#503 From Data to Differentiation: Ron Green on Building Real AI in the Enterprise

[00:00:00] 

Mehmet: Hello and welcome back to a new episode of the CTO Show with Mehmet today. I'm very pleased joining me, uh, Ron Green from Kungfu ai. You are the CTO there, Ron. So thank you very much for joining me. The way I love [00:01:00] to do it is I keep to my guest to introduce themself. Tell us a more about, you know, your journey, your background, and what you're currently up to.

And then we are gonna take it. From there, just, you know, a hint to, to our audience, we're gonna talk a lot, as you can imagine about ai, about the trends and, you know, some hints, uh, for entrepreneurs maybe also in that domain. So, Ron, the floor is your, 

Ron : thank you. Thank you so much for having me on. I'm, I'm really delighted to be here.

I'll, I'll give a real, real quick, uh, brief background on myself. I've, I've been in the tech industry now, um, over 30 years. Kind of, kind of amazing to say that. Um, I've got a background in computer science and I did a, a master's in artificial intelligence way back in the nineties. I. I have people every now and then ask me, you know, they're amazed.

I've been, you know, in AI since before chat, GPT, and I'm like, man, I would, you know, I, I was back in AI when, you know, Clinton was president. So been doing this a really long time. Um. I call myself a serial entrepreneur. I've done nine [00:02:00] startups. Um, most of those on the product sides, Kung fu ai is a, uh, management, consulting and engineering firm.

So we basically help, um, enterprises with their entire AI journey, whether it's, you know, figuring out how to get started, how to hire, how, how to build a AI strategy roadmap, how to, uh. Execute how to actually build custom models, how to take advantage of the proprietary data they have to build, uh, predictive systems or generative systems, uh, whatever it may be.

We, we really can help them across that, that entire journey. Um, and, you know, on a personal basis, I'm just as happy as, as possible to be professionally, um, after working in AI in the nineties. And, uh, you know, seeing the promise, but never really quite being able to build production systems, right? We, we, we always lack the data.

We always lacked the compute. We, we knew we needed [00:03:00] more. I don't think we knew how much more we needed. And, uh, uh, we've spent the last eight years at Kku ai, you know, helping our clients build production grade systems. And so I consider myself just very blessed and lucky to be doing, uh, professionally what I've always dreamed of.

Mehmet: Great, and thank you for being here with me today. Ron, just out of curiosity, quick question, like why kung fu? Like why the name? 

Ron : Yeah, that's a great question. All right, so we, we, you know, first off you have to get a name that ends in ai, so that's gonna limit a little bit. But we, we also just wanted something really provocative.

And something memorable. And a lot of people think that kung fu means, um, like martial art, but it, it doesn't, it's actually Mandarin and it means, um, any skill acquired through like hard work and discipline. And we thought that was a perfect, uh, encapsulation of what it, what's needed to see succeed in artificial intelligence.

Mehmet: That's interesting. Yeah. But you know the name. Even when the [00:04:00] team reached out to me, I said, okay, what these guys, they do. But of course I then checked. But to your point, yeah, these names should be, um, remember also as well. Exactly now to to to what you do today with companies, uh, around, um, so you help them, as you mentioned, building the strategy, like taking that you, you talk about data and we always.

You know when I discuss with my guest about ai, so the first thing comes to mind, of course, data and data readiness. Now maybe kind of a traditional question, but because you've been in the field for a long time. We've been talking about digital transformation, getting our data for more than a, I would say two decades, very, very easily.

Uh, do you still think, you know, like companies struggle in that field or No. Now, because, you know, with the AI imperative coming even from board level, you start to see companies [00:05:00] are more ready, at least from data perspective, so they can. You know, take that to the next level. 

Ron : It's a great question, and it's a little bit of both.

Um, you know, uh, the digital transformation initiatives have really made it much more easy to go into companies and it, it's pretty rare. We'll walk into a situation where. They are not either fully in the cloud or, um, are, are in the, in the midst of some type of migration. So that helps a lot right there.

The, the main thing we run into though, is very often the data that is necessary to perform the, the type of, uh, AI modeling to build the, the type of AI systems that companies want. They may not have it. Um, either collected or um, available in sort of, um, volumes that are necessary or it's, it's spread out.

And so, you know, the dominant technique within [00:06:00] artificial intelligence right now, of course, is something called supervised learning where. You, you, you, you build models and you show them a bunch of examples and they're labeled examples. Meaning, you know, here's an input and you know what the right answer is.

