TRANSKRYPCJA 🔗
Ilya Sakharov: So we're going to have the spec at the beginning. We're going to have the input. We're going to have the output, but that magic in the middle, okay? makes us think, assume. So we make mistakes and redo is disappearing.
Sebastian: Welcome to another episode of our new season of Beyond Vibecoding partnering with ImpalaSearch.
André: The go to tech and executive search agency in Germany.
Sebastian: In this podcast, we explore the transformational change in software engineering and knowledge work in general. I'm Sebastian Heidemeyer zu Erpen, CTO at Ecosia.
André: And I'm Andre, CTPO at Trusted Shops. Great to have you back. In this episode, which is again a somewhat different, we take a look at how AI will influence our way of thinking and how we remain responsible for the work we do with AI.
Sebastian: Yes, we're speaking with Ilya Sakharov, Executive Director of Engineering of Redcare Pharmacy, and we are exchanging different perspectives. Nevertheless, we end with a joint verdict and we hope you enjoy the discussion as much as we did. welcome to our podcast, Ilja. Excited to have you on and as usual, like to ask you to introduce yourself to our listeners.
Ilya Sakharov: Hi, Sebastian. Hi, Andre. And thank you very much for having me. I'll be honest upfront. This is the first podcast I ever do. So I cannot be held responsible for all the different slips that can happen. But
Sebastian: It's our fault.
Ilya Sakharov: yeah, it's your fault. So my name is Ilya. I'm based in Berlin. I've been in software engineering and software engineering industry long enough to remember that everybody was afraid of a cloud. And Kubernetes was somehow a good idea. my career started around quality and testing, which probably explains a lot about me being quite critical about stuff. But I've spent a couple of recent details, decades with platform engineering, with site reliability, developer productivity, architecture and engineering leadership. Basically. moving through my career and continuously asking the same question with increasingly more expensive answer is how to make engineers move faster without setting the production on fire. So this is also my passion. So this also took me through work with Microsoft, with RSA, with HelloFresh, Ketir, Codility. And today I'm executive director of engineering for RedCare. Red Care Pharmacy is the biggest e-pharmacy in Europe, which you probably know by the name Shopapotec. So that would be
André: true that.
Ilya Sakharov: us. But that's only just one brand out of many. Platform engineering was always a part. It's always a weak spot. I love it. I kind of like build roads and not just drive cars, but now we are adding the autonomous cars to those roads. And those are the agents, I believe. And those are occasionally hallucinate and drive into the lake here and there, but eventually platform is in my opinion, is underlying layer for everything. So I'm happy to be here and see where the discussion leads.
Sebastian: Thank you. Yes, we are also curious where this discussion leads. But before we go to the main part and the main discussion, we always ask our guests to talk a little bit about how they utilize AI in their day-to-day work and would like to do the same with you. So what kind of tools, what kind of tech are you using?
Ilya Sakharov: Yeah. So as the name of a podcast, beyond the hype, Beyond the vibe coding. I think we are beyond the models. are beyond the, is Opus 5.x better than whatever model? I think we are into durable working tools. And... Me, I've been, as a, as an engineering leader and also with managerial responsibilities, I'm finding myself in this, weird connection between utilizing AI for my, kind of corporate managerial work, but also, making sure that the organization follows, and adopts and really takes the benefit of, those tools. we moved, I mean, AI explosion happened from individual engineers, right? So they did the one engineer who. brought the GPT 3.2, was it like in March 23, right? And try to do something with it. And then they use the personal subscriptions for corporate code, right? And then someone said, okay, let's buy everybody's subscription. then, so it grew kind of bottom up from engineering, right? And I think we are in a very high level of adoption already everywhere and it's time to really consolidate. the tools that we use into something that can be controlled, usable financially, financially reasonable, right? So, and then we talk about skills, we talk about architecture, document architectural decisions. In RedCare, for example, we do have access to both frontier providers of... of AI tooling, which is OpenAI Anthropic, and we use them for different tasks. We do have internal LLM MCP gateway that allows us to separate different models for different tasks, but this is mostly purely for financial aspect because eventually we are quite big. And the spend is of course growing exponentially the more we utilize those things, right? So, and I think it doesn't really matter what you use today, right? So we have both, our main product is Codex, we use Codex mostly, Claude is happening. in parallel plus in other functions because sometimes Claude is actually better for functions that outside of engineering, our marketing department, the pricing, all of those, they're all using AI tools. And for them, we have an option to offer Claude for engineering codecs pretty much fills the gap that we need. And yeah, sorry, just
André: Are you also, are you?
