Subscribe
Share
Share
Embed
If you are seeking new ways to increase your ROI on marketing with your commerce platform, or you may be an entrepreneur who wants to grow your team and be more efficient with your online business.
Talk Commerce with Brent W. Peterson draws stories from merchants, marketers, and entrepreneurs who share their experiences in the trenches to help you learn what works and what may not in your business.
Keep up with the current news on commerce platforms, marketing trends, and what is new in the entrepreneurial world. Episodes drop every Tuesday with the occasional bonus episodes.
You can check out our daily blog post and signup for our newsletter here https://talk-commerce.com
Brent Peterson (00:00.578)
Welcome to this special AI edition of Talk Commerce that I have Paul Byrne is the author of Adapt or Die Paul, go ahead and do an introduction for yourself. us your day-to-day role outside of writing your book and one of your passions.
Paul Byrne (00:06.298)
Great to be here.
Thank you.
Paul Byrne (00:18.104)
Right.
Yeah, so I am the president of Razoyo We are a long time development firm. We started in the e-commerce space, is where we met Brent and partnered on some Magento things. Your company really kind of wrote some fantastic extensions for Magento. So we sold quite a few of those to our clients. And
I, day to day, I mostly focus on, you know, kind of what an executive might focus on, you making sure that the business is improving itself, handling any specific issues that might be, you know, require extra resources or something along those lines. And just, you know, the admin role of overseeing finance, accounting, marketing, all that sort of stuff. But I do spend a fair amount of time
talking to our clients and that's sort of what led me to write the book. I saw that there was a lot of kind of concern about where AI is going and everything kind of seems up in the air.
And I know a lot of development companies, including, including Razoyo, saw a lot of clients kind of sitting on the sidelines, wondering if the projects they're working on or what they were doing was going to be sort of overshadowed by AI or sort of engulfed in it. And would they get any return on investment for writing stuff? And which is a legit question because that's definitely.
Paul Byrne (02:02.062)
That's definitely the case. You have to be able to pick and choose and know what's going on. I spent a lot of time researching what AI is and the history of it and where it's going and what the limitations are. writing the book was made just to help people make better decisions who are either founders or maybe the product owners or their CEOs of a
of a smaller company that doesn't have a bunch of people on staff that are AI experts. And so just, you know, providing people some guidance in that area was my main goal.
Brent Peterson (02:39.086)
Yeah, good. right. All right. But before we talk about the book, I'm going to we have to do the free joke project. I'm going to tell you joke. There you go. Yeah. You give me a rating eight to 13. So here we go. There are two types of people in the world, those who can't relax when there are dishes in the sink and those don't and those who don't even notice them and they end up marrying each other.
Paul Byrne (02:43.022)
Yeah. Okay. All right. All right. I'm on joke welfare here.
8 to 13.
Paul Byrne (03:06.766)
That one reads a little true. I'm going to say that's a 12. I'll give that one a 12.
Brent Peterson (03:12.044)
Yeah, I picked that one because that is my life. Yeah, absolutely. I won't I'll let people out there guess whether I'm the dishwasher or I'm the sink not looking-er
Paul Byrne (03:23.955)
Yeah, if my wife's watching this, she knows exactly which one I am.
Brent Peterson (03:30.614)
Okay, good. right, well, so, you know, this is such a great topic. And, you know, there's still so much pushback on using AI and even what AI can be used for. And I literally had a call an hour ago on that very subject on, you know, what do you use for workflows? But so kind of give us a premise of the book and how and the 10,000 foot view to start.
Paul Byrne (03:53.718)
Yeah, I think as I was writing the book, which was a follow up really, and sort of intellectually and research wise to several sessions that I taught at the University of North Texas, where I was teaching the students in digital commerce, so the e-commerce majors.