So if you're trying to detect fraud, you would go get a bunch of, um, let's say you're trying to detect credit card fraud. You would collect a bunch of credit card transaction history and you know which ones were fraudulent and you know which ones were legitimate. And you train a model to learn to identify.

The traits based upon, um, things that maybe even as humans, we, we don't know. Um, the challenge is we will very frequently walk into situations where. Not all of that data's available. Or sometimes they'll have expert processes where humans are performing tasks, but really all of the knowledge is in their head.

And so, um, the, the, the good news is all of those systems can still be built with ai. You [00:07:00] can augment humans. You, you can eventually get there, but it does mean, I'd say, you know, probably about 20% of the time, um. The enterprise needs to start collecting that data. Um, or, or, um, um, um. Pulling that data together in a way where it's unified for training.

But for the most part, just like you said, the good news is most enterprises have gone through the digital transformation and they already, it's, it's, it's they already, and there's so much low hanging fruit out there regarding AI initiatives that it's very rare while walking into a company and we can't find literally hundreds of meaningful AI initiatives that they could, uh, take on.

Mehmet: That's cool. Uh, Ron, now maybe a question I never asked before, but it came to my mind because, you know, some, some of the stuff that you do and, uh, what you just mentioned, uh, let me wonder. So we know there are like [00:08:00] two aspects. So there's the business aspects of adopting an AI strategy and there's, you know, the hands on, you know, what, what we do from adopt systems, maybe implementing LLMs or whatever the technology is now.

What have you seen more successful with the businesses that you work with? Is it like when the imperatives come from up to down, let's say the board say, Hey guys, like you know, everyone's talking about ai. We, we need to start to work on ai, or is it the other way around where the people who are on the ground, they start to understand the business cases?

Is that maybe I, and then they go and they suggest to. To, to the board, you know, the, the management. Yeah. Like, okay. So, so which approach have you found? Or is it like a blend, a mix of both. Like what have you seen successful? 

Ron : All right. Nobody's ever asked me that question before and I love it because it is very clearly the second.

It's when it's sort of bottom up. We, when, when, you know, we've been doing this for like eight years now. And [00:09:00] all the, all the clients we would work with in the beginning stage of our, of our company's existence would, they would come to us and they would say, well, first off, they had to be early adopters.

They had to really have a hard problem they couldn't solve with traditional methodologies. And they would come to us and they would say, Hey, would it be possible to, you know, build a computer vision, uh, system to detect manufacturing deficits or something like that? And we would say, yeah, that's absolutely possible.

Then chat, GBT comes out. And everything goes crazy. I'd say for a couple years we, we were, we were getting, um, inbound, uh, requests all the time. Like, Hey, can you build us some ai? And we would say. Yeah, sure. Well, you know, what are you thinking? And they would say, oh, I don't really care. But the board is telling us we need ai.

We need ai. And that is a terrible way to go about it because, um, you are essentially, you know, you, you, you have a hammer and you're looking around for nails and everything starts to look like a nail. And you can waste a lot of [00:10:00] time and a lot of money doing that. What's really much, much more effective is, and this is what our strategy team does, is when our strategy team embeds with the company.

And it can be six weeks, it can be three months. De depending upon the complexity of the enterprise, they will go and talk to stakeholders across different departments and surface, um, all the different initiatives. But then they're also vetting the data availability because, you know, it's never been more true that garbage in, garbage out, uh, regarding data.

Ai. Um, and then those initiatives then can be vetted for our ROI, meaning don't just go do it because it's cool or interesting. Um, make sure it has a real business impact. And over and over and over again, we see that it's not the top down ideas, it's the bottom up ideas where all the real valuable initiatives come from.

Mehmet: And here, I think this is where the company culture will, will [00:11:00] play some role also as well. Because, you know, if, if the culture is kind of a, you know, that doesn't, uh, embrace like, uh, people coming from the field telling their first line managers and then these first line managers, they take it up. So if you don't have this culture, like I think things might not move at, at a fast pace.

The other thing like, but. We know for a fact that there is this fear among even people who are talking about, yeah, AI will replace the blue, blue collar jobs. Now we're talking about AI replacing the white collar jobs also as well. So the balance between someone, like let's say if, if I'm someone who works in the logistics department or maybe in, uh, I don't know, operations department, and I see, you know, there's a bottleneck that AI.