Ilya Sakharov: for me, I'm using mostly for, again, for my managerial tasks and for my corporate tasks, I'm utilizing both tools and just find what works better.
André: Okay. That actually answered my question. was wondering if you're, or I wanted to ask if you're also contributing to
Ilya Sakharov: Hmm.
André: the software projects you're running or how you utilize AI or maybe the only follow-up question could be, there a harness you are using for your managerial work?
Load full transcript
Is there, I don't know, a special setup?
Ilya Sakharov: So, no, not specifically. I have my own GPTs that were created. They helped me to perform certain tasks that are relatively repeatable. The issue is that I'm also not really a big fan of calling this automation because I hear a lot of... usage of the word automation in correlation with AI, right? There is a task you need to automate, you use AI to do it. Right? And in my mind that, you know, coming from quality and from automated tests 26 years ago, right? Automation is a deterministic process, right? And whatever AI is doing isn't, right? So we invest quite a lot of effort of creating those harnesses and guardrails and, you know, agents checking agents and the... One LLM validates the output of the other LLM in all this just to make sure that we are not deviating too much from our target. But eventually those machines, though the LLMs in general, they are not made to be deterministic. So I would not call this automation.
Sebastian: I would argue that humans are not deterministic either, right? And that standard operating procedures are essentially like specs, right? So you define a process that humans, or in the past, that humans carried out. And they have been doing this not in a very deterministic way. That's why there are these process descriptions in order to limit the. pretty much the amplitude of deviation of results. And I would say that that's fairly similar to what is being done with AI, at least under the label of automation, which is not, I think, everything that is being done with AI, right?
Ilya Sakharov: Yeah, yes, but I'm not saying that humans are better in determinism than LLMs.
Sebastian: You
Ilya Sakharov: What I'm saying is that we invented automation to achieve deterministic results because humans aren't.
Sebastian: Yes, yeah, yeah, yeah.
Ilya Sakharov: So instead of repeating the same error prone, like human error prone process by hand, we introduced automation, right? And then we know we can rely on this process, okay, to come to certain conclusion in a very deterministic way. If you automate with LLMs, you don't.
Sebastian: I see where you're coming from. And I think that's also an
Ilya Sakharov: you
Sebastian: interesting philosophical discussion. Maybe we can even continue it the main section
Ilya Sakharov: Okay.
Sebastian: of our discussion. first I would say, let's move over to the main section of our podcast, to the meat as we call it. there you created or came up with a quite interesting topic. which is about cognitive. We already had in the previous discussion a bit of a discussion on the terms cognitive load leading to cognitive debt And yeah, please introduce our listeners to your ideas around this.
Ilya Sakharov: Yes, so I'm trying to look at what AI does, not only on the assisted coding or generating the next software and stuff like that, but in general, what does it do to us as humans? And it can be, whatever I have to say can be applied to engineering, but also can be similarly applied to any other. type cognitive work, right? So what happens is that we started to see lot of products and media and texts generated by LMS. And we call them among ourselves the slope, right? LinkedIn even introduced a beautiful tag that you can... You can mark a post you don't like, it looks like AI slope, and it actually gets a hit in exposure. But it also means that we are outsourcing the understanding. And this is something that I'm thinking about quite a lot. And I'm thinking about this process in which... What thinking is, right? What made a specialist? What made a person who understands how system works? Right? It's a person who wasn't involved in the building of that system. It's the person who thought about how to build that system, right? But we are closing, we are closer and closer to the point in which we're going to outsource even that. So we're going to have the spec at the beginning. We're going to have the input. We're going to have the output, but that magic in the middle, okay? makes us think, assume. So we make mistakes and redo is disappearing. We are starting to have input and output and we are sourcing in the middle. Which means that as eventually we are responsible for the things that we build, we actually are creating a system that we don't know how it works. And we don't know how... it evolved up and from the moment of conception or idea towards the product it is today. We don't know this chain of architectural decisions and this chain of changes that were made. And practically what we have is a black box. And this black box is a very difficult thing to own, right? Because I think ownership is about your ability to influence, right? and your ability to control, right? Because if you're owning something you cannot control, right? You cannot really own it. It's going to be like a strange type of ownership. It's going to be mostly scapegoating. and in this case, we're creating those systems and those systems are very far from understanding of everyone who is involved in building it. And we have no idea what that is. And we are even stepping into this, I call it the fantasy concept of magic. Right? Because if you know fantasy books, they are talking about some kind of force, some kind of, it's always some kind of force, some kind of magic that is present in this realm. Right? And some people can use it. Some people cannot use it, but nobody knows really how it works. Right? And this is exactly where we're heading. And my biggest concern with what AI is doing to us is that thing. Okay, is that we're accumulating just like technical depth, something that we made decisions that were not completely right, but we made it under the pressure of time and the circumstances. We're accumulating a cognitive depth that we will not be able even to process, even if we wanted to comprehend the system that we created. And then the question is who owns the system, who is responsible for system behavior, and what
do we need to do if we need to change the system And in some cases, we do want to change not only evolutionary as an adding additional feature, but maybe we need to fix a problem. So we're going to outsource that too. Right? So there is a problem, there is a misbehavior. Here is the stack trace. I mean, we won't have stack traces anymore. Right? And here is the problem, here is the behavior, fix it. And if it didn't fix it upfront, it will fix it on the next generation or the next generation on the next generation, while we still don't understand what the, what on earth is actually happening in our systems.
