idea of what's your thing. ⁓ is and it's more than just like an open claw type of system. It's there's actually a little bit of a magic sauce behind the platform and granted right now it's in beta but hopefully over the next little while it'll come out of beta and you know I think the pricing is probably going to be really fair. ⁓ And it's something that you can connect up with all your stuff. I mean, it uses Composio on the backend. So whatever you got on the backend, you can hook it up and then it can help you to execute on a wide variety of daily and routine tasks. And like I said, because of that magic sauce in the backend, ⁓ it's something that it understands what you're doing. It understands what you're trying to do. It... The memory system on it really works very well. Particularly if you hook it up to something like Google Sheets, it'll turn that into a database and then away you go with anything that needs doing on a regular and routine basis. For example, it's literally my email client right now. I mean, I don't have an email client anymore. I find out from Sure thing what people are asking me and I... instantly reply. It prepares a reply that I can just instantly click and away it goes. An incredible platform and I think we're going to see more and more of these ⁓ types of interfaces pop up. ⁓ Everything from the Rabbit R1 which is a handheld version with open claw. It can do that ⁓ on up to something like this. So you know the industry is definitely very much maturing. And everybody from KeloCode to ByteDance to pick somebody has an open claw running KeloCode though, when in a slightly different direction. And I think they're going to find something extremely powerful with a technology called Gastown. And I know the name of it is like wild and the way they named everything, it's very oil industry. But I mean, there is something to be said for the oil industry and that they figured out how to ⁓ elementally break things down and to do things. First is something like open bang, which you kind of have to do a lot of the orchestration yourself in executing the agents. And yes, you can take open claw as a overseer and push something down from that, but that doesn't quite work as well as you might think. And so they're in like the plot. It's like, how do you... Get a swarm of agents working for you. And yes, I know there's others like Brightmine He's working on I guess called Brights swarm Hopefully, you know wish him well he can get that out ⁓ But as we all need I mean we can't have 32 different screens like, know, Pete Stein ⁓ I don't know how he does it to do that. ⁓ We need to have you know, really like one kind of an interface and whatever the system needs, it does. ⁓ And I think that's where Gastown comes in. I think something like that will be able to, ⁓ will work, right? And I've been experimenting with it a bit. And so here's, you know, an architecture. Y'all can take it, use it, don't, don't care. ⁓ But if you start with, like an open cloth, sure thing, kilo code, whatever on the head end, start there. That's your main prompt. You need to add a memory system to it that works. There are several. If you use something like Memu or Membase, any one of those, ⁓ you'll be able to ⁓ have a memory system that works basically. If you use sure thing, then it's gonna work natively. Now, the way that you... you know, link that up to something else. ⁓ I think it's, that's the hard part, right? And I find honestly, out of everything, and I've tried a lot of different things, I find linear works best. ⁓ And I'm somebody that honestly, I hate having to buy things. But for example, menu works because well, it works, right? And I've tried a lot of different self-hosted solutions and I mean, I know there's Quadrant and so forth, but Quadrant and so forth, those almost end up like a specialized kind of a system. If you need something indexed and it's a static or even a slightly dynamic index, meaning you update it on a regular basis, know, Quadrant, yes, that makes sense. You know, you have an index of documents that you get all the time. You have an index of even emails you get all the time. That sort of thing works best. But when you're trying to... go that's like almost like horizontal when you're trying to go vertical over time. ⁓ You know, you need something like a membase system that ⁓ also has an AI overseer that also has a variety of different indexing techniques that can take all this chaos that is our life and to index that over time. And I'm sorry, but MD documents is not enough. ⁓ That's almost like crazy when you go that level. All right. So now you go to the next level down and that's where Gastown works awesome, I think, ⁓ over OpenBank. We'll find out. And the idea there is you want to point that to linear and if all this pointing stuff gets too much Composio actually is great. I mean, yes, you can do it yourself. Composio has a fantastic free tier and for 90 % of y'all, the free tier is probably all you'll ever need. ⁓ If you happen to exceed the free tier, it's 29 bucks a month for their, ⁓ you know, you're going really crazy pro mode, know, Claude Maxx and burning out the subscription mode. ⁓ I do not think anybody on a personal basis will be able, except for peacetime, will be able to exceed the $29 a month plan. ⁓ And so what the nice thing about that is, it's kind of like your gateway to everything. And it's better than MCPs. MCPs are great. It's not ready. All right? Because the, well, the everything about it ⁓ in terms of the connectivity and the dynamic-ness, if you will, of the outside world, everything's changing. It really takes a human touch to keep a system ⁓ of that scale steady. And so like I said, what I'm describing here is it's not about, you know, so much about cost. Cost is a concern, it's about speed, right? Time, quality, know, whatever. This is speed here, all right? ⁓ But cost is a concern. I didn't say no cost, I said some cost, right? And it's not gonna