Might solve and with automation, but look like maybe I will lose my team and if I lose my team, I'll not. So, so how [00:12:00] much is this, you know, balance between going from this bottom up approach while also having in mind. What we hear and see in the media and you know, from thought leaders sometime like, yeah, AI is replacing jobs, so there must be a balance there.

Ron, what is it? 

Ron : Yeah, yeah. You know, it's really interesting. Uh, you, you bring up this really great point, I think. I think most of us thought, you know, when we finally got to this point in the history of the development of artificial intelligence, it might be blue collar jobs that were most effective initially, but it's really the opposite.

It's, it's, like you said, it's white collar jobs and probably the most effective broad use of sort of AI automation was, is with AI coding assistance. I think it's like. Unbelievably ironic that, uh, you know, it's the, it's the very, you know, sort of software work, um, that makes [00:13:00] AI possible. That those positions are being, in some ways, um, uh, automated first in, in, in the world.

It's just unbelievably ironic to me. That said, um. I'm, I'm, you know, I, I say this only just slightly joking. Look, artificial intelligence is coming for all of our jobs eventually, right? Um, and what I mean by that is we are going to get to the point during our lifetimes where we have AI systems with really broad general capabilities, whether we'll get to artificial super intelligence, you know, AI capable of performing.

Not only all sort of human level intellectual tasks, um, and doing original scientific research, but going, you know, far beyond human capabilities. Uh, that's, uh, I, I'm actually pretty optimistic about that, but, um, that's less clear what I, the, what I do, the way I think [00:14:00] about this is this is this technology trend.

There's nothing that any of us can do to stop it because the benefits are so huge. Um, I really, truly would not be working in artificial intelligence if I didn't think that this is going to be a net benefit for society, for mankind, for the planet. And it's going to help, um, cure diseases. It's going to help cure hunger.

It is going to lead us into this world where, um. Eventually when we have humanoid robotics, which I think are gonna be in the next 10 years, where a lot of the backbreaking labor that humans perform right now can be done, you know, by machines. And it doesn't mean that all of these humans are going to be unemployed and destitute.

I think it's gonna free up once there is, once there is more. Of everything, um, uh, scarcity will become less of a [00:15:00] factor. And I know I'm getting, you know, very philosophical here, but I really do believe that whether all of this will happen in our last lifetime, I think this is where we're headed. And then, so to kind of come back as far as, you know, individual workers today, this is what I would say.

Um. At Kung fu ai, we're really big believers in building AI systems that augment humans that don't replace humans. And we al I would say, almost without exception, all of the projects that we've done and we've done. Well over 120 production deployments at this point for companies, um, almost without exception, if there's a human involved.

Um, we built a human in the loop system where the AI can do the boring, repetitive tasks or it can do the tasks. That, um, allow a human to essentially get a second opinion, but ultimately the human is in the loop vetting the decisions, and there is a checkpoint on the [00:16:00] system. And so I'm a real, really big believer.

You know, maybe 10, 20, 30 years from now we can move away from that. But at this point, I'm not really interested in building systems to take people's jobs away. I'm build, I'm interested in building systems that will make. Uh, their jobs easier and make them more effective and more efficient. 

Mehmet: Yeah, this is what I say.

Ron and I repeated it. A lot of times I write about it. Also, my humble opinion is go back to what humanity used to do before the industrial revolution. And I say like, do impactful thing. And I ask people like, have you ever tried to go to a farm one day? And they ask me why. And they say, because you know, when.

Because I, when I was a kid, like, you know, it's not like a big farm, but we, we used to go like to, to these small farms where you see the farmers and they work in agriculture and you feel, you know, this tie with what you do, and then you think, okay, I'm, I'm planting a seed today. It's gonna become, I [00:17:00] don't know, like.

A fruit, a vegetable, whatever is that, and then I gotta feed someone because I did this right. And you take care during that. I said like, wasn't this better than letting someone sit behind a screen all day, just, you know, doing repetitive tasks and they hate it. Let's, let's be honest, like people hate doing these jobs.

But yeah, they have no other choice. Of course, I know it's a little bit philosophical, but this is my take. Like I'm not afraid of AI taking. Jobs, I would say, because as you said, it's gonna augment, elevate, taking, taking us to the next level. So indeed it's not there. You mentioned few terms, which are interesting.

You talked about, uh, super intelligence, a GI, and I know you say like generative AI is just the opening act. Um, so what's act two? How does that look like? Yeah. And are we ready for it from business perspective? 