And I had done some research at that point trying to figure out what was going on with AI, cetera. And I did some follow-up research. And one of the key insights as I tried to distill this down to make it useful was the difference between type one and type two learning and where we are in that development process. And so.
or one of the early chapters in the book introduces that concept. And the basic concept is type one learning is more like reactive. So you already know the answer. It's a multiple choice question on a test, something or a math problem that you already know how to solve. And you just kind of.
You don't really have to think about it. You just go into solve mode. You go into answer mode and that's that's type one learning. You think about it as like a glorified Google search. The second type two has some very specific requirements and it's more like reflective thinking. And so it's the type of answer where you have to consider the question, you know, think about it.
You haven't encountered this in the past, in the way it is now. There's no way you could answer it from memory or anything like that. And so the type two learning is more along those lines. And that's like kind of considering human terms, right? In machine terms, it's a little bit different. But we know what
Paul Byrne (06:01.164)
of what the LLMs and the machine learning does right now is it just kind of predicts answers, right? It's not really trying to get the right answer. It's trying to get the one that's most likely to be kind of acceptable based on what it's been trained on. And it's highly dependent on the data set that it's been trained on, which now is the entire text of the internet.
And so that kind of learning is what I would call really glorified search, Google search. Even though it seems like it's talking to you and it seems like it's really planning. The type two.
Type two learning and type two thinking in AI is going to be then, well, it has to have a plan, has to have persistent memory, it has to have, you know, be essentially goal seeking. And there are a couple of other requirements in there, but I think it gives you, you know, a good idea of just going over that.
what the type two learning is and our type two thinking is from an AI perspective. And we're just not there yet. I think, you know, I follow a gentleman named Yann LeCun who is a French researcher. used to be the head of Metta's artificial intelligence unit or whatever they call the business unit. And he's now left there to run his own shop.
And so maybe he's getting close. He's the guy that kind of is, I would say, sort of the most direct in his assessment of how we're getting there. And in his point of view, you would say that the LLMs that we have are kind of an off ramp from AI research. And they're a good off ramp, right? Like there's so much.
Paul Byrne (08:02.81)
that we can do with them. And I think we're just kind of scratching the surface of the current models and the type of machine learning we have now, but it is going to be limited. so, you know, in my book, I try to explain how to understand what those limitations are so that you can evaluate, oh, is this project I'm working on likely to be, you know, become part of an LLM in the future? And I think a lot of the early, you know,
projects in that space, including e-commerce, have sort of figured out, like, I just wasted a bunch of money on a project because now it's just incorporated in LLMs. So that's really kind of what the book explains. It takes longer than 20 minutes. But you could probably read the book in a few hours. And I'm sorry about that.
Brent Peterson (08:47.775)
Hahaha.
Brent Peterson (08:53.036)
Yeah, that's awesome. I think there's a couple of things to dig into there. I think maybe even people knowing what to use AI for, are we going into that? There's so many use cases, and the typical use case is creating social media posts. But that is such a side project of what people should be using it for. think Claude code is probably the biggest thing that's going to stick around.
Paul Byrne (09:10.809)
Yeah.
Paul Byrne (09:20.666)
Yeah, we use that extensively. Yeah, so that, yeah, and Claude Code, and there's another one I can't remember its name right now, but these CLI tools for developers that are agents that can produce sub-agents that can kind of work together. Those are, you know, probably the, those have made AI tools useful for developers.
Brent Peterson (09:22.721)
Yeah.
Paul Byrne (09:47.623)
and in fact, we just, recorded a podcast earlier today with, our CTO and one of my product managers and, you he, he went over kind of how developers use Claude code and what it can and can't do, you know, and, sort of what we expect from it. But yeah, you hit it on the nose. Like Claude code is, is, is big like that. That definitely changes things.
It doesn't mean I can replace 10 developers with Claude code, but it does mean that my senior developers and my leaders can move more quickly, especially in code review. know, code review is, you know, checking for security and performance, et cetera. Uh, you know, it's, it's able to do that stuff. That's kind of tedium really quickly. And in fact, that's what LLMs are really good at. Like the tedium, right. Taking away the T.
