Viktor
00:00:00.050
exactly. So let's say those managers that of teams of people of five who don't have people below them, they're in trouble, right? Because the teams of five is now my, my, are my, my agents. So I am that manager of a team of five. But actually that's old news. That was half a year ago. Now I'm the manager of those managers because now I work on, with five teams in parallel, 25 of them in total. So that's two layers of management that I'm performing right now I just replaced six managers plus 25 coders all those are me now This is DevOps Paradox, episode number 375. AI writes the code. Who understands it?
Darin
00:01:46.135
An agent can write 200, 300 lines a minute of code. As a human, you've got capacity to maybe review 40 of those a minute. That's a pretty big gap in between. But bottom line, it's now becoming our jobs as software developers, as engineers, to comprehend and understand what's being put in front of us That's the end story. Is that even a true statement, Viktor?
Viktor
00:02:12.755
Yes. And I think that excluding the difference in quantity is still the same, let's say be- with before AI Would you review if… Let's say that your job is to review other people PRs. How do you approach that? Do you review it line by line?
Darin
00:02:34.042
what time of day is it? What day of week is it? Am I even on today and I'm just being helpful? I don't know. There's… I don't have a constant way of reviewing PRs
Viktor
00:02:44.616
unless I'm really interested, I tend to focus on certain things that matter, right? Interfaces, for example, I would spend much more time, if not all time, on interfaces rather than actual functions. But because if I see the interface, whether that's a API or user interface or something like that, and I see that you nailed it, that means that you understood what we're doing, There might be a critical function that I can look, but I… This is before AI, I wouldn't look at every single line. That feels like not productive
Darin
00:03:19.975
there's also a way to think about it too is code got cheap, but the understanding got expensive. But I think that phrase, code has always been cheap and the understanding has always been expensive
Viktor
00:03:32.615
Correct. The only reason why it wasn't expense- in some cases it wasn't expensive in the past is because we were not trying to understand it
Darin
00:03:41.645
Because a human wrote it, so therefore we trusted it more than having a machine write the code
Viktor
00:03:47.845
Yeah. But I g- I get questions all the time, "Oh, but how now we do we, do we review my answer is just, is the same. Conceptually, you review it in the same way you were reviewing it before, Except that maybe you would be doing more of the critical part of the review than you were doing before, and less of noise, If you rev- if you were reviewing every, every line of diff before, that's mainly because you could, not because you should And now you can't because of the quantity. But it's not that you should have review every single line. Oh, let me find the uh, missing comma in, in the documentation Let me go really in detail word by word and find orthographical errors That's a waste of your time. Now, it wasn't before, again, because you could, not because you should
Darin
00:04:48.182
Or we're looking at tests staying green. The only thing is, what you don't know is the AI decided, "Hey, that test really isn't helpful. Let's get rid of it. L- let's stub it back out
Viktor
00:05:03.712
you know how many sonar errors were fixed like that in the past because a manager said, we need 80% coverage," and everything always green. You put unrealistic demands, and that's what people do
Darin
00:05:18.752
Yep, but then it surfaces back six, 18 months later, whatever, and it's like, "Oh, yeah, we commented that test out. Why?" I don't know."
Viktor
00:05:26.892
The problem with AI is that AI is very much trained on our behavior and is doing very much things that we normally do, and we don't like it it's Corivo, right?
Darin
00:05:42.562
Well, it's reflecting back to us who we actually are, and we don't like seeing the bad part of ourselves
Viktor
00:05:48.962
I often complain how AI will always agree with you or almost always agree with you unless you explicitly say no, be critical. Really try to disprove me." Unless you explicitly say that, it will agree with you, right? You know who behaves like that? Most of the workforce in companies When a big shot manager comes and says "No, no, no, we decided to do this," most people will agree even if it's stupid There is that I'm not sure how to qualify it. Let's say fear of authority, and when authority comes, "Oh, he's a super senior engineer. Who am I to disagree with that person?" So AI does the same things as, as we do. What's the acronym?
Darin
00:06:29.970
The acronym is HiPPO, right? Or something like that. Highest paid person in the room
Darin
00:06:35.288
that you're going to, to defer to them because, hey, they're the highest paid person, so they must know what's going on.
Darin
00:06:41.846
Hint, they don't. a lot of times they don't we would like for them to be, but it's those decisions that then fall out everywhere else. The other way to think about it too is used to we had a gate. If you were a senior reviewing junior's work, junior could not produce as much work that the senior could actually review, even with the senior going to a bunch of meetings and going to play golf. But now it's flipped. If you use AI now as the junior, which knows more than the senior, but it doesn't understand the business as well as the senior, now it's flipped.
Viktor
00:07:16.716
Again, this is human behavior. Very often those seniors know less than you, but you will agree with them
Darin
00:07:22.886
Correct. But now I'm able to, as the junior, I'm able, as AI, I'm able to generate more code than I was as a human junior. If I'm a junior ju- a human junior using AI, I'm still going to overtake that senior with things to review.
