Mackenzie Jackson: Hello everyone. Welcome to another episode of Cyber and Sake. I'm here with Matthias Feys. Welcome to the show.
Matthias Feys: Thank you. Nice to be here.
Mackenzie Jackson: So, probably no surprise for those that know you and know what you're about, that we're going to be talking about AI and all things. You've been in this space for a whole lot longer than most people. So, before we get into it, I almost forgot, but we do have some sake here to try. So, we're going with our first 0% sake. Now, sake famously from Japan. This sake is from the very famous brewing state of Arkansas in the United States.
Matthias Feys: You popped it.
Mackenzie Jackson: But it's actually my preferred 0% so we'll try a little bit of this. All right. Cheers. Cheers.
Matthias Feys: I like it. I like it.
Mackenzie Jackson: It's quite interesting. Yeah. Okay. So, now that we're ready to roll, I want to start with the beginning of ML6. You guys founded in 2014, and you're not a company that pivoted into AI when it became cool. You've always been an AI company, an AI startup. Maybe you want to talk to the vision that was the why you started that company, and really what you do today.
Matthias Feys: So, it goes way back. The company I joined first was called Datatonic, and they did data analytics. Basically, there we did first AI projects. Based on the first few experiences, that company got split up and we co-founded ML6, me and Nicolas, with a pure dedicated focus on AI projects. We did the first few projects and also with my background we really saw a lot more things to do and we just wanted to focus purely on that. From that point on, we kept that focus. In the beginning it was quite niche, now it's very broad, so that changed, but otherwise most of the things remain.
Mackenzie Jackson: Yeah. So I kind of want to go back to the beginning of that, because machine learning, and I'm assuming ML stands for machine learning?
Matthias Feys: Yes.
Mackenzie Jackson: Yes, which was kind of what we, back then, it's really what we referred to as AI. We were talking more about machine learning. Is there still kind of a distinction between machine learning and AI as we think about it now, and what that is?
Matthias Feys: Yes. AI is a broader term which comprises a lot of different types of tools to use. Machine learning is one. Then Gen AI is the most popular format right now. Ultimately, it's also based on neural networks, but it's quite different in how you use it. So with machine learning, you get from the data you have, a few models you train yourself. With Gen AI, which is typically what everyone talks about when they talk about AI, these models are pre-trained and you don't need to do anything, you can just use them. So, in the early days, it was mostly about how to train these models, what data we need to collect to train those models, what architectures are best. For most applications, the answer is transformer-based models, and you don't need to train it yourself. You just prompt it. And that's this big distinction between typical machine learning and deep learning projects, and Gen AI projects, in which the model is solved for you. For some use cases, it's still very useful to train or fine-tune the models, but typically that's not needed anymore.
Mackenzie Jackson: Right. I want to get back to kind of some of the different models and training in a minute, but I kind of, before we get there, I want to focus on just the journey that you've been through, because I find it fascinating. To give you some context, the first viral piece that I ever wrote is something that I wish I could remove from the internet, because I wrote this opinion piece for a tech magazine after I went to a conference and I saw everyone had AI on everything, and this was pre-ChatGPT. The article I wrote was titled Artificial Intelligence is More A Than I. My whole piece was, it's just a marketing ploy and there's nothing really real here. Six months later, ChatGPT was released. My opinion has changed significantly since then. When you started, did you anticipate where we would get to with AI and the journey? Did it take longer? Was it faster? I'm curious to know, in 2014, did you expect to end up here? Did you think it would look different?
Matthias Feys: The understanding that it would become much bigger, I think that belief was always there. So that's not like, oh, what happened? How it happened in the classical pattern of it will become easier for people to train their model, and that will become easier. That was the thinking, like it will become so easy everyone will do it. Well, ultimately what happened is like no one trains their models because it's done for you. So that's quite different. The typical applications or things you can do, I think there are changes on that, but I think it's mainly on people don't need to use it or to understand it in order to use it. And that's a big unlock, which is quite different. But aside from everyone using it in some formats, I think that was the common belief which we rooted for as well.