Brent Peterson (10:37.9)
Yeah, yeah, they're good at the tedium, but they're also good at the confusion. And I was just thinking about this in my work, in my workout this morning on, you know, I built a little Claude code workspace where I have a sprint and I start my sprints and it creates, you know, my documents. And, this morning I, I do a retrospect every other Friday and I went and looked at it. It's created a ton of documents that mean nothing that are in all over, all over the place. And there's duplicates and like, it is just, and I have.
Paul Byrne (10:43.032)
Yeah.
Yeah.
Paul Byrne (10:58.532)
Mm-hmm.
Brent Peterson (11:07.21)
I have an MCP memory running and I write a contextual file as I'm going along. it keeps, you know, I can continue across context, but man, it's like, and I think about this in my head, like it is thinking in its own way and we need to think in our way. And I almost feel like we need to have a way to like have a skill that's a human skill that says humans aren't going to think with 10,000 random files in a folder.
We need the folders that are in a structure so we can understand what's there.
Paul Byrne (11:41.112)
Yeah, yeah, exactly. So another AI educator, you're probably familiar with Andrej Karpathy. He was in the news when the whole OpenAI thing blew up a few years ago and Sam Altman was there and he wasn't there and came back and everything. But Andrej Karpathy left just shortly after that. He was one of the founders of OpenAI and in my mind, one of the best educators about it, about what
what AI does and can do and especially the type one thinking models that we have. But his point of view and the way that we've kind of started designing our software is there needs to be like this human slider. So AI becomes more and more capable in certain ways and in certain things. And so you have to know when to insert the human into the loop within
a program. we've kind of done that. We start replacing things that humans used to do with a little bit of AI, right? And then you start doing more and more of that. And so we've actually started kind of designing our software around this idea that there's this slider and it's going to keep going up, at least for now. Because there a lot of headwinds for AI.
Brent Peterson (13:03.086)
Yeah, yeah, I mean, going back to the idea of humans in the loop, and even writing content, it's great. AI can create all your content. could write your book probably now. But is it going to write it in a way that anybody wants to read? That's the biggest problem. And especially in software, it's going to do it.
Paul Byrne (13:20.706)
Not really.
Brent Peterson (13:28.76)
but is it really gonna do it the way anybody wants to see it? know, like there's so many variables that have to have a human in the loop there. Talk a little bit about how you've been able to bridge some of that and what you talk about specifically in your book around humans in the loop.
Paul Byrne (13:38.328)
for sure.
Paul Byrne (13:47.192)
Yeah. So it's, it's a question of, you know, an example I could use is like the AI chatbot, right? At first, it seems really cool. You go to a website and, there's this chatbot and you don't have to wait for a human to show up. You don't have to wait until they're not answering 20 other chats, you know, to get back to you. Like you can get really useful information and, know, and often it's good enough for a lot of inquiries.
But then you get to things that it can't answer for you or that are clearly exceptions. And then it just starts spitting out garbage, right? And unless it, so what you have to do is you have to design your bot so that the bot realizes or your monitor of that bot realizes when this is something that requires, that is an exception and is gonna require a human and then transfer the chat over to the human.
So we're in the middle of, in the next few months, we're actually launching a new messaging service for a customer service for B2B e-commerce. It's specifically designed for that because there's tons of interaction that really just needs to be handled by a human. And it really was built on the idea of kind of empowering the human, like using AI to bring context in while you're talking to somebody.
And, you know, so that the human doesn't have to have 50 tabs open and, you know, be researching stuff and Googling like that can all happen in the background. So that's like, that's one way, you know, of kind of adjusting that slider. The other thing is in the way that we write code and the way that we do code review. And this morning as I was speaking to William, our CTO.
We were talking about how he's incorporated it into code review and he's built these various agents using Claude code that go through and check for all of the common things. And so before one of the developers submits something for human code review, has to go through AI code review. And the way he set up the agencies, they provide him with a specific report.