Ron : Yeah. Um, you know, I think, [00:18:00] um, the, I think most people would, would agree that, um, agentic AI is, is probably the really next big thing that's gonna happen.

It's got a lot of buzz, so it, it doesn't need me to hype it up. I, you know, I'm, I'm, I'm, I'm, overall, I'm very excited about Agentic ai and I think it's going to be very effective. We caution our clients strongly to be, uh, to be cautious on executing agen AI systems and, uh, frankly, generative systems for, for really the same reason.

Generative systems can be incredibly powerful, but for the most part, generative systems really need a human in the loop. Still, there really does need to be somebody in there, uh, vetting the output for the most part. Um, and the same thing with AG agentic ai. And the problem is the whole point of agentic ai, right, is that you are building, you are building, uh, [00:19:00] systems that have some autonomy that you can give them a high level task and they can work through uncertainty, make decisions, and accomplish goals.

And so, um, by the, by very definition, um, it will be making decisions not necessarily with the human involved. And so. The failure rate there is just too high for most enterprises. I think that that failure rate is coming down dramatically in the, the, the, not only the duration of the tasks that agents can complete is increasing, um, exponentially.

I think the most recent report that came out last month said that. Across sort of frontier models. Um, there's a 50 50% success rate on tasks that are roughly about 80 minutes, 90 minutes. So an hour and a half, that's a giant improvement. Um, but as you string together tasks, [00:20:00] um, the chance of failure increases pretty significantly.

So I'm very, very bullish on agentic ai and that is, you know, sort of deeply tied together with. Uh, reasoning models. Um, personally what I'm most excited about is what's going on with reinforcement learning on what's called verifiable uh, domains where. There. There we have empirical proof now that if you take base, you know, high quality based pre-trained language models, and then you use reinforcement learning or verifiable tasks like programming where you can give it a task, it writes a program, and then you can run the program to see if it works.

You can test it or you give it math questions and you can see if the answer is correct on those verifiable domains. Um. After a certain amount of training using reinforcement learning, there is this emergent behavior within these models. They literally develop emergent reasoning capabilities, and I think that's the [00:21:00] thing that I'm most excited about in artificial intelligence right now because that means that we have potentially, and, and I'm very confident about this, but I think we have potentially.

Solve the data issue because there is more and more evidence that using this technique we will be able to have models, not only solve problems, but then give themselves new unsolved challenges, test themselves, learn for 'em from it, and then enter this sort of feedback loop. And there's really, really strong evidence in just the last few months that that approach will work.

Mehmet: It's very exciting times with the agentic AI and, and, um, I'm fine with, with not only ai, with automation also as well, and I know it's not like agentic AI in the sense that you talked about I was doing a test because I never had. Time, but the other day I went, I wanted to go back technical. I come, [00:22:00] it's been a while.

I didn't touch anything and I just tried like few things with like MCP servers and, you know, all this stuff. And my mind was blown really because the, the amount of things you can do with the reasoning of the model, uh, it's not like. You know, you just say, Hey, like, go check my schedule and give me a summary.

You check, for example, uh, you instructed to go to the CRM, for example and say, Hey, like, find, you know, what was the main reason? Like some of these, I don't know, deals were stuck for a long time. Check the emails, check the, everything you know in, in our system and then. You know, it can come back with, with really good insights and it asks you if you want me to take actions, do you want me to send an email?

Do you want me to send a reminder? Should I put it as a task? And, and this is like, you know, you have literally a, a wizard sitting next to you. It's true. Tell you what to, so really, really, you know, fantastic and [00:23:00] amazing. And I'm excited also about it because. You know, the, the, you mentioned about like saving the time and I always tell people, like, think about it, uh, from, you know, time is money also as well and you know, if, if you are able to save time on some of the, you know, projects or tasks you, you are, uh, having on, on, on your plate.

So it's, it's absolutely fantastic now. You mentioned also about, um, you know, working with different businesses. So what industries are, let's say, seeing the most, let's call it quite disruption because of all of that? Ron, 

Ron : we, that's, gosh, another good question. We, we have worked, um, so broadly. In fact, we, for years we have thought about, um.

The need to potentially specialize and just focus on certain industries. Like do we, should we focus on, you know, [00:24:00] retail or you know, finance and banking or. Um, you know, I don't know, insurance or something like that. Um, I, I think because most companies are just now, even in, even in 2025, they're just now really taking their first steps into adopting ai.