Mackenzie Jackson: Yeah, yeah. And it's so easy to think back in hindsight to be like, oh yeah, of course that would have happened. But literally, I was not a believer until I actually got down and saw. I knew that there was going to be implications of it. I knew that it was going to happen, but I didn't think that it was going to revolutionize everything, because it had been promised for so long and all that seemed to be happening was marketing. Again, I was wrong. You were right. Do you agree with where we ended up? So, what ended up happening was OpenAI released their model, and it had a follow-on effect, because Google also had very similar generative AI models that they were working on. Same with Meta. OpenAI released theirs, some would argue prematurely, you know, in terms of some safety features and areas like that, but it pushed everyone to then kind of move forward. Do you agree with the process?
Matthias Feys: I'm not sure if I can agree or not agree, but I think they made it useful or more public, and specifically with applications on top of that with ChatGPT, because the model itself is the transformers model architecture, which was invented by Google, but they didn't do that much as in building new applications like ChatGPT and exposing it. So I think it's more about OpenAI really opening it up in this way, not an open model, but then an application that people can use which made a lot of people believers like, okay, this can do a lot of new things. And I think it's still a net positive, so I don't think I would say this is totally wrong because I think it's still a net positive to go forward.
Mackenzie Jackson: Yeah, and mistakes have been made and prices have been paid, but I think it's still a net positive where we ended up. I totally agree. Yeah, well let's talk about where we are now then. This is a good segue. One thing I'm kind of interested in, and I really wanted to talk to you and we've kind of touched on it, is this idea of when we talk about AI now, we're almost always talking about, at least in my circles of AI, the GPT type. There are a whole bunch of different types of applications for artificial intelligence that isn't Gen AI. Maybe can you explain what they are and when they're useful?
Matthias Feys: Yes, for general public, as for most applications. Yeah, it also covers most typical applications what you can do with Gen AI. So I think that's really a big chunk in there as well. Ultimately, these are still neural networks. So they predict the next token words, and that concept is a broader thing. Then you have a bit of a sister to that, is like the typical deep learning project, in which you have not the super large models, but like the smaller modules you train yourself, which is still useful in specific applications which are more local. So if you look at, for instance, computer vision in manufactories, you don't go with a large language model or vision model, you have a fine-tuned model there. You have your typical time series forecasting models, which are also still prevalent. And there's research on using LLMs for that, but you can just train it yourself with the data points. So that's a separate type as well. And there's also new research on new types of architectures, more like symbolic, because ultimately AI is the concept of being intelligent artificially, and there's lots of different ways of getting there. And in the very early days were more about expert models or like the symbolic reasoning things, which I think will come back, but at this point it's still quite small.
Mackenzie Jackson: Right. One of the things that you do at ML6, you have lots of different partners that you've been working with. You've got lots of, and this is interesting for me, because when we're dealing with AI in my area, typically dealing, we're in security, it's typically a very specific thing. You have much broader knowledge of it. Where are you most excited about AI being implemented? And maybe this is kind of logical or what we would expect. And are there areas where you look at it and it's kind of the wrong tool for it or something that's just kind of, you know, because just to preface this, one of the things that I feel is that we kind of use AI as a bit of a sledgehammer to hammer in the nail sometimes. You know, and the computational cost of doing something that could be done in traditional methods, sometimes feels very redundant.
Matthias Feys: I think I'll maybe start with misuse, because I think you need to understand the application or the use rather than the technology for that. I think you're totally right, like it's the default answer to a lot of different things. And it's an overreliance, it's like just AI as a default answer to a lot of different things and a lot of use cases in which people just assume or they just trust it. And that's really a big no. You don't trust it. You can use it, you use it but you don't trust it. So it's like that's a big one. A lot of mistakes being made there and like it doesn't work. No, you just use it. You just don't trust it directly. So I think that's something that's generic to a lot of different applications. And then what's interesting is, if you use AI to help generate code, that's where you have the benefit and downsides mixing together which could end up with a positive outcome. Meaning like, you could use LLMs for doing a lot of different things, but if you have LLM generate the code to do more useful things, you typically use it in a better way, right? So instead of saying, hey, just do this calculation, you ask it to generate the code to do some calculations. That's typically like a nicer way and there's a lot of nicer things to do. And I think there there's a lot of more gains to leverage than we're currently doing because it can do a lot more with that paradigm. And if you purely use it as an LLM doing the thing in itself, you kind of also misuse it for the wrong things.