Paul Byrne (16:06.778)
and information on it, and then he knows how to go in and look. And he said, yeah, sometimes it gets it wrong. There are some false positives, et cetera. But for the most part, it really enables him to do more in-depth code review. And so it improves the quality of our software. And that's kind one of my predictions in the book as well, is that we're going to see a lot of improvement in quality. There's like the AI slop.
And then, which is asking the AI to actually generate something and create something. And then there's the AI code review or the AI quality improvement on the other side where the human's overseeing everything, but the AI enables that human to work more quickly and to do more.
Brent Peterson (16:53.996)
Yeah, I'll make two comments because I'm a big Claude Code user and I've created a Claudes comms, a Claude comms skill where if you have two windows open and they suddenly need to talk to each other, you can either copy and paste or I've created a common file in my Claude folder that they write to and then they talk to each other, which has really sped it up.
Paul Byrne (17:12.122)
Mm-hmm.
Brent Peterson (17:21.226)
sped up my work and I actually have them name themselves as well. The only downside is I get a lot of Claude Beethoven's for some reason. Like I just looked and I had like four Claude Beethoven's. I don't know why they, obviously I need a rule that says don't create the same name. But, that one, the other part of this is the democratization of doing a lot of things that people couldn't do. let's talk about coding because everybody can be a coder now, right?
Paul Byrne (17:29.562)
you
Paul Byrne (17:36.76)
Yeah, that's funny.
Brent Peterson (17:50.604)
you know, part of that is that they don't know what they're doing.
Paul Byrne (17:51.354)
Thank
Yeah. Yeah, they don't. They don't. In fact, I was on a recent podcast of Fires, founders firing on all cylinders. And I was talking to Christian Hamilton of Acoustic Orange and he was, he was, he's not technical at all, right? He doesn't write code. And he got into vibe coding and using the tools. And I think he had the experience that most people do that don't
you know, understand how to develop software. And that is he would get to a point where then he would just, he would just ask it to like to change the title on the page and you know, would nuke his entire repo, right? Start over. So, so the tools are getting better. I'm looking forward to a conversation with a gentleman that I was on a kind of on a panel discussion yesterday.
Brent Peterson (18:36.75)
Right? Yeah. Yes. my God. Yes.
Paul Byrne (18:52.022)
who's also not technical and he's taught himself not to use Claude code, but to use cursor in a very limited way. And he's got like this specific project that he wants to do and he's going to record the whole process. And that's going to be fun. Like I'm going to recommend that video when comes out. Cause first of all, this guy's pretty interesting and you know, kind of a transversal thinker, but also, you know, I haven't really recorded anything like that.
Brent Peterson (19:06.06)
Okay.
Paul Byrne (19:20.378)
But I'm guessing there will be a lot of foul language involved.
Brent Peterson (19:24.082)
yeah. I have a weekly newsletter called Tuesdays with Claude. I have a number of them. Well, I have at least one where I have a lot of emotion in it. And then I have another one. I do have one where I asked Claude to go ahead and clean up some of my files in a folder. And it ended up deleting about 1,000 files from my system for absolutely no reason.
Paul Byrne (19:31.374)
Yeah.
Paul Byrne (19:36.9)
Yeah.
Paul Byrne (19:44.388)
Yeah.
Brent Peterson (19:53.527)
I mean, thank goodness had a repo and I have a remote repo, but yeah, was just, I'm like, what? And I started going and restoring and I'm like, wait, no, let's just do a hard reset. This is crazy. I also have a violations log that I keep running at all times.
Paul Byrne (20:06.126)
Yeah. Yeah.