We have found really broad, um, um, um, benefits across all verticals. There's really no industry. Um, that I have experience with yet that, um, can't benefit from it. And, and it, and if you step back and think about it for just a little bit, it makes perfect sense. What is AI bringing to the table? It's bringing the ability to perceive, uh, in the way that humans can so understand images and video and, and text and, and speech for the first time well.

That's amazing and what business couldn't leverage that. But it can also [00:25:00] find patterns and identify, um, uh, really complicated, um, correlations that might be beyond not only humans, but other existing statistical approaches and that those insights can be leveraged for your business. Again, you know, who wouldn't want, want that?

And when you couple this. These two together with many other things. We're seeing adoption everywhere. In fact, you know, even within, let's say agriculture, um, uh, uh, which, you know, you might think that, uh, might be a laggard from this respect. Not at all. We have many, many clients within the agricultural space, and it is already revolutionizing the way.

Farmers plant, um, how they manage the soil, how they identify, uh, pests and bugs, how they treat it, um, coupled with drones, um, and satellite imagery, it's, it's just a completely revolutionizing that sector alone. So I feel like [00:26:00] when, when AI can come. To the table and affect, um, a space as diverse as you know, agriculture and banking, it's probably gonna be applicable to your business no matter what that is.

Mehmet: Cool. Now, one thing I wonder about when we talk about these different industries and building the systems run, um. When it comes to the engineering part of it and you know, so, so if you can walk us through maybe the phases they have, so I know of course they're gonna start from a. Business case or a use case where they want to build an AI product for that.

And, but, but then, you know, we start the actual thing. So I'm curious about like, to understand, you know, also, like probably this is, will be initiatives coming from, from the CTO in the company. Uh, so what are the phases from, from identify? We talk about how to identify the business case, but I mean, starting doing maybe a proof of concept.

Mm-hmm. And [00:27:00] then taking it to a. Like production, great AI ready product or Oh, I'm excited to talk about this solutions. Yeah, this 

Ron : is great. So when we started the company back in 20 17, 1 of my big concerns was, is it going to be possible to even offer ai, um, engineering services to companies? Um, and the reason is really simple, as we talked about most.

AI systems today are still supervised learning base. That means that you need data, you need labeled examples, you need inputs and the correct answer to go along with them. Um, and that is, um. Not necessarily a given. So you're talking to somebody who wants to, um, uh, let's say build a predictive, um, a predictive AI model.

Um, I can't know beforehand if they have the data, uh, of sufficient volume and quality to build [00:28:00] a system. And the clients might even have, um, certain, um. Um, sort of accuracy numbers. Like they may say like, Hey, we want to build a system to extract information from documents, but it has to be 96% accurate on handwriting recognition.

And it's, again, it's difficult to know if they have, uh, the data. To accomplish this. So what we did was come up with a methodology that has worked fantastically over all these years, and I'm, I'm happy to share it. We break engagements down really into four stages. 

Mehmet: Mm-hmm. 

Ron : Once an initiative has been identified, um, and vetted that if it was successful there would be good ROI, we go and we do, uh.

Um, data exploration and analysis. We need to go and understand the data at a deep level. Is there enough data is from a quality and volume perspective, that can take as little as a week to, you know, several weeks to [00:29:00] do. Once you're done with that phase, and only when you're done with that phase can you go to the next phase, which is sort of model prototyping.

You take the data, but you don't know necessarily what all aspects of the data you need. You may have. You know, uh, you may have, if, if you're dealing with structured data, you may have thousands of columns, but, but not all of them really are bringing much information to the table. It may just be a handful that are really important.

So it's sort of feature selection, but you also have different modeling approaches. Um, unlike traditional software, it's not necessarily very clear exactly what the best, uh, architecture from a modeling perspective you should go with. So you prototype with the models. And you, you, you come out of this phase and you understand exactly what architecture, exactly what data features, and probably most importantly, you come out knowing what your what, what your likely overall [00:30:00] performance is gonna be.

So I will be able to go to a client and say, yes, I think we can hit. 95, 90 6% accuracy or no, we don't think we can. We, you know, we, we can only hit we think maybe 85, 80 6% accuracy and each of these checkpoints de-risks the engagement. The third phase is the actual training of the model. Um, you, you, you build, you train, you tune it.