be, you know, 5,000 a month, forget that, you know? I'm looking at you, Twitter, thank you for updating your API. ⁓ Cost, I mean, 5,000 a month, really? So. to take it to the next level down, it's like, great, you got a hundred different Asians now, well, you gotta hook it all up, that's where Composio comes in. All right, you can hook it all up to your world, linear. In linear, you can set a couple of rules, and one of those rules can be to implement, and that's where you get your cycle of how to implement. ⁓ You know, it's simple enough to have like maybe two, three, well, three different statuses at most. like maybe, you know, getting ready if you want to have like a to-do or something. And then you have an implement that's for all the agents, you know, hook up ⁓ as a separate layer. You can have a QA where you have your hyper browser, you know, hooked up ⁓ to monitor. And then you have your, okay, human review. Four steps. It's probably more than enough. And that's what you define in your labels. All right. And then on linear, then you go to hook it all up to the rest of your world, which might be, for example, EXA for web research. I find it's, it's great. ⁓ if you need real time, you.com is incredible for real time. All these APIs, ⁓ you might spend a few dollars on, ⁓ it's really not expensive, ⁓ to, know, get real time and up to date information. And I say that because one of the first things you want to do, ⁓ is to actually have ⁓ A researcher go out and research the task and define what is, and the prompt is simple, what is the most effective way to do blah, whatever that task is, and come up with the five best ways to do it. ⁓ Even if you use ⁓ some of the worst, like GLM 47, ⁓ AIs, just asking that question in a web search ⁓ tends to greatly improve. the quality of your code. And the reason being it goes out there and actually re-sensitizes the state-of-the-art information. Particularly if you add in things like ⁓ I call it ARBICS, it's A-R-X-I-D. ⁓ But if you go into ARBICS and grab the information, it will, you know, from a research task, it will, like I said, it will synthesize all this and say, we think this is the best right. Next step down, that's where you throw the OPUS for the, ⁓ well, OPUS is great at you know, rationale, I would call it. ⁓ G54 is actually great at reasoning. There's a distinct difference between the two, all right? ⁓ And sometimes, sometimes rationale is important in what you do, right? So, and if you want to do both, then you get the best of both worlds. So you take that, put the two together, now you go to next. And the next is... relatively straightforward and that's just to break it down into elemental pieces. And this is the piece I'm actually going to be trying out today. I'll let everybody know tomorrow what that's all about. And there is a project called TinyAGI slash fractals up on GitHub. Fractals as in Mandelbrot, right? So it's interesting in that it'll take any task you give it and break it down into the pieces. I'm going to try to have an agent grab this project and execute it after the research task. ⁓ And then from there, once it's in elemental pieces, you throw your swarm at it. Get it all done. ⁓ And then you QA gate that at the first level. Get all that done. Send it off to the whatever hyper browsers, whatever it is. visually look at it and get all that done and then you send it off back to you, Human Review. So that's the goal or the synthesis of this form and the process by which you can get things done. if you hear what I'm saying, a lot of this is just, you know, prompt, configure, prompt, configure, prompt, and you're done. You know, none of this ⁓ having to sit there at a command line, see a line, trying to do it all yourself. There is the gotcha. If you try to use ⁓ Claude, Claude, directly on an open claw, for example, ⁓ or you try to use now Google anti-gravity directly on, right? ⁓ They may ban you. They're the workaround. And that's where you use, know, call it the agent, whether it's, you know, open claw, whatever it is, ⁓ you have that, you use the harness below it. And that gives you actually a few distinct advantages if you do it this way. Alright, so if you have the agent actually configured to only use, you know, what I call the execution layer, right? Now you get the best of both worlds. You get an overseer, you know, which has a better prompt. That'll be like, ⁓ hey, make sure this is actually done. This is the criteria by which you make sure this gets done, right? You have that kind of a prompt. And then below that you have the ⁓ execution layer, which can be, for example, open code ⁓ with all but you know those gotcha ones ⁓ and below that you can have you can actually combine that with ⁓ my open code which i think it's a fantastic add-on to open code to actually if you're listening go work it out with this guy hire him ⁓ bring him onto the fold as a superior partner whatever you want to call it ⁓ you know franchise partner you know awesome brand partner whatever it takes you know that harness or you know know, set of sub-processes inside of OpenCode is ⁓ brilliant, I think better than Claude. Claude can be great. I find Claude really, really great at research tasks, all right? And synthesizing large bodies of information and especially with, you know, their latest, Sonnet 4.6, it's with the window and everything, it could actually very rapidly get through all of that. That said, there are limits. to ⁓ Claude in coding. I know everybody's different. For me and some of the technical levels of code, it's not so great. I think for visual design, know, laying out pages and things like that, Claude is a lot better, right, than anything else out there. But, and so for example, you can say give me 15 different variants of blah, and it will actually give you a great, you know, series like that. But when it comes to detail coding, Codex 5.3, it's so much better at that seriously level. You know, you have to go one by