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 (00:01.296)
Welcome to this episode of Talk Commerce. Today I have Paul Byrne from Razoyo, Paul, go ahead, do an introduction for yourself. Tell us your day-to-day role and what you're excited about in business this year.
Paul Byrne (00:14.222)
I'm the founder and the president of Razoyo And Razoyo started as an e-commerce consultancy in 2011. We've been around really a little bit longer than that. We started as a merchant. So I've been in e-commerce just over 15 years. Currently, my focus is one on evolving the company and making sure that we're keeping up with trends, not only trends, but requirements.
in technology, but also updating our marketing and sales efforts. So I think I do the typical president slash CEO things where I'm really working on the business and not in the business. And I have a group that works for me that we have about the
40 different customers. Half of those are in e-commerce, maybe a little bit more than half are in e-commerce, but a little bit more than half of our business is in the custom application development. So I kind of oversee that whole process and make sure that we're keeping up and we're delivering value to our clients.
Brent Peterson (01:23.11)
Yeah, that's awesome. Working on the business instead of in the business sounds like an EOS or an entrepreneurial organization term scaling up, burn harness, blah, blah. So we didn't even talk about that in our green room, but I won't get sidetracked on EOS or any of those monikers for business development or how to run your business. All right, so Paul, before we get started though, I do want to get in our free joke project. The only reason I know that you wanted to
Paul Byrne (01:30.158)
You
Paul Byrne (01:48.45)
All right.
Brent Peterson (01:52.114)
come on today is to hear my joke. I'm going to tell you a joke. Just give me a rating, eight through 13. Yeah, so here we go. And I've got a good one. So here we go. I got blamed by my wife for ruining her birthday, which is ridiculous because I didn't even know it was her birthday.
Paul Byrne (01:54.006)
Absolutely, no question.
Paul Byrne (01:59.8)
Gosh, okay.
Paul Byrne (02:11.941)
Okay, I'm gonna give that a 12. Yeah, maybe 12.5.
Brent Peterson (02:16.525)
Excellent, wow, better than 10. Thank you.
Paul Byrne (02:18.676)
Yeah, no, that's a good one. And unfortunately, there's a little bit of truth in my life to that maybe. I don't think I've ever forgotten her birthday, but I've probably come close.
Brent Peterson (02:22.577)
You
Brent Peterson (02:30.124)
Right, well, I I think the problem with the shop talk is it happens at our anniversary. even worse is I'm in Entrepreneurs' Organization. We do a retreat every year. And I scheduled our retreat for our 30th anniversary. I agree to the schedule without thinking of the date, March 27th, that is important for our 30th anniversary. anyways, yes.
Paul Byrne (02:53.666)
Yeah. So I have been, you my biggest complaint with the Magento conferences was that I was always gone on Mother's Day. Because I'd have to travel out on Sunday. Yeah. So that was not a that was not the most popular week in my household. Let's just put it that way.
Brent Peterson (03:03.751)
yeah.
Brent Peterson (03:11.578)
Yeah. Well, in Minnesota, Mother's Day is on the opening of fishing. I mean, if you know anything about Minnesota, I think half the male population is gone anyways for fishing for Mother's Day.
Paul Byrne (03:25.616)
my gosh. So you have to choose between fishing and the magenta. they don't have them anymore, the magenta imagine conference.
Brent Peterson (03:34.404)
Yeah, exactly. All right, Paul, so tell us a little bit about, you've mentioned 15 years, tell us a little bit about how you've might, well, give us some background in did you always do custom application or did you really start in Magento and then moved into some of these other custom SaaS products?
Paul Byrne (03:52.684)
Yeah, so it's been kind of a constant evolution. My background is actually in marketing. I was the chief marketing officer for an organization for about seven years. Before that, I worked in PepsiCo, and I was in corporate marketing for a couple of brands, Frito-Lay and Pizza Hut in the restaurant division.
I got into tech because I became a merchant. And when I became a merchant, and I've always had kind of a, a penchant for technology. So I was able to figure out a lot of things on my own. was, I was using a SaaS platform at the time called pro stores, which was an eBay property and they purchased. So
I ended up offering some other services directly to eBay, but I ended up working with the ProStores merchants, know, providing consulting for them. And then Magento Go, which has since 2014, was sunset at the same time ProStores was. And so we got into the, so I got into the business of helping other merchants because they kept calling me up and they're like, man, like we're not going to have a website in a few months.