Yeah. So, so my point of view is that the, won't really be good at software engineering and creative software development and design until there are breakthroughs in type two thinking. And right now, but
The other thing that I think a lot of people expected and one of the reasons, well, you know, we're going to like the CEO of Anthropic Davos makes the comment that software engineering is going to be automated away in 12 months. I'm like, like, I don't know if you're just breathing your own exhaust or what's going on there, but it's clearly not happening. And they're, not going to get there. In fact, if anything, like parts of it are getting a little bit worse with the new models.
So, yeah, I think we've kind of reached the point of marginal returns on the LLMs, at least in terms of capabilities. Now, in terms of implementation, I think we have a long way to go and there's a lot of money to be made there. But in terms of raw capabilities, they're proving not to be that much better.
Brent Peterson (21:21.646)
Yeah, yeah, and my son is a front end developer in Magento. Yeah, don't say I'm sorry, but that's what he does for a living. And he's embraced cursor and using Claude code. And he said he gets four times as much done now. And if you don't embrace it, I think your boss is going to look at you like, should be using some of these tools to do the mundane tasks. He is much more.
Paul Byrne (21:29.818)
you
Yeah.
Paul Byrne (21:37.892)
Mm-hmm.
Brent Peterson (21:49.423)
Cautious of letting it do more than it he thinks it should where I'm a little more liberal on that So I think that I think you're right I think you hit you got it You know I Roy Rubin had said last year that websites are going to go away, and we're all just going to be buying through ChatGPT I also don't think that's going to happen So I think there's a lot of predictions coming out right now that we'll see what happens in the at the end
Paul Byrne (22:10.957)
I, yeah, and I.
Paul Byrne (22:16.502)
Well, and I think, you know, like there was a prediction that brick and mortar retail was going to go away, but that doesn't happen. you know, part of it is, is just the format. You know, a of people like to go to the store, browse it, touch the products, et cetera. Same thing on, you know, if you go to a website, an e-commerce website that is specialized, has good data, et cetera. Yeah. You're going to get a lot more done there. It's going to be much more satisfying experience.
because you don't know where, especially right now in the early days, you don't know where the LLMs getting their products are, the price is really good or not. You've seen so many misfires. I've seen LLMs, they continue to do this, like making up websites and their information because the training takes so long before a new model is implemented. You've got to train it and then they got to QA it and then they've got to...
Brent Peterson (23:01.425)
Right.
Paul Byrne (23:11.598)
you know, check it for security and all that sort of stuff. by the time you get that model, you know, it's six months behind. Well, prices have changed. Even if it's using the web search techniques, it's not even going to be aware of like new stores. It's sort of, yeah, there's just so many gaps there that I, I don't, I think we're a ways away from that. Now I do believe there's going to be a lot more commerce through LLMs than there is now. And honestly,
You know, Google's losing out on that. You know, there are lot of websites that are kind of decreasing in importance because of the LLMs, but replacement some things over a very long period of time, but it's not going to happen tomorrow.
Brent Peterson (23:57.283)
Yeah, think a lot of these, well, mean, I think everything's gonna exponentially get better. I also, I agree that we've kind of plateaued on some things and then we're gonna break out in other things. In the few minutes we have left, where do you think in the development space as an agency you should be sitting in terms of billing and is it bringing down the prices of software or custom spot software, especially for people or?
Is it still, you still have to have that high level person to kind of watch things and maybe they're a little bit more efficient.
Paul Byrne (24:33.188)
So the rates don't come down. We're getting more done with less time, and we generally charge for our time. I'm looking at ways to move away from that model, not because I'm greedy and I want to charge the same amount as we did before, but because the tools do come with certain costs. And there's a certain amount of expertise required to use the AI tools that's a little bit more expensive.
than say a junior developer. we haven't found the model on that. I think we're going to end up a little more fixed price on a lot of projects. We've already gone to fixed price on discovery. So we'll probably do some of that as well from an agency standpoint. I think that...
Every part of the agency has to embrace it, not just the developers. All of our business analysts, our project managers, everybody is using it. And I think we're going to have to become accustomed to running more quickly than would have been comfortable a year ago.