André: Yeah, I think we're already in prediction land here, but so it's really hard to tell
Ilya Sakharov: Hahaha
André: how this all will work out. I hear what you're saying. I tend to agree on the other side. I think there are also some similarities if you look back. So, for example, you mentioned in the beginning, certain technologies, right? I think not everyone is really aware how these technologies are really functioning under the hood. thought, think also today there is not 100 % control. And then
Ilya Sakharov: Mm-hmm.
André: I think the second argument, but this is at least from my point of view, already a weak one, it's always going up the ladder. So the languages, for example, we're talking about like the programming languages today, right? Like they haven't been there back in the days. Like back in the days it was very low level programming. And then people came up with abstraction, abstraction, abstractions. Like you could also say this is another abstraction. However, it's not deterministic. I think that is my main concern with that.
Ilya Sakharov: Not only that, let me address the two points. So I think we are on this very thin edge in which AI is treated as a tool, but also as already not only. So we started from a point in which AI assisted coding and all those code generators that we had, they were the tool, tool to generate. a function or a class or a file or even a small system from the spec and then a human would actually look at it, understand, fix, maybe optimize and release. So we use the tool. Usually people who are really like AI pioneers and enthusiasts, they say, okay, but you didn't complain about the calculator. Right? Calculator was introduced and they usually post this picture of people, you know, demonstrating with bandic calculators and stuff like that from the last century. It's not the same thing. Calculator is a tool, right? That performs a complicated action. But it is one, it is deterministic. And second, it is taking away the technicality. It doesn't take away the thinking. think AI is threatening us to take away the thinking. And we are overly relying on those systems to take away the thinking and the best way to do something. I think this should remain with human beings. So this is one argument, the first one. And the second one about the abstraction, it practically stems from the same place, right? Abstraction is simplification. Simplification on account of complexity. So you increase complexity, but you achieve simplified way, simpler way to do something. And here is where my problem is, is what is this something that we want to achieve to do? If it is code generation, I'm perfectly fine with it. If it is the idea behind how we build a certain engine, a 3D graphics engine, finally beat the unreal one, if it's even possible. I think it's a problem because then nobody understands how it works. And the cognitive load and as you mentioned, Sebastian, of inability to process the amount of things that AI is producing, okay, leads to cognitive depth in which even if we wanted, okay, to understand that we would open this in the files and start reading the code. and profiling it, we will not be able to. Because of the size, because of the complexity, because of the verbosity of LLMs. And then, what's left for us?
Sebastian: That's the interesting question here, right? So I also think, and there's, think, discussion currently going on that there is a risk of us humans losing the ability to think through things, right? So that we're outsourcing our thinking and are just,
Ilya Sakharov: Mm-hmm.
Sebastian: I don't know, thinking very cheaply, very, very simple in simple terms. having the AI do the complex stuff, right? And that's something that definitely right now would not work because the AI doesn't understand the concepts and doesn't have the context that is normally needed in order to do that. Maybe it could work in the future. But also they're not sure if this is something we would actually want as humans, because I think this is really pretty much a big part of what sets us apart from other. living beings,
Ilya Sakharov: You
Sebastian: living organisms. And I think in code, exactly as you mentioned, or in software engineering, we still have code as the deterministic, at least intermediate result from what we're doing with LLMs right now with agents. And I think that it's important to understand the systems. do not treat them as a black box.