Mackenzie Jackson: Yeah, it makes a whole bunch of sense because I mean it's, I guess that's when you have a tool that's so powerful, we really need to figure out how to best use it. And I like that analogy of, and I think what you said can really be applied to security in our things. It's like we know we, we scan code for vulnerabilities. If you use an LLM to do that, well, one, it's non-deterministic, so it may give you the wrong answer. But it also, the computational cost of that is huge. Currently, OpenAI is covering most of that bill. Thanks. But, but it could create rule sets that a static engine could use that would be much more appropriate.
Matthias Feys: Yeah, I think there's much more there to exploit and explore because indeed, yeah, like the summing is the simplest example which is very generic, right? Like asking it to make a sum of numbers, it's really bad at it, but let it create code to do the sum. It's very easy. And there's a lot of those things and specific applications you can use for the better. Yeah.
Mackenzie Jackson: And what about on the other side of it, like where is something that you're really excited that AI is totally changing the game in your opinion? Maybe it's a project that you've worked on that's really changed things.
Matthias Feys: So, I think it's more on the boring work. So there's always nice to talk about very cool things to do. And I think it's more about the boring things like under the hood, which makes it a lot easier. So, back in the day we had a lot of NLP, natural language processing projects, and we're a lot about literally building a custom model for a lot of different use cases and that's completely obliterated. It's really done. But it doesn't mean like these types of use cases are not useful anymore. It's kind of the scale at which you can deploy it is quite different. So, for a lot of different, I think like the one like I always go back to, is during COVID, we had a prediction model to help forecast demand for certain jobs. And at that point in time, there's a lot of time series forecasting which required previous historic data. We just finished that model, training it, when COVID hit, but then the situation hit dramatically, that the historic data wasn't that useful at all. So it's kind of like, yeah, this model will predict what's happening, but it can't predict if it was a quite new situation. So that was interesting. And then we pivoted to basically look at data, have specific NLP models to understand articles or things to extract where there's probably more jobs to be created or lost. But those were specific models, specific for every language, specific for lots of variation for that. And if we would be able to do that right now, it's kind of just one prompt away and you have the same thing. So I think the lowering of the bar and where you can apply it quickly, it's quite different.
Mackenzie Jackson: Yeah. Yeah. I think that makes a whole lot of sense. I love what you said about the boring work, too, because when AI first came out, it seemed to think that the only thing it couldn't do was boring work in terms of like it would create great pitches. It could write stories until everything's fine. But I think at its core, we will get down to a point where we can actually automate a lot of this stuff. It's boring, repetitive jobs, because back in the day it was like the 3Ds: dull, dangerous, and something else, you know. But then GBT is more like now it creates music. I don't know.
Matthias Feys: Yeah. The multimodality is quite interesting.
Mackenzie Jackson: Yeah. Absolutely. But I think we will get back there because that's useful to automate for us. This is mostly a security podcast, but we have all kinds of interesting people on here. I do want to maybe pivot a little bit into just some areas of security that you can think of. When you, you know, you're a technical person yourself, do you ever worry about the security implications of these AIs? For example, you know, we've released here this Agentic AI penetration testing. So AI hackers.
Matthias Feys: Yeah.
Mackenzie Jackson: I love our product. Freaks me out a little bit. I'm not going to lie. Do you worry about that as someone that's in here, or what?
Matthias Feys: In a natural way, I think as mentioned, I don't think you should trust things like the zero-trust thing. I think it applies to AI systems as well. I think it's best to design it so that you don't trust it. You don't have to trust it because you can govern it in a proper way. So I think in this way it's contained. And if the application is built well, I think the application if it's built well, you can trust it. But just saying trust AI, that really doesn't work. And if you talk about these agentic systems doing things, it's again about how do you guard it so that within it, the things that are possible to do, it's fine, whatever it does. But if you give it access to everything to do whatever, I wouldn't trust that either.