It's the new space we're in. The river is now moving more quickly. I guess you could say we're going into the rapids, right? And we just have to adapt to that, like we always have. It's just this has kind of accelerated things at a quicker pace, but in part because of all the hype. There's been so much investment and so much
you know, money going into creating this model. My biggest fear is that we're going to become dependent on them and then they're going to start charging us what it really costs to generate tokens, which is what the models do. Uh, and I'm, know, right now, I don't know what we're paying 20%, 30%. Uh, nobody's making any money in that space. So, you know,
Paul Byrne (26:34.522)
And the rate of innovation is slowing down and there's more pressure. You look at OpenAI, they're practically bankrupt. There's all these different headwinds that I think may increase the cost of AI. And I just hope it's still worth it at that point.
Brent Peterson (26:57.218)
Yeah, I do think that contextual length will keep increasing and the amount of tokens we can use in a chat will keep increasing, that type of thing. But you're right. Yeah, I think that the cost will probably continue to increase as well, right?
Paul Byrne (27:16.152)
Yeah, well, there are even a lot of headwinds for increasing the size of the contextual windows, because increasing just a little bit doubles the cost, can triple the cost. It increases the size of the model that you have to build. So it's not just a matter of adding more memory. It's, I have to have that many more inputs on my entry nodes in my.
you know, in, in my large language model or my neural net, right. And like, that means like just adding a few more nodes, I got to expand that thing enormously. And so I, it's yeah, you know, it's not just increasing server memory and that stuff. I think one thing that is going to make, continue to make a difference is moving as much of the, of the actual AI to the edge.
So whether it's actually running on your machine and then collaborating with the LLM on the back end, think there's going to be, yeah, that there's probably some juice to be squeezed out of that over the next couple of years.
Brent Peterson (28:29.954)
Yeah, my hope is that we can run the models on our own machines, but my hope is there'd be things like Claude code that you just download and keep running that is running internally and you're not sharing and it's using its own internal resources. And certainly it's not going to be as big, but it's also going to learn from your project. And hopefully you can download as much as you need. And I do believe it's probably going to go that way. That's kind of where IDEs came from, And even Photoshop, that's kind of how it...
Paul Byrne (28:38.756)
Mm-hmm.
Paul Byrne (28:55.352)
Yeah, absolutely. Mm-hmm. Yeah, that's an excellent analogy, Photoshop.
Brent Peterson (28:59.404)
It started.
Brent Peterson (29:04.169)
Mm-hmm. So Paul, we have a few minutes left as it close out. You know, I'm going to give you a shameless plug. I'm assuming I know what you're going to plug. But go ahead, anything you'd like to plug today.
Paul Byrne (29:10.05)
Appreciate it.
Well, it just so happens that I released a book last week, and it's called Adapt or Die The Real AI Playbook. You can get this, really the only place you can get it right now is at Amazon. But if you're looking for development work and you reach out to me, there's a good chance I'll send you a signed copy for free. Would be more than happy to do that for your listeners. And just, you
If it's useful to you, let me know. me a, you know, best thing to do is give me a review on Amazon because I really, I've got the second book in my head already. They told me this was going to happen, that as soon as I released my first book, I was going to have my, you know, the second book ready to go. And that's exactly what happened. So, you know, I'd love to learn from this and make something, you know, even better, more useful to as many people as possible.
Brent Peterson (30:05.698)
Perfect, yes, and I did get it. I have not read it yet, but I have a 13 hour flight coming up, so I plan on getting some reading in on the airplane. All right, Paul Byrne, it's been such a great conversation. Adapt or Die is your book. It's available on Amazon. We'll get those into the show notes. Thanks so much for being here today.
Paul Byrne (30:11.226)
There you go. Perfect opportunity. Yeah, cool.
Paul Byrne (30:19.13)
Peace, Brent.
Paul Byrne (30:25.22)
Thank you very much, Brent. It was such a pleasure.