Ilya Sakharov: Mm-hmm.
Sebastian: However, I also do see areas where we already pretty much have these very complicated black boxes. And I do think that this will like become a very normal element in our daily lives going forward. I think there are autopilots for planes and autonomous. So AVs are pretty much
Ilya Sakharov: Mm-hmm.
Sebastian: advancing, right? We already utilize machine learning of different variations in different kinds of, be it fraud detection or, I don't know, be it KYC processes, right? So there are numerous situations where we already are utilizing these systems in a like black box kind of way where we pretty much, train these things and then verify that the algorithms with a specific data set work according to what we expect, right? And then we let them out in the wild and then we see if they behave similarly good in the wild, right?
Ilya Sakharov: Yes, but just one comment on your
first statement about this discussion that goes about us losing the ability to think. think the bigger problem there is that we're losing the ability to learn. Right. And that already started before when the screen generation came up and all this, right. So the educational system was already kind of ringing the bell, like people, our kids are having trouble to learn. And I think this is exactly what's going to be amplified because that thinking thing in the middle is what makes us remember stuff. Right now we operate on input output. So even if the output is correct, and it is, let's say, deterministic, we will not remember the same output when we will face this problem reasonably in a longer period of time from now. If I face the problem again, I will automatically go and ask the LLM again. Okay. I will not remember because I didn't go through the hurdle, okay, of try and understand and comprehend and dissect and combine in my head, which actually creates the connection between the neurons. So I think that the problem, the physiological problem is very real. Okay. We are really losing that ability if we're going to rely on this. Regarding your second comment, I agree with you. And I think it's also inevitable. think we are going there regardless of what I think about, is it good or bad? We are going there anyway. Right. And when you started to talk, I thought about you, you're talking about complicated libraries, right? We, know, someone wrote something amazing and like, you know, again, the 3d rendering engine, right? I don't want to write it. I, I buy it. I integrate my library and here we are. I don't need to understand how it works. But you, went even more. kind of specific with things outside of coding. And I agree with you. It goes and goes and goes. I still think we are in a place where those pieces are Lego blocks and not complete solution. So right now I say, okay, I need the KYC model. Okay, I put it here, right? I need the engine. I put it here, right? And together they achieve what I need to achieve. I think we are getting to the point in which all this thinking about what I need to achieve and what what blocks is it made of, okay, is going to be completely irrelevant. And honestly, maybe I'm old, but it kind of scares me.
Sebastian: Yeah, I mean, can hardly argue with that. So it is to some extent scary to fully understand that. would say that given also your example of you put the KC bottle there than the complicated other system there, and you still have the full understanding of the system. To me, I would draw a connection to what you said earlier, that it's somewhat like magic what we're experiencing with this technology, right? And everything or every new technology initially feels like magic and later on then becomes pretty much normal, right? And I would say it's the same thing with what we experienced today. Initially, like, let's take the phone, right? So it was someone had the idea that we could send signals through wires,
right? And that on the other end, these signals could be received and then like we could have these Morse codes and could pretty much decrypt the messages, right? Which was a very like strange thing and only few people could do it. And then these could translate them, these messages into normal human messages, right? And out of that, we created landlines, we created the internet, Wi-Fi, sorry, the smartphones and whatnot, right? And nowadays it's... It's just of course, everybody has a phone
Ilya Sakharov: you
Sebastian: and that's very normal. And so I do see a world where this is happening with pretty much any
Ilya Sakharov: Mm-hmm.
Sebastian: artificial intelligence that we're utilizing in the day to day work.
Ilya Sakharov: And yeah, I think it was a Graham Bell, right? Yeah. And you know,
Sebastian: Yes, he. exactly.
Ilya Sakharov: when he built the telephone, right? The first telephone that he built, he already had two missed calls from Chuck Norris.
André: Ha ha ha.
Sebastian: We now have our outtake.