Mackenzie Jackson: No. Yeah. No, I don't think that's the correct approach. But it's so confident when you start using it, too. And there's this kind of dilemma that I talk to people about and is kind of, when you know a subject really, really well and you ask an AI model, then you realize that it's giving you wrong information a lot of the time, but in a very confident way. But are you asking an AI model to produce something that you know already or something that you don't know? So therefore, it's much harder to understand the areas that you get wrong. And there are security implications that I'm worried about. I'm also worried about a little bit of lack of knowledge in terms of, if I was a developer writing code now, my approach and my learning path would be very different than the one that I took.
Matthias Feys: Yes, yeah.
Mackenzie Jackson: I wonder, and I'm the type of person that if I can avoid learning how JavaScript works but still make a JavaScript application, I'm probably going to do that.
Matthias Feys: Yeah, probably most people.
Mackenzie Jackson: Do you worry a little bit about a lack of knowledge?
Matthias Feys: Yeah, exactly. I think that's a big one. I think that's a really big one. So I think, but also education, right? And as a whole it has always been the case that you have tools, but you still learn how to do one even though you have tools to help you with just calculation. I think the same holds true for a lot of more abstract things and with helping with a lot of work to be done. I think it's essential for people still to understand how to do things properly than to use systems to help them with that. But just relying and trusting that another system that they don't know how it works, just trusting that, I don't think that's a good idea. Because as long as it just helps you accelerate on what you want to do, I think it's totally fine. But if you rely on it in a way that you just spill a net spin, then I don't think that's a good idea and we need to really avoid that situation.
Mackenzie Jackson: Yeah, absolutely. It kind of reminds me of a funny story. I have a friend that I went to university with that decided to become very proficient in an old programming language called Cobol. And all the bank mainframes are on Cobol and there's this real big dilemma is that no one wants to rebuild that infrastructure, but most Cobol developers are 60 years old right now. I wonder if we're having a similar conversation in 20, 30 years when all the developers that built everything are now looking to retire. So look, it will be interesting, but hopefully we will get there. Yeah.
Matthias Feys: And it's also not necessarily a technical problem, I think, because if you look at press, but also look at systems of just engineering systems, for engineering systems you also have the bus, or for any company you have the bus factor. You don't want to be relying on one person to do everything. And it's a bit of the same if it's relying on one tool for everything. It's also a bit of a risk. And then for the news, if you read the news and you're in the article and you're like, yeah, but they did this and this would be different, a bit differently. And if you're not in the context, you feel like this is the truth. And it's the same for the AI systems, or if you use the tools, if you're really in it, you're like, yeah, but you missed this and that. But most people are not in the specific or you're not the expert in most things, you're expert in specific things. So if you ask a random person something and they're more expert in that topic than you, they might still be wrong but coarsely correct. The problem is with these large chatbots, they are proficient in all the things, but they're not an expert in all the things. And that's an interesting distinction.
Mackenzie Jackson: Well, let's talk about the future then, because they're not an expert in all the things now. Where do you think we're going with this? Will we get to a point, you know, I, I don't want to talk about super intelligence or general intelligence just yet, but at least in the very short term, before we kind of get to that mythical area, you know, are we, are things improving at a rate where you think that for example, it will be able to write perfect code with nothing wrong with it? It will be able to research topics without hallucinating any answers. Do you think we'll get there or is the technology kind of going to prohibit that kind of perfection? Let's say.
Matthias Feys: I think two answers for me. I try to not have to answer that question in a sense that I think the things that you can do that are useful in both situations are certainly the things I want to do. So if models get better and there's things that it will do automatically, there's still a lot of integration with people. A lot of different things still to solve and I think that's certainly no regrets as in, is it just a matter of scaling? I don't believe that's the case. So I think there's a lot of steps still to take to get to some level, but I think it's also very asymptotic. So it's not like it will be there with one more thing.