There are, you know, it can be very complicated. There can be, um, parts of it that are not very scientific because, um, you know, we're still in the early stages of the, of this. There's hyper parameter tuning. Sometimes these models, you know, they're not deterministic. They're probabilistic. That's a good thing, but it also can make it challenging.

Um, and then once you're done and you have. And this is the bulk of the work by far. This is, you know, 80% of the work. Once you have a model that, um, you have thoroughly, thoroughly tested and you think [00:31:00] is generalized. With acceptable behavior in, in the domain you're in, then you can go to Productionization and that, that varies wildly based upon the specific um, case.

But those first three phases I mentioned, mentioned are, are really unchanged no matter the vertical or the end goal that you're trying to accomplish. 

Mehmet: So this is the baseline and then. After that, it'll be like, just, I like this approach. 'cause you know, you build really a, a good foundation first before taking it to right to production later.

And it looks like also you crack the cord run, so it works across multiple verticals, which is uh, which is good. Which is actually, yeah, like the only thing, maybe it'll defer, you know, the way the interface probably, or maybe the core of the LLM itself, to your point, like before about. You know, like, uh, with the, how, how you, you do the reinforcement, [00:32:00] uh, you know, reinforcement learning.

Exactly. Yeah, exactly. So, so this is, this is completely makes sense to me. Yeah. Like, this makes me excited also as well because, you know, the, the way you do it, like it's kind of. Uh, a customized AI application for customized use case within that vertical in that customer, which is, that's exactly right.

That's exactly right. Yeah. Yes. So, so, so, so that makes complete sense. And this is, you know, why I tell people I'm not, I'm not a consultant or an AI consultant, but people sometimes come say, Hey, like, you talk to a lot of people in the field, and they say like, yeah, you, you need like to dissect it in a way, and thank you, Ron, because now.

I can, I can use this approach and say, Hey, or you go check the episode because you dissected it in a very clear way where people can understand it in a layman terms. Let's, let's put it this way. So, so really fantastic. Now I want to shift gears a little bit, and, you know, you've been an, a, a, a serial entrepreneur yourself also on now when it comes to startups, [00:33:00] right?

So, um, do you think. Startups nowadays, and I know you are a mentor also as well, so do you think startups nowadays are leveraging AI really the way they should be doing, or they're just, you know, following the hype and Yeah, let's, let's just put an AI and say we, we are an AI company. What, what you're seeing from, from the work you do also in with the startups.

Ron : Oh, that is, that is, yeah. I think it's difficult maybe to broadly characterize, you know, all startups. So I would say, you know, like, like almost any tech trend when you're on that, that hype cycle, everybody all of a sudden is embracing it. Um, it feels like, you know, every new startup is now an AI startup.

Um, I, you mentioned I mentor, you know, um, I do a lot of mentoring of early stage companies. [00:34:00] You know, a lot of companies are saying they're an AI based startup when they're really not. You know, it's, it's really just more aspirational. Um. I like the idea Fundamentally though, um, I, I think that if you are starting a company now from scratch, if you're building a startup and you have the opportunity to rethink without all of the legacy concerns that entrenched players have to deal with, and you can rethink from scratch how to build your product or services to put AI at the core.

It is absolutely a no-brainer that you should do that because it will change your entire go-to-market strategy. Just my, my one word of caution would be don't let you know your aspirations. Um, get ahead of what you are capable of doing. Be ambitious, but also be realistic, um, [00:35:00] and don't. Um, I see too many early stage companies, um, thinking that if they just fake it till they make it.

That eventually a miracle will happen and the AI will become a reality, and it's, and they're kind of hollow at the actual core of the technology. That's the one mistake that I would deeply encourage, uh, uh, early stage companies to avoid, make AI the core of the way you think about it, but grounded in reality.

And I think you can have the best of both worlds. 

Mehmet: Cool. Now one, one more thing, like, is it now. In your opinion, especially I'm talking about not every company, but especially the companies who have to deal with a lot of data. They have to actually, everyone have to deal with a lot of data nowadays. But I mean.

When they're building their foundation, uh, is that they must have to [00:36:00] have a machine learning or an AI engineer on the team. So this is the first, I know it's kind of a loaded question, and we started to see that this domain. Mainly AI engineers. We have kind of a feeling we have scarcity because all what's happening, right?

We, we saw like meta trying to Yeah. Get talents from competitions. Uh, and we're hearing, you know, it's like the wild west, right? So as a startup, like if I'm a founder today, uh, how should I prioritize hiring my AI engineers? Like is it like necessity at the beginning? And what kind of maybe knowledge I should be looking for.