one, make sure it's correct and get through something that works. And especially on the first try without having to do it over and over and over again. So for those reasons, I think Codex is much better. Now, Codex, you can throw into anything you want because, you know, they hired V9, away you go. That said, you head on over to Gemini or the others, sorry, not Gemini. Well, yes, but no. You have to use Gemini for Google Anti-Gravity. I think it works with the anti-gravity subscription, so within the CLI. And like I said, if you use the claw as the harness, as the overseer, you can use the coder as the harness and away you go. You don't buy a light subscription because... they do work in a CLI basis. And by the way, you should use something like TMUX. That way you have persistent sessions, right? And make sure you get the prompt of, clean up your TMUX. Otherwise you end up with 500 TMUXs with coders running and your system crashes. So that's the overall kind of a process. And I think if you put all that together, you end up with something that's extraordinarily great. As soon as I get everything kind of tested, I'll throw it up on my... GitHub repo as a, here's the install script or something basic, here's how it all works, here's what I had found to work and the idea is to give you guys a leg up and to do things on a quicker basis. So if you all have any questions, email me at jefferson at jefferson9.com, complaints, whatever it may be, I welcome anything and everything. In the interim, other news that are out there, which again, I find... rather compelling and interesting. There's, like I said, a lot of adoption. They had an event at Tencent HQ. There's events coming up all over the place with lots of likes and impressions. And we're growing up with the millions of impressions on every, I'm sorry, views on every event. is continuing to just grow like crazy. All right, and... to include on AI and I think what we will find is that tokens are limited and there's a guy out there I've been really following Hot Isle, H-O-T-A-I-S-L-E. He's got some A-M-Ds ⁓ out there and everybody's really been burning tokens and he's one of the ones that said, yeah, I find it's true. You shouldn't be... ⁓ having to pay per token, which is ridiculous. You should have monthly subscriptions for tokens, just like in the early days of cell phones, and I tend to agree. That is something that ⁓ we really should be demanding of our AI partners. It should be ⁓ a true subscription that doesn't have unlimited. It should be unlimited, just like a cell phone's are unlimited, right? So that's something to think about. And I tend to agree on something like that. It's just kind of been crazy. There's a lot of people complaining about AI slop even in the news and people are hearing that. Human-based articles, I think, are going to be making a huge comeback, which would be fantastic because I agree AI slop when it comes to writing tends to be very sloppy. It's not quite there yet. I think over time that will improve. I really do. in terms of the quality of AI writing. That said, you cannot replace humans in the world of reporting on humans. And that's the difference. It can give you kind of trend lines and analytics, but it doesn't replace the human, you know, the extraordinary rationale, the creativity, that kind of thing. That said, you know, in the world of finance, ⁓ and I say finance in the world of I would say blockchain. ⁓ I personally, I think Bitcoin and Ethereum are great, but when it comes to the larger audience, I tend to disagree with the ⁓ crap tokens that are everywhere because a lot of them tend to be influenced by pumps and dumpers. And I really have gotten tired of that over the last decade. And again, not financial advice. You do you. I think blockchain is going to be an incredible feature. And that said, I Kraken has been making some incredible news and incredible moves to get into the marketplace and to make this work for us. And I think over time, we'll see a lot of improvement there. Outside of that, there's an ecosystem that's been evolving around OpenClaw. I'm going to continue to monitor that. I'm not quite sure. where that's going to evolve to at the moment. But that said, think autonomous, true autonomous, it's going to be where we are headed. And that's why I was spending a lot of emphasis and time on today's episode to talk about how, right? How to execute. And then as we go, I will share news and tips and tricks on what I find in terms of the execution. of these different projects and of how to, you know, just make something work. You know, you want to create, I don't know, a gardening app. Well, how, right? Or sort of just sitting there at the command line or sort of sitting there on the cloud webpage. How do you just get this done? And this is what I was discussing here. There'll be quite a few more tips and tricks that I'll be covering in the upcoming episodes. Unlike Well, today I wanted to cover the broad level, like how this is put together. As we move on in the upcoming episodes, I will go into detail. I will share my screen. I will show you what I have seen and what I have learned and show how this works. ⁓ I will share as much as I can on the GitHub. ⁓ So you can just go do it for yourself and get that, you know, little bit of acceleration. ⁓ I'm going to make it as painless as possible so that you don't have to sit there and install 50 different applications and go, I spent all day or all week installing things and configuring things and I didn't get anything done. The goal is to get you guys quickly and efficiently up to speed so that you can just get done what you need to get done. So that's the goal of Fast AI Dev and Thank you for listening. Again, there are questions, jefferson at jefferson9.com. Like, subscribe, share with your friends or don't. It's up to you. I'm doing this as a little bit of a way to document what I've learned as I learn it, as I do it. Take care. Thanks for listening.