Mackenzie Jackson: Yeah. So I think one more version of G.
Matthias Feys: Yeah, one more version. So I think it will be rather asymptotically to get better and better, which is, so I don't feel like everything will be solved. And there's still ultimately using tools is where the magic happens and it will still be people using the tools. So there's a lot of interaction still to be solved and it's not like everything will be resolved by one chat interface. Yeah. Because you still need a lot of different things to govern that, a lot of processes to offer the securities on top of that. So a lot of different things still to do. So on one hand, I think it's not for tomorrow, and regardless of that, I think there's still things to add on top of it.
Mackenzie Jackson: And let's talk about kind of the super intelligence component. Now, this is the one that kind of really, I guess, really freaks me out is the point. And there's a commentator that I listen to a lot that every time AI comes up, he goes, yes, but has it solved cancer yet? Has it killed cancer yet? Because until it does that, I don't really, I don't really get. But my thoughts is like, yeah, but when it solves that, it's kind of like, I don't know what I'm going to be doing. Probably not podcasting. Let's start off with the obvious one. Are we going to get there? Are we going to get to super intelligence, general intelligence? Will it happen in the next 10 years, our lifetime?
Matthias Feys: I think definitions always change. So, I'll just start there because I think as in AGI, on where are we? If you look at the test of just the Chat GV answering questions, I think it's very difficult to say if it's a person or if it's a model. I think that's already a lot there, but it's not like a magical thing happened and everybody's like, oh, my job is done. Yeah. So I think it's, it will keep on changing what we expect of the things to do in order to be experienced as such. So I think that's part of the test that will just keep on moving it. Personally, as with ML6, really personally, I think that the enterprise super intelligence will happen first in a sense that the thing of being on par with a person and being better than a person, I don't think as in general like for everything that makes sense. If you look at it from a company perspective and having processes in place and trying to improve processes and different tasks to be done, I think that's something that's much more confined as in what does that mean? It's like you have a goal to achieve and as a person or organization have different people, different contexts. And I think these systems are capable of having a lot more context than one person to be used. And I think that's something that we're really close to. There's a lot of things going wrong and it's like if only a person knew this and then you talk with another person, then that person also knows. And there's really constrained my bandwidth, like if people all know the same thing, they would do a lot of different things, but that's impossible. And I think that bandwidth problem is something that these systems are able to get much closer to, but it doesn't mean they still need any guidance from people involved. But I do really feel like there's huge unlocks possible there, because the context problem of like people knowing what they need to know, it's really, really constrained and really causing a lot of friction which can be resolved soon.
Mackenzie Jackson: One of the interesting things that I learned about you doing my digging into you preparing for this, was that you were actually kind of on in the process of doing a PhD. And I guess how it is, how I kind of found out about it is that you were kind of so convinced in the real world applications that you left that early to pursue it. What was that PhD that you were kind of looking into?
Matthias Feys: Yeah. So the PhD was on summarization of events or of texts, and using deep learning. By that point in time, deep learning wasn't that hyped up. Yeah. It wasn't Gen AI hyped as it is like Gen AI now.
Mackenzie Jackson: Yeah.
Matthias Feys: But deep learning was like the thing, like a few years later. So I think there's two parts. One, I really wanted to be it used at companies because I really felt like this can do crazy stuff. And I just wanted to not wait four years in order to apply it somewhere. So that's one. But also like, I really, really enjoyed my time as doing research, but also felt like there was the time horizon is quite different. And I'm more made of like doing for the short-term time horizon. I'm working on things which get to production. So that's the other side. Like I really also felt like I'm not maybe a researcher, really more of like making it real in real life applications. And I was, when I was presenting my work as well, I felt like people are like, yeah, so what's like in industry, like if I was going to industry and like this is what I'm doing, I was like.