Ron : Mm-hmm. I do think it is a necessity, um, because you know, if you follow the idea that. You're gonna build a startup, you know, in this day and age, and you're going to be sort of [00:37:00] AI first, you're, you're going to think about, um, building functionality and not layering on AI or adding in ai later, you're going to take an AI first perspective, which I really think, uh, new companies should, um.

You really need people that have the ability to build those systems and, um, handle all the complexity of all the things we've talked about already. Um, unlike, I think unlike most of the tech transitions, the, the big moments that we've seen over the last 30 years. So if you think about. The, the, the, uh, worldwide web exploding in 1995, or the move to, uh, mobile, um, after the iPhone development in 2007.

You know, AI is the third really giant technology movement during that period, but it's different [00:38:00] because there is so much math and there is so much, um, uh, uh, intuition that you have to come. Into engagements with, around how to manage and successfully train and build these very, very complex systems, these very complex probabilistic systems.

The learning curve is much, much steeper than most software engineers are used to. So it's not like, Hey, the the iPhone just came out. I'm just, I'm gonna go learn how to build iPhone apps, or the, or the web just came out. I'm gonna go learn H tml and CSS. It is. Deeply, deeply, uh, uh, um, complicated, uh, endeavors to build state-of-the-art enterprise grade AI application.

So my advice is. Absolutely hire machine learning engineers that have production experience. There is, you know, we always talk about the difference between like a [00:39:00] proof of concept and a production application that has never been a bigger Gulf than between a proof of concept AI model that works like.

It's 60% accurate and one that can get to, you know, 98% accurate for production. That can be years of work, even if getting to 60% took you five days. And so people underestimate the complexity of building production grade applications. 

Mehmet: Cool. Just one thing because you talked about software engineers and, and all that, um, people are talking now that, you know, the infrastructure of AI is commoditized, like in a sense like the LLMs, the models, it's just like matter how you bring people that can train them.

So is it like, do you think like really AI is commoditized now or like No, especially 'cause I had a conversation the other day and, and I think. When, when the time comes to air, both episodes, or the episodes before this one. So we're talking about that software [00:40:00] isn't like the, the mode anymore, it's like more on something else as well.

Mm-hmm. So we talk about deep tech, we talk about, you know, merging the AI with, I don't know, maybe quantum computing with blockchain. Mm-hmm. Or something like this, um, from you. Me and Julie Ron, like we've been here for long times and you know, because always I give the same examples about the internet, the mobile revolution, and you know, the social revolution also as well.

So if, if we want like to, to put kind of predictions maybe, of course we, we don't have crystal balls, but. For, for differentiation beyond software and beyond ai, what do you think, you know, the next generation of startups would be disrupting in? Mm-hmm. Which technologies? 

Ron : Mm-hmm. Um, it's, this is a really difficult question because we're at the stage where, you know, I think maybe three years ago, most people would've thought that the, [00:41:00] the sort of frontier labs that we're building.

Um, the, the most sophisticated large language models that, that was gonna be the differentiation. It was who could build the smartest language model, uh, the most capable model. And you could look at companies like, um, Microsoft or um, AWS or Apple who weren't. Building those models. The themselves, they might've been investing in companies, but they weren't building it.

And it looked like it could be a liability. It could actually potentially, you know, undermine their entire, um, value stack. Now in 2025, it seems pretty clear at, at least at this stage. That there's parody, uh, on the frontier, uh, model level. And, you know, if you're working with AWS you can go through Bedrock and pick which one you wanna use.

And, you know, um, it's not really so much of a differentiator anymore. And so I, my my belief is that that's [00:42:00] probably gonna continue. I think that there is too much of a culture of sort of scientific sharing within the AI community. Um. Uh, amazingly, uh, you know, a lot of the really interesting breakthroughs, uh, have, have happened in China this year and the research there, researchers there have published the specifics of how they did that.

Unlike a lot of, you know, American companies, and I think that's really, really valuable. And so my guess is that. It's going to be many years before things start settling down. Um, and what, what companies should do is not worry about any one AI capability or any one model. Or some magic point where we're gonna hit a JI and now this one model can do everything.

I think we're quite a few years away from that. And what I, [00:43:00] what I tell companies is when I'm working in consulting with them, I tell them, you have proprietary data and that is your. Competitive advantage. Don't worry about building a front, you know, a billion dollar frontier model. Don't worry about the infrastructure and the power issues.