Ilya Sakharov: So yes, I think you are correct. We're going that way. Although, I look back and I look at those magic moments that happened, I think people still asked how they were. Right. And there were still people who were actually digging into the kind of nitty gritty details of how things work, how the electronics works. have electronics enthusiasts. We have those people who are creating systems at home. I don't think this conversation actually happens regarding AI. It is very, very niche. Right. is a quite specific groups that are really building those things. Right. But the consumers. I don't think they're bothering themselves to really ask the question, how does it work? And you can say the regular iPhone user also doesn't ask how the iPhone works, but
Sebastian: I was about to say that.
Ilya Sakharov: I still have hopes.
Sebastian: You
André: But also for your work context, I could imagine that we will see... a future where the organization is divided in people who are building these kinds of systems, like
Ilya Sakharov: Mm-hmm.
André: driven with automations and AI and people who are just using this without having a greater understanding what is, how this technology actually works. So to some extent, you could also say we will see some kind of Taylorism, where you have the people. working on the system, so building the system, having a good understanding for those. think critical thinking is super much needed because we're not taking from their shoulders, right? They need to build the system. And then there are people who are utilizing that. think that was also one of the latest episodes when we talked with Bastian, a friend of us. We actually worked together at Immobilien Scout. I think he made a point, or he made the observation during the AI transformation at the Getaway Group that maybe it's not the right approach that everyone needs to deal with AI technologies because not everyone is also capable of understanding that. Actually, for some people, it's already digitization is too much.
Ilya Sakharov: Yeah, we are, I mean, in our company, we are facing this quite a lot. We are trying to digitize healthcare, right? And we are seeing a very interesting groups of people who are really kind of up native people who really see this as a certain normal evolution of how you perceive, how you receive your medication, how you receive your prescriptions. but some groups are still going to local pharmacy, right? Because this is what needs to happen. regulations and government doesn't really help. yes, but just to extend, I agree with you, but just to extend this, and I think this is the, we're building the systems that are allowing people to use it in a more secure and... and confident way is exactly where this transition from current wild AI market towards platforms is something that needs to happen. And again, it started from engineering. Engineering is consuming AI and agents in a very natural and organic way. But then it spills over, spills over in other functions and other functions are not really technically savvy to create their own agents, right? So this is why the companies like N8n or something came up with low code, no code situations that allow someone to create whatever they need. I don't like those approaches because eventually the engineering is going to own that thing that they created. But I think it's also up to engineering to... to allow marketing to use AI in the proper way, which was the second point that we decided not to discuss. And that was the... Platform engineering now serves not only engineering. It's a revolutionary moment. It is sales, it is marketing, it is finance, it is legal. All of those functions are now going to rely on AI tailored products delivered to them by the technical department and within the guardrails of security, privacy and NPI protection.
Sebastian: And there the question is number one, it's maybe a leading question. Is this inevitable or avoidable?
Ilya Sakharov: I think. While we are in this wild arms race, okay, between the two AI superpowers in the world that are not really stopping, I think this is inevitable.
Sebastian: I agree.
Ilya Sakharov: I think, yeah, I think and we need to make sure that we protect what needs to be protected on our own because there will not be a solution coming from any one of the frontier model producers.
Sebastian: Yeah, this is exactly also what I thought in a sense that we need to face this situation, like we need to face the changes that are currently happening, right? And we actually need to be at the forefront so that we can steer them in the right direction in our own spheres of influences.
Ilya Sakharov: Yeah. look what's happening with security recently, right? Look at the PHP thing that happened and all of this. The amount of new vulnerabilities being found today is 10, 100 times faster or bigger than it was before the latest frontier models appeared. Right? So everyone is scanning
now things, right? Everyone is finding things and the amount is unprecedented. Now, one would say that, you know, it's actually great because we will then fix all of them. And then eventually we will get to the point where there is none, right? Because we will find all the problems and we will fix them and it's great that it's being exposed. But I think the amount of software and buggy and leaky and vulnerable software produced right now out there by the same LLMs is exponentially higher. So that's the spiral of death.
André: Yep, I was just about to ask how you see that point.
Ilya Sakharov: Spiral of Death.
Sebastian: That would be a very positive ending for the main section. Just kidding. I mean, in the end, I think that's it, right? That's the reality we have to face. I mean, we're all in different tech orgs, right? And yes, of course, two most recent weeks have been filled with like scrambling to mitigate any or to... first of all assess if there are any vulnerabilities in the systems then scrambling to mitigate I would say the eMudra was probably even depending on your tech stack but these were the two biggest ones right and yes we also have discussions going on okay how do we improve our posture so that we can react quicker how can we limit the amount of effort that this is generating so that we can still concentrate with the vast majority of our capacity on value delivering stuff as opposed to defensive security mitigations. Yeah, for sure.