Mackenzie Jackson: Good for you. I was like, how is this possible? I feel like that kind of happens a lot with research is there's like an abstraction there, you know, and then there's kind of like some people kind of have this chip on their shoulders like, yeah, but in the real world outside of your lab, you know, that doesn't, that doesn't quite work. I can really understand the need to get out there and prove it in a way. The other thing that I kind of want to dive into a little bit is because you've been on the forefront of this journey for so long, you were dealing in AI in 2014, which is like a lifetime ago now. And I went back, I almost, so many companies have pivoted to AI now and you would never believe that they were something else. So I, when I was doing this, I kind of assumed that ML6 was the same. I went on the Way Back Machine. I wanted to look at the first versions of your, and you didn't pivot. You've been on AI for a long time. So back then, what were some of the things that you predicted would happen with AI that you guys were right about and what were some of the areas where you know, you really got it wrong or some of the some of the assumptions that you made that didn't come true?
Matthias Feys: Yeah, I'll maybe start where we were wrong and it's always about trying to bet with progress and not against it, which turned out right. At one time, I think we did it wrong in a sense that the application of LLMs back in 2022 or something, just pre-Chat GPT, when we were doing projects in that context, we always had to fine-tune those models, just to make it useful. We made an assumption and we also made investments in open source that work, to make it easier for people to fine-tune their own models because we thought like, hey, this is great and with a bit of fine-tuning you can just use this, will make it easier for people to fine-tune their own models to help with collecting the data and making all that work so much easier. People don't know yet they need to do that because they think everything will work without doing any fine-tuning. We were proven wrong. So I think it will come back. So it's about making the same useful with less. So with less big models, yeah, you still need to have some fine-tuning. But up until now, the, it's not like the applications need you to fine-tune. I think we did a lot of work in that regard and also did a few projects to help people with that. And I think we're too early or too late. I'm not sure. Still what? So we're at the wrong point in time. So I think that was a big mistake in the sense that we thought that this would need to be solved first versus something that comes way later. I think for cost-wise it's still useful to do that, but for just getting it to work, totally not. I think the opposite of the things like betting with, I think most of the time we bet it with like, hey, it's not good now but it will become much greater soon and that will always turn out pretty useful to be there. So a lot of the projects we were very early to doing have a first proof point of that and then we repeated that. So I think in 2014, 2015, TensorFlow came out from Google, the open-source framework to help with deep learning to train our models and do the inference. And we're really hardcore on that. So really kind of co-branded and I think our first versions of our gear were also about TensorFlow and ML6 and really thought it really closely together and that also made for a lot of first projects that really if they wanted to use that framework, that tool, we're also very close to that. Yeah.
Mackenzie Jackson: Yeah. I think that's, yeah. And it's a pretty safe bet to bet on Google in that case because over there was like some nonprofit, I don't think anyone predicted that that was going to become. And who knows where it will go in the future. You work on a lot of different projects. I feel like there's a lot of companies that want to use AI, maybe don't know how to do it. Don't know the best approach. Don't know what models, AI, other areas of AI, and what all this stuff that we talked about. What do you, what do you hope to do at ML6? How do you kind of work on projects and help people kind of build into, build, build AI into there in a meaningful way?
Matthias Feys: There's a lot of reasons where to use it best and it's really depending on your company on like where you want to use it. If you want to use tools for that, if you want to do something custom-built, or if you don't want to use it at all and it's like it needs to be a thing so that you can say no to things as well in order to say yes to the good things. So I think it's a good starting point to help with that. So typically we also advise people or sorry, companies with like just mapping it and also seeing where you don't want to use it. Because it's good like to not put any more focus on that. So you can put your focus on where you do want to use it. And then there as well, like we help either how to use tools that exist or how to build your work tools. And again there it's like a big difference between if it's really differentiating for you as a company or if you just want to use it because everyone is using it. It's like the default and then you don't want to differentiate and just buy the tools. So there's like different things that we do to help them to to get to that point. And then if they really want to differentiate and do something custom, that's where we help them build it together as well.