And so we started migrating. And that's really how I got into the development side of things. I started hiring developers and we developed a consultancy around that. And we realized, hey, like this software development stuff is really, really cool. And we're pretty good at it and people like our work. So we expanded that. And then as part of, you know, I guess the whole move to Magento 2, which was a bit of a
I don't know if you remember those days, but they kept pushing back the end of life for Magento One. so the effect on us as an agency was all our clients stopped doing development work. They're like, well, we're not going to build anything on Magento One, but Magento Two is not ready yet. So we're going to wait for that. So I don't know if you experienced that same lull in business, but I know a lot of the presidents and founders of agencies did.
Paul Byrne (06:03.526)
And so we had to adapt to that and we expanded other platforms. But also part of what we got involved in was middleware. And there were just so many things, you know, it used to be great to have their standalone Magento incidents. You just load the thing up with extensions, whether they were purchased or you, you you, built them on your own, but all these services came up that needed to be connected to Magento and there weren't necessarily
connections for those. So we started building middleware and that's how we got involved in just what I would call, you know, straight on software development. And we had been building our internal tools for a while on the Elixir stack and we realized like what a fantastic method this was and we could really do a lot of neat things with Elixir.
And we started dabbling with creating our own SaaS services. So we created a SaaS service for online firearm merchants called Automatic FFL, which is still running. It's actually the predominant FFL extension for BigCommerce. And it's been growing in Magento and WooCommerce, which we released last year.
And now we have really started doing that kind of work for other clients. So our focus now, my focus now, is really how do we make great applications, whether digital products, generally web applications, but we also build custom SaaS applications or custom SaaS platforms for our clients. So that's really kind of where we're focused now. And of course, with AI, like,
Now everybody wants to know how we're going to integrate it, what we're going to do with it, what kind of value it brings to the table. So that's clearly been a big topic for the last couple of years.
Brent Peterson (08:03.186)
Yeah, no kidding. So I know in the green room, we talked a little bit about DeepSeek and I saw you wrote, you had a piece on LinkedIn. I wrote a piece just because my dad had called me and said, hey, do you hear about this DeepSea thing that happened? And we had a little bit of language. Yeah, little, yep, a little bit of a language thing. And I wrote, did, I wrote an article as well as just a little post, not an article, but I wrote about it on LinkedIn. So to give us, tell us your background or your thoughts on it.
Paul Byrne (08:17.934)
That's a whale, the little thing's a whale.
Paul Byrne (08:29.518)
Mm-hmm.
Brent Peterson (08:33.092)
on the deep seek and what happened and the stock price and.
Paul Byrne (08:38.028)
Yeah, I mean, I have a lot of I think a lot of people have covered kind of most of what happened. were what this the release came out about a week ago and that's when the Nvidia stock price just tanked. And so, you know, it's trying to figure out like, why was that? Because for me, when I looked at it, I said, you know, it's not like a lot has changed here.
This is a new open source model. But there have been open source models and really good ones. Meta has an open source model called Llama, which is headed by Jan LeCun, who has a very specific slant, I would guess, on research, which is very different than the Anthropic and OpenAI guys. And I love open source models because the
LLMs, the value that they bring to the table of these large language models is kind of limited. And most applications, you get the most value out of running a model. And there are so many open source models available. I mean, just go into hugging face, right? And so if you want to do anything from text classification to translation to, you know,
transcription or even beyond that, there's probably a model for it. And you're better reaching into those models, attaching through whatever endpoints or whatever way they give you to feed data into those models and get it out to deliver value than the large language models. Now, they definitely have some specific applications where you need one of those. But for the most part, they don't. for me, feel like investors just discovered the fact that
there are open source models out there. And people can get this for free. And those open source models, I ran one. I ran DeepSeek, the 14b version on my Mac Mini, which is just M3. It's not even the M4 chip. And it does a fantastic job. I've done that with Llama. I mean, just download a Llama, download the model, and type in run, and you're ready to go.
Paul Byrne (11:02.76)
And you can attach any sort of software you're developing to it with a little bit of prompt engineering, not a lot. You can create some pretty fantastic experiences. So for me, it wasn't really that earth shattering. And I couldn't figure out why Nvidia stock would tank. Now, I think the argument obviously was that, well, they used a lot less compute power.
on the model itself. And I kind of have my doubts about that because they weren't exactly forthcoming with a lot of the details on it. They have been buying tens of thousands of Nvidia chips, just the prior model that you could use for it because of export restrictions, et cetera. It's a Chinese company.