André: Maybe.
Ilya Sakharov: On this, just one comment on this. think it is also a wake up call for security groups within the organizations, right? Usually security groups within the organizations, especially the large organizations that are not dealing with security on day to day, and this is not their bread and butter or business. They are reactive, right? It is the faster patch. is finding vulnerability that's already there. It's bug bounties. It is external researchers and stuff like This is how it works. I think it's time to build internal offensive security that is actually finding those things in the software that those companies produce before it reaches the customer.
Sebastian: Yes, I would argue that this is the one of the main responsibilities of every software development company already. Right. So it should actually not be new. But the amount of investment probably should increase relative relative to the overall expenditure.
Ilya Sakharov: We all know the meme of security budget before the breach and after the breach.
Sebastian: Yes.
Ilya Sakharov: I think this offensive, it's not just app scan or app penetration test that usually people do just to get the certification. It's actually putting this as part of your SDLC with proper systems tuned to attack your software as it would. be attacked in the wild.
Sebastian: Yes, definitely. so you're talking about proactive security, Yeah, fully agree there. I think, Andre, you also wanted to say something.
André: Well, like earlier to the points, I'm just saying, maybe, maybe let me, let me try to sum it up. I think everything I've heard more leads me more to the point that we as tech people should do everything to guide the organizations we are working with to... make use of the capabilities, but just to a certain extent there is a model. And I don't mean a frontier model, more a cognitive model, which gives some kind of direction how to handle that situation. think that is for me the takeaway because I think what we discussed in the last two minutes is it was actually the opposite from what we discussed before. At the moment, we are in this war where on the one side, people are really using the capabilities without having a clue what are the consequences, right? Because they don't know.
Ilya Sakharov: Mm-hmm.
André: They do this in the best interest, for sure. But at the end, doesn't matter whether engineering is responsible or not, but at a certain point, they will ask you, right? So something went down the road. What can we do? How can we fix that? On the other side, you see that even the most technical person, those who should understand how this is working, right? Are struggling because they don't know. Like we try to find like patterns, like back from what we have learned, but there is actually no pattern because this technology, this change is so massive that there is no cognitive model yet. And then think until then. We need to somehow guard this. And I'm also not a fan of this low-code thing. I'm more a fan of that forward deployed engineers. think that is one, this is a model to somehow overcome that situation. But I'm really curious if we maybe meet in a year or two, how this evolves. And if there are, I'm more or less always looking for this best practice, how to build the organization to ensure. you make use of the capabilities, but in a way that it puts your company not fully at risk. Sorry, that was a long summary.
Ilya Sakharov: No, I agree with you and I think the risk is real and because of the, and you said that the change is massive and I think the change haven't happened yet, right? Because I, right?
André: Don't scare me.
Ilya Sakharov: I think this change is just evolving and unwrapping, okay, in a form that we cannot really
André: Yeah, you're right.
Ilya Sakharov: understand what is going, what's the end state, right? Because again, if we take the internet or the telephone or the iPhone moments, they are kind of momentarily in time, right? Nobody thought it's possible, now it is. Boom, okay? And then you evolve, you make it better, bigger, brighter, battery longer, blah, right? Here, next Monday, we will wake up to the new world.
Sebastian: And that's that's actually a key point here also to give a little bit my summary or or to to your summaries. I think that number one, there is no end state like also there was no end state with previous changes, right? So the Internet has evolved several times already. The speed of change is massively increasing, so
it's being faster and faster, right? That's number one and number two. That also means that there actually is no continuous best practice, right? If you also compare to previous times, you should do extreme programming and then you should do Scrum and you have Scrum Masters and product owners. then suddenly Scrum Masters are actually not needed anymore because the companies are agile. And then you go into Kanban and then POs are not really... POs, but PMs and you follow maybe, I don't know, the Silicon Valley product, whatever. Right. So it's constantly evolving and the speed of change is increasing. And that's just the reality that we have to face and that we based on what we know in this moment, we need to make the best decisions. Right. And yes, next Monday, there is a different situation and we need to make different
Ilya Sakharov: Yeah,
Sebastian: decisions.
Ilya Sakharov: by all means, yes. And I think we still need to do extreme programming. All the rest is not worth it.