Mackenzie Jackson: Yeah. Yeah. And I think that there's a gap that's, there's so many mistakes that I think people make out of excitement, AI, you know. Like for example, at Aikido, one of the areas that we're using AI to analyze a whole bunch of vulnerabilities that we find in the open space. And our first approach, our first thought was, let's build a model that's going to be really good at doing this. So, you know, you burn huge amounts of cash to do that, only for the next minor upgrade of GPT to literally smash your model. And I think you just need sometimes some advice of to take a step back and actually know, this is what it's going to be good for. Here's where you can use a frontier model. You know, here are some different uses.
Matthias Feys: Yeah. And it's also what I meant, like if you bet with technology and assume things will get better, what can you do that still makes sense? And then you're happy with progress. For sure. Every upgrade of the frontier model is an upgrade for your product. Yes, exactly. And if you do it the way around, you're like, ah, they will just suck and I'll have to solve it for them.
Mackenzie Jackson: Called Jensen, I need some more GPUs. Yeah, it's difficult to win. All right. So, before we kind of finish up here, we have a fun game that I want to play. It's called Would You Rather. I'm going to give you two very crappy options, and you have to pick one. But before we get into that, we'll try our second RC that we have here. I think this is the right. Looks like the right color. So, cheers. Cheers.
Matthias Feys: Yeah.
Mackenzie Jackson: It's like a soury juice. Not really sake. I'll cut that out. But it's like my favorite Japanese drink that I found. I was looking for alternatives of like sometimes I do this at 9:00 AM and you look for reasons to. All right. So to the game, would you rather, would you rather have an AI system plan all of your agendas, everything that you can do, or you cannot use AI at all in your professional life? So either have an AI system that plans absolutely everything for you, answers your emails, replies.
Matthias Feys: The first one. Yeah. Yes.
Mackenzie Jackson: Okay. Well, let's, let's move it up a level then. Let's talk about writing code and everything. If you were to write, build an application, you either completely vibe code the application. So it is entirely on AI. Now you can use Cursor or Windswept or one of these tools, but you cannot write any code yourself. It has to all be done through the prompt, through the help, or it's 2014 again and you're building it entirely by hand, which would you rather?
Matthias Feys: You're making it much tougher now. This is terrible. It depends on the application usage. I would say if it would be really enterprise-y, I would do it myself without any tools. Yeah. For sure. Yeah.
Mackenzie Jackson: Yeah. I think that there is a bit of a, there is a difference between a banking app that's modernized. And I'll tell you a funny story that we actually, in the marketing team, we opened up a position called Vibe Coder. And basically we wanted someone to, they didn't need to be a developer, but they needed to know these tools. And we just basically wanted them to be able to help us build little projects, little things unconnected to the main thing, just little marketing projects. We actually had to get rid of that position because there's a lot of, a lot of potential customers didn't like the fact that we had a Vibe Coder. So even though it was completely like detached, I think the people's perception of that was kind of like, yeah, so security product, don't want it to be vibe coded. Checks out, checks out. All right, let's move on to kind of the last area. Would you rather, you've gotten your application, you've built it now entirely. And I'm going to put forward a timeline, two years from now. Would you rather face off against the latest APT, so a state-sponsored hacking group that's trying to attack you, disrupt you, gain access into your databases or the latest AI models with an unlimited budget in terms of like tokens and credits trying to attack your site. Which would you rather?
Matthias Feys: In two years, right? In two years. Well, let's go now and in two years. I think that would pivot the situation. Yeah. I think for now, the question is what would you rather? So the question is what do you think will perform more, right? Just answer. Yeah. I think at this point, in two years, I do feel confident that the AI system should all work better for that because it has access to a lot more, I think within two years. At this point, I'm not the expert in that regard. So I don't feel like, I think it's a lot of early days for that application and I think it's at this point I'm more confident that the tool will limit its access to what it has seen and that's limited. But in two years, I have a totally different view on that.
Mackenzie Jackson: It is, it is a pretty wild area that we're now, but it's certainly going to change and it's going to be very interesting. But I've really enjoyed your perspective. Thanks so much for coming here. Thanks for answering the questions at the end. And yeah, I really look forward to watching your progress continue.
Matthias Feys: Likewise.