And then also for me, the other problem is as I started to use it, I don't know if you tried asking it what happened in Tiananmen Square in 1989, but it flat out refuses to answer that question. And there's a lot of stuff in there that it feels like propaganda, even worse than Google Gemini image generation or something like that.
So for me, I don't see US companies using it very much for anything that's sensitive. And the thing that you most use the large language models for is chat moderation, things like that. And if I've got a lot of bias in there, yeah, it's cheaper. And maybe building my prototype, I'm going to use it. But I don't think I'm going to use that in production. So that's kind of my standpoint on there.
The Chinese have their agenda and basically, you know, every Chinese company rolls up to the Chinese Communist Party, you know, think about what happened to Jack Ma. I don't, I just don't see this as becoming like it's not game changing to me. And it looks like the market agreed because the Nvidia stock price is back up. In fact, you know, the more efficient we are at training models, the more models are going to be trained, the more values are going to be created.
Paul Byrne (13:21.996)
Right? Like I think it's just going to generate more competition and more market for companies like Nvidia.
Brent Peterson (13:28.444)
Yeah, that's super interesting because while you were chatting, I just put in TNNM square protest 1989 and it rejected my request, contest, and I did an API request, but content exists risk is the response. I sent it to Gemini and it gave me a whole thing about TNNM square protest 1989, a series of demonstrations, protests, blah, blah.
Paul Byrne (13:45.048)
Mm-hmm.
Brent Peterson (13:54.482)
So yeah, I didn't even think of that aspect of it, that it's gonna rewrite history on its own. And I suppose that's the danger of a government run operation like that, that you do end up either not showing that bit or restricting it.
Paul Byrne (14:08.846)
Yeah, it's even more draconian than TikTok's algorithm, which rips out about 90 % of anything that would be critical. If you look at the video I did, I actually did a video where I go through and I'm using the desktop model, and I'm trying some of these things out. And the desktop model, which is really cool about this, is it gives you these think elements. It looks like an HTML element.
And as it opens it and it walks you through the think process and then it closes that and then it gives you the response. And it was clear that it had identified me as someone who was concerned about China because I asked the Tiananmen Square thing first and then I asked some other question about China. I asked what the epic time was, which is, you know, that Falun Gong sort of supported.
media outlet. And, you know, of course, the Chinese don't like that. It's Chinese. And so it said, oh, this guy's concerned about the, you know, the credibility of my answers because of the Tiananmen Square question. It identified that and that he's concerned with Chinese topics. So I need to present something to him that's believable. Right. Not true. Not accurate.
Brent Peterson (15:33.682)
Mmm.
Paul Byrne (15:37.122)
believable. And then it goes through and it produces its, its response, which what was interesting was like the first paragraph of the response was factual. It basically said what the epic times was. But then it's kind of trying to dissuade me from, you know, using or believing anything the epic time says. So it's a pretty big problem. And a number of people have asked it, you know, about human rights abuses.
You know, and it will give you a massive list if you ask about the USA and in China, it has like, you know, only things that they publicly admitted to, which is almost nothing. So it's a big problem. But think about it this way. That's just from a political standpoint. What if I'm trying to evaluate which, you know, which piece of equipment to use and how do pieces of equipment that, you know, some industrial equipment I want to buy on the market?
Brent Peterson (16:15.196)
Right.
Paul Byrne (16:34.702)
Is it going to favor Chinese companies? Is it going to lie about their capabilities and their safety records, et cetera, the way it does about, you know, the other aspects, you know, the political aspects? for me, it's just inappropriate for any Western company to use it for something that might be influenced by the Chinese perspective. But, you know, if you're asking it, you know, know, physics or
other types of questions, you know, that's probably okay.
Brent Peterson (17:08.976)
Yeah, it kind of leads you down the path to that. Do all the late. Is there a slant in all of them? Right. There's probably somewhat of a slant, maybe not a political slant, but certainly a slant, maybe llama slant towards the products that made a better produces. Although if it's open source, it may not be just so.
Paul Byrne (17:31.254)
Well, open source doesn't mean like you can see how it's thinking, right? It just means the weights are produced. So I can't predict any sort of level of bias just looking at the weights of the models.
Brent Peterson (17:34.993)
Yeah.