Sebastian: I cannot disagree with that statement necessarily. All right. But
André: .
Ilya Sakharov: Hahaha
Sebastian: I think that's a great ending for the main segment of our podcast.
Ilya Sakharov: Yeah, sure.
Sebastian: So let's go to the two small recurring segments towards the end. Yeah. So we always ask our guests what the fuck moments they have still with AI and also what the hottest stuff is that they see. Yeah. Do you have some for our listeners?
Ilya Sakharov: Yeah. So I think the first for the what the fuck moment is exactly that the misinterpretation of the user intent. Right. And I think it happens to all of us quite a lot when we mean something specific and we want the AI agent or a chat or whatever to do. And it just does absolutely different. And it's it sees into our soul, our true intent, right? So that we didn't express the words. But recent what the fuck moment was, again, we use also copilot with modern models and engineer just added a copilot for advice, tagged agent, copilot agent into their code to validate the states of the tests that were created for that. a particular change and that agent just interpreted it as an invitation to completely rewrite the function. And it just adjusted the function to pass the existing tests and really spiraled out of the way. So I think the learning was that be careful what you wish for and who you call and who you tag. And if you do it, do it with intention. So that would be probably the what the fact moment. It's not catastrophic, it's just weird to observe and it just amplifies the non-deterministic nature. And also maybe inability of human beings to express their intent in a clear way. And I think about the hottest stuff, again, as we passed the models, we passed this, I would say... the, not the hype, but maybe the overall enthusiasm about things. now, now kind of AI becomes part of our, our world, part of our toolbox and everything. So I think it's super important to understand the financial implication. And
it becomes more and more important, not only in engineering, again, as we said, engineering may be the owner of the tools of those tools, but it's the whole organization that uses it. And I think measuring measuring the impact not by adoption, percentage of adoption or consumption of tokens, right? That the people are kind of token maxing and everything. think it's going away because of the price now. But recently measuring how to build the economics and phenops and measuring the value of your AI usage in the end value to the customer. And I haven't found yet a comprehensive way to do it. There is a simplistic way, there are very complicated mathematical ways, but none of them feels right. And I think we are still circling around it and I would really love to see more and more thoughts about it. So for me, it's kind of hot stuff because I own the P &L.
Sebastian: Makes sense.
André: Yeah.
Sebastian: Right. Thank you. Yeah. And then last but not least, do also have a prediction for our listeners?
Ilya Sakharov: Yeah. Something that I already shared maybe with LinkedIn and others, think, and it's kind of a derivative our conversation. We are entering what I believe is the era of the disposable code in which the only thing that is constant is probably going to be the architecture, right? The architectural... scaffolding or whatever, right? And every part of the software that we want will be completely regenerated every time. So there will no longer be builds. Okay, there will no longer be CI CDs, right? Okay, there will be like regeneration of the model, maybe directly into production. So it's going to be like ultimate CD. And the issue is that if we take it further and say, okay, and people don't need to understand what's in it, then maybe extension of the existing function with the new capability that we want is going to be just another regeneration. Fix is going to be regeneration. Now, with this regeneration, because of non-deterministic behavior, the core of your function will change. And I foresee that as, if you remember some experiments were done by Microsoft and others of having two bots talking to each other. and they invented their own language, this might be the death of
André: Yep.
Ilya Sakharov: the programming language. There will be a language
André: might be.
Ilya Sakharov: created between machines, understood by machines, and probably not going to be binary. Okay, but we are going to be completely squeezed out of that process.
Sebastian: thanks for the prediction, right? closing word from my side, when we started this podcast season. So we had a podcast, three years ago, and then we stopped it and we restarted it earlier this year because of AI. we said, okay, we need to have an outlet. We need to discuss this. our first guest was, Stefan Schmidt, and he is a CTO coach in Europe. Quite, quite well known. his claim in our first episode, think was that code will go away. So that was exactly this prediction that you were just making. All right. Thanks a lot, Elyeh.
André: Thank you.
Ilya Sakharov: It was a pleasure. Thank you so much for
André: Same here.
Ilya Sakharov: having me and being patient and leading this discussion also with me. I come out a little bit smarter than I came in. So thank you for that.
Sebastian: Well, thank
André: Same here.
Sebastian: you. Yeah, we always come out a little bit smarter. Definitely. Thanks, Ilja.