Take the proprietary data you have and build capabilities that give you either, um, competitive advantage from new capabilities. Um, uh, streamline enterprise workflows, new predictive capabilities, new generative capabilities, whatever it may be. But that proprietary data, um, is, is is gold in, in the landscape of against your competitors.

And we see this over and over and over again at Kungfu AI is you can go take this data that you've had for 20 years and build a system that can. Transform your business. I, I'll give you one example really quickly. Sure. We built a, it took a couple years, but we built a model [00:44:00] and it can do, uh, uh, a loan decisioning automatically, and it transformed this business from being, um, you know, 48 hour decision times to nine seconds.

Wow. And. Fraud is down, chargebacks are down. And when they went public with this new model, and they, they're a publicly traded company, when they announced the release of it, they saw a quarter billion dollar increase in market capitalization because, um, they were taking this proprietary data and using it for competitive advantages.

Mehmet: That's fantastic actually, just, you know, uh, if I might suggest something, I didn't finish it yet. I'm, I'm halfway. Uh, because you mentioned about, uh, before we talked about data and you mentioned now about the China. So there's a book, which I'm reading now called AI Superpowers by Kai Foodie. Uh, highly, highly recommend it.

I didn't finish it yet. I'm halfway. But this, this gives. You know, everyone [00:45:00] established companies and startups as well. Exactly. Your point, Ron, where? It's discussed like why China was able to have like little bit advantage when it comes to ai. It all comes down. Of course, the author described many things, but data is the main thing and people would be surprised, like, because that's right.

You know, he describes how I, I don't want to tell all the whole book, but you know. It's like we used to think of Chinese companies as they just copy what they see abroad, but it ended up with time because the mass population, the mass data they have. So they were able to have more data that, you know, US companies, European companies have, and they are like mainly in the fields that touch human beings every day.

So ride sharing, food delivery, just to name few. Uh, transactions, you know, wallets [00:46:00] and guess what? Like this, give them a huge edge. So this is why when I start to think about it from a startup or like even a established business perspective, to your point, Ron. Yeah. Yeah. Your data should be your sacred place and you should try to also think, in my opinion, correct me if I'm wrong, how we can get more data that.

Can be useful for us. So maybe there are some data points we're not collecting today, we are not monitoring how we can bring this and maybe find a co, co uh, relation with that. So a hundred percent on this one. Um, Ron, as we are almost come close to an end, any final thoughts, anything that you know, maybe I should have asked you I didn't, you wanna share and where people can get in touch?

Ron : If pe people can find me, uh, uh, uh, online, we, um, kung fu.ai is our website. Obviously, reach out. Uh, my email is very, very [00:47:00] simple. Ron, RON, at kung fu.ai. Um, if you're interested in. Uh, uh, AI and, and we have a podcast called Hidden Layers where every month we cover the, the latest breakthroughs and we'll do a little bit of, a little bit of a, a deep dive on the technology, um, but we try to make it broadly applicable.

Um, check that out if you're interesting, uh, interested in that. And, um, uh, and then I'm on all the, you know, the normal social media stuff. 

Mehmet: Great. And thank you again for being here with me today, Ron. Really, I enjoyed the discussion. It's very enlightening, I think, for both business people and technologists, uh, as well.

And you know, this is why the call the show is called CTO Show, right? So because we try to bridge the business and tech together. So thank you very much Ron for being here with me today. For the audience, the, you know, everything will be in the show notes so you don't have to. Look for, and this is usually how I end my episodes.

Uh, this is for the audience. If you just discovered our podcast by luck, thank you for passing by. I hope you enjoyed [00:48:00] it. If you did, so, give me a favor, subscribe and share it with your friends and colleagues, and if you are one of the people who keeps coming again and again, thank you for all the support, for the feedback, for the, thank you very much for taking the podcast this year.

I'm repeating it every time. I know, but this is just a sign of. You know, being grateful for my audience, who with their support, the podcast is trending in the top 200 chart on the Apple Podcast across multiple countries at the same time. Eight to be specific, hoping to reach the nine and 10 marks very soon.

And this cannot happen without all the support of you, my audience, and of course all my guests, including your own also as well. So thank you very much, and as I say, always. Hope you enjoyed this episode today. We will have a new episode very soon. Thank you. Bye-bye.