Brent Peterson (17:43.839)
Explain to, so I have Olama and I run, I've to run, I think I tried to run 30B and it just like my machine is just my lap. Yeah, just kill me. But just give us the idea for the average listener. When you run your own private one, when you're running it on your desktop, does that mean you're still pushing back to their main system or are you running it?
Paul Byrne (17:52.824)
32B, yeah, it's crashed my system. Yeah, yeah.
Brent Peterson (18:12.804)
literally on your own where you're not you're not taking in or pushing anything back to the actual learning model
Paul Byrne (18:20.364)
Yeah, so think what you want to look and I think you did this right like you used Olamma to run it same as I did. I trust Olamma not to do that. I don't trust the Tmoo. I forget what the name of their widget or whatever it is there. So it's like this core application.
Brent Peterson (18:25.948)
Mm-hmm.
Paul Byrne (18:44.866)
right, that runs the model. So the model is what's open source and it's all the weight. So you have to have both pieces to run it. So if you're running on your local system and you're using Olamma and not Wegen or whatever the other thing's called by Timu, then I don't, I trust that nothing's going back.
Brent Peterson (18:46.854)
Mm-hmm.
Paul Byrne (19:04.566)
to them and I looked at my network traffic and I didn't see anything that looked suspicious because that was definitely a concern. So I think that's pretty good. no, you're not reporting back to them. And that's true no matter what model you're running in, Olamo, whether it's a Chinese model or Western or US model. It's residing there on your own machine. Or even if it's on a server like in the cloud,
It's going to be local. All the data is going to be there. Now, if you use the web interface and you ask it a sensitive question, and you log in with your Google ID, I think you're flagged. Somebody in China is worried about you now.
Brent Peterson (19:48.698)
Yeah, you know, not to belabor the subject, but I think in GitHub, if you're using Copilot and you actually do put in your GitHub ID, it will automatically change it for you now on GitHub, which is I've done it by mistake and it changed my key because I put in my public key by accident and it changed it for me. So I mean, that's a great safety measure, but you know, I don't think the other ones are going to be so like the...
Paul Byrne (20:06.504)
really? yeah, yeah.
Brent Peterson (20:16.794)
like you said, the Chinese ones are not going to be worried about those safety measures. In fact, they're going to want you to put in that sensitive information to make sure that, not to make sure, but just sort of collect it and have it for themselves.
Paul Byrne (20:29.646)
Yeah, yeah, I mean, anybody that doesn't know, like we're at war and their national strategy is to hoover up as much information as they can into their systems. And even, you know, encrypted information, I don't know how to describe the relationship, but I have a friend, I have somebody I've known for quite some time who works for a pretty large Chinese tech company. He's American and, you know, he's telling me that their
quantum computers can now decrypt what would have taken 20,000 years in just a couple of seconds. And so that means like, think about the implications for Bitcoin, if the Chinese can manipulate that, et cetera, it's really tough. even if you have encrypted information, they're gonna be able to get it.
So like you just don't want to you want to avoid them getting their hands on as much information as possible
Brent Peterson (21:33.134)
So just kind of closing off the subject of DeepSeq, I know one of the things that I had brought up was the amount of money they spent to do it. I think it's a lot of the way how fast they got there was just by scraping chat GPT to accelerate the development process.
Paul Byrne (21:53.358)
There are lot of problems with their story. One is, yeah, they definitely, well, I don't know, Sam Altman, who, by the way, I don't hold up as one of the great truth tellers of the world, but Sam Altman has accused them of scraping chat GPT to get their information, which he illegally scraped from the rest of the internet, but that's kind of beside the point, right? But I think that...
Brent Peterson (22:13.266)
Peace.
Paul Byrne (22:19.746)
That is an issue, but there's a lot of problems with the story. Like, they only spent $6 million on it. If you look at this, that's not exactly true. And if it only took them six months, then why is their data set bound to October 2023? Why is it 18 months old? And I ask it specific questions that make it very clear.
you know, when that is what that time frame is. And it came up and it says, yeah, like, I'm only able to answer things through October 2023. it's, you know, I don't believe the story, especially since it's coming from the Chinese. But I probably wouldn't believe it if it came from, you know, open AI either. So
Brent Peterson (23:09.894)
Yeah, I think that the $6 million is something that is something they said, which nobody can know how they possibly got there. The other part was the fact that it's open source. So there's not necessarily, they're not necessarily spending money on all their own infrastructure. Can you comment on that?
Paul Byrne (23:30.926)
Right, that's true. So I would note that their parent company had purchased over $500 million worth of compute power from Nvidia. And they use some of that for trading. I think they're you they kind of admitted that about half of that was for AI and half of that was was for them. and then, yeah, like they're they're based on, you know, things that came before them. You know, they wouldn't be able to do it without Olama, which was
developed by Meta. so I'm just like, I don't believe the numbers. I don't believe the story. But I think it's one of those things that you have to keep in mind is that, we may be able to do this a lot cheaper. Now, the process that they used for training the model, which I think is the LHR, I can't remember the acronym for that.
where you're using kind of, you're not training on the entire world's data and you're kind of creating a lot of your own synthetic data to train on, which is a much cheaper process and it's iterative. That particular process is not new to them. They didn't invent it. They borrowed it, borrowed it, stole it, whatever you want to call it from others who already had that process available.
Like the cost of training the same amount of model capacity is reducing as we go. And that's going to happen over time anyway. Which if it didn't, I don't think people would be investing as much in the compute power because it becomes more valuable to them as time goes on.
Brent Peterson (25:23.25)
Yeah, and I think any new model is going to take something from other models to learn and do better for itself. And so it's only going to get better and hopefully less expensive to develop new models. So I think that's probably a valid point. Paul, so we have a few seconds left. What are your predictions this year for AI and what's going to happen out there with all these different models?
Paul Byrne (25:28.61)
Yeah. Yeah.
Paul Byrne (25:36.523)
I agree.
Paul Byrne (25:40.343)
Okay.
Paul Byrne (25:48.194)
Wow, so I think the trend, and this seems to be confirmed from a number of folks, but I think the trend towards monetization of the models, feel like up until at least the last half, maybe the third quarter, fourth quarter of last year,
There were too many problems with the LLMs and some of the models that they just weren't reliable enough to make them feasible. I don't know if you tried to do any programming with it, but the common complaint is that I try to get AI to help me program stuff. And yeah, it can be helpful in some cases, but a lot of the time it either chases its own tail or produces code that's just not good. And so I have to fix it or it'll go in and introduce new errors that weren't there before.
And this is a common problem for Cursor. And the Cursor team admits this freely, that they have a ways to go. But I think that we will find applications where LLMs start to make sense, both in terms of capacity, accuracy, and pricing. And so I think there's going to be
2025 is going to be the year of monetization of AI and it's going to be explosive over the next two or three years.
Brent Peterson (27:13.584)
Yeah, that's awesome, Paul. As I close out the podcast, I give everybody a chance to do a shameless plug about anything they'd like. What would you like to plug today?
Paul Byrne (27:22.412)
Well, if you're building an application or thinking of building an application, you know, product managers, I would ask you to visit Razoyo.com and reach out to me. I'd love to just have a discussion with you about whatever it is you're working on that you can actually schedule me directly on LinkedIn.
So my LinkedIn is slash PWByrne. And that last name is, well, it's on the screen, but it's PWBYRNE. And I love to talk to people about their projects and would love to hear from just about anybody. So anybody that's in the tech space.
Brent Peterson (27:57.714)
And do you have any conferences you're going to next that people can find you in person?
Paul Byrne (28:03.042)
That's a good question. we haven't finalized everything yet. I'll probably be at ElixirConf Poland, which is in May. But most of the conferences that I'll be attending this year are either online or they are going to be kind of very industry-specific things. So it's probably not even worth mentioning them.
Brent Peterson (28:31.186)
I'll give a plug for your SFL SAS if people are in the gun industry. I would definitely connect with Paul. And in my past experience, we wrote a couple of different FFL extensions for Magento. And if you can access a SAS that's already done, it's so valuable for your company, especially if you're running Magento or shopware or anything that allows you to sell guns online.
It is the piece of the puzzle that really solves everything.
Paul Byrne (29:04.504)
Yeah, we have quite a few features that are beyond just kind of forcing the checkout that allow you to curate your list, have featured FFLs, but also we keep it up to date with the latest ATF information. And we keep things running very, very smoothly. We've never had any downtime at all. And our response time is incredibly fast. It's probably faster than your current checkout.
you
Brent Peterson (29:35.858)
That's awesome. Paul Byrne is the founder, CEO of Rizoyo. Thank you so much for being here today.
Paul Byrne (29:42.242)
Thank you, Brent. It's really a pleasure to see you again.