Darin
00:07:45.078
because hey, I don't understand how this … Let's use it this way. I don't know how this Power Builder app actually works because it was created before I was born leave that there for a minute. Let's let that breathe. But hey, this AI's helping me figure it out. And you as the senior that's been in that company for however many years, you're familiar with that PowerBuilder app, but now you've got this junior that's making changes to it that you have to review, but you know all the places where the bodies are buried And the AI doesn't know it as well. So now you really have to spend time going through it because somebody said, "Hey, we need to update this app and we don't want to re-platform it Now you've got a whole… You've got an AI that doesn't understand the business part of it. You've got a junior using an AI that's trying to figure out tech that's 25-plus years old, and you've got a senior that's trying to deal with the fallout of what the h- human junior used AI to produce That just sounds like a great Monday to me
Darin
00:08:48.174
What are we supposed to do? I mean, As a senior, we're supposed to know what's going on. And assuming that we do, we're still finite humans that can only process so much data within a finite amir- amount of time
Viktor
00:09:03.105
what's going on depends on the level from which you're looking at. Let's say, okay, you're an AWS expert, right? Would you agree that you need to know what's going on? At least I would. But at what level? When something goes wrong, do you understand what's happening with hypervisors in AWS?
Viktor
00:09:25.565
There we go. Exactly, So you say, okay that's not the level that I care about, right? That's somebody else's problem. Somebody in AWS or machine or I don't know. So some- somebody's taking care of. I don't care about those things. I care, I care about higher level issues in AWS, Those that are surfaced for, to me, those that are, that are accessible to me, right? And even if you have access to low-level details, you… Let's say observability, right? The way how you figure out what's wrong is that you start from high up and then you drill down, down, further down, But at one moment stop, you stop and say, "I found it." And you keep going down only if you don't find it, right? Most of the time, you're not at the lowest level in case of observability, right? You got just progress down down, down, down, down and, and then you stop and say, "I got it." So it-- There, there's nothing wrong not going into all the details when you don't have to, and that's the key
Viktor
00:10:34.747
Probably don't. And even if they do matter, right? there will be a person who says "No, no, no, details matter." And I will ask that person, "Okay, which details?" Kind of, You're writing Java," "Yes." "Oh, you need to know this and that." But do you know JVM inside out? Most people don't, and they never, they would never consider that even, So nobody goes as low as you can go. But then we complain that oh no, my level of lowness is the correct one. Who says that?
Darin
00:11:06.839
Until you get to the point where you're talking about transistors, you're not going deep enough.
Viktor
00:11:12.043
Yeah. Exactly. Yeah. Oh, I ca- can observe inside the chip what the heck is going on I'm, I'm a bit
Darin
00:11:20.571
but even that becomes flawed because every chip produced off the same line is not the same. There's this concept called binning, right? To make it s- simple you test the chip afterwards, and the chip was either 100% correct or it was 80% correct. 100% correct goes on down to retail. 80% correct, it goes into a different bin that we'll use for something else, 'cause it's 80% good for something else, just not for retail It, it becomes a ridiculous argument that people just want to be correct and want to win. And at some point, that's just becoming useless of actually solving the real problems at hand. Again, if we are building things today, if you think about some of the apps today, let me ask you this. Okay, Viktor, you have AgentDeck, open source project Do you fully comprehend everything that is happening within Agentix?
Viktor
00:12:23.679
no, it doesn't bother me. It did bother me in the past. It doesn't bother me now
Darin
00:12:28.499
Why? Why is that? Because it seems like at some point, because you're using Rust to write that. Rust is a language, and Rust is not your native language
Darin
00:12:40.189
So you, going back to the Java example you had a moment ago, you do not understand the internals of Rust
Darin
00:12:49.049
But if you did understand, let's say the version of Rust that you're using right now, you did understand the internals and you felt comfortable doing that, but now you're going to upgrade to a new version of Rust. Are you going to feel comfortable with that new version of Rust internals?
Viktor
00:13:01.379
The key over here is that something switched in me that I understood that I'm not a coder anymore. I'm a manager of choose any type you want. I don't care, I think we spoke about that before. were you a lead or something like that, manager, whatever, Darin?
Viktor
00:13:21.100
Okay. Did you go into every single detail produced by every single person you, in your team?
Viktor
00:13:30.354
Exactly. And you probably made sure that actually you do your part well so that they can do their part well because their part partly depends on you, It's a different type of job, right? And you couldn't because if you had, let's say it was a team of five people if you could analyze everything that each one of the- of them really did in detail, that means that you would be doing work for five people. You couldn't do that. You're not superhuman. I'm sorry I said it. You're not, Something switches when you decide to change the role
Darin
00:14:04.069
Because once you understand that shift, if you can make that shift to where you're no longer just a senior software developer, but now you are a senior software agent manager. I'm making up new titles today. the point, once you understand that, once you make that switch, it's like, "Oh, I'm just telling them what to do without getting any back talk from them." Hopefully
Viktor
00:14:28.734
Correct. but there is a transition period, right? I'm now saying for myself, I don't know how other people experience it, but it took me a while to get there, right? Start with one agent, doesn't produce what you want. Is it fault of the agent? Sometimes. Model, sometimes. Sometimes mine. How can we improve this? Better skills, better PRDs, and so on and so forth, right? you reach the point and say, "Okay, now a month or a year later, you and me, we are doing actually a good job. We are not failing as frequently as we did before." You and me, You say, "Okay, maybe actually there can be two of you." Now I'm ready for two of you now And then that does not work as well as, as one of you. And then we go through the same cycle and say, "Okay, now we're a good team. It's a team of three. Let's make it four. Let's make it five." Until you eventually reach the limit and say "I, I cannot have more than five of you." But then you keep improving, and maybe you reach six, and maybe you reach seven, maybe hundred, who knows? Because the more I trust you as a person or you as an agent, the more I trust my ability to convey what is needed, and the more I trust you to do it the less you and me need to interact together, right?
Darin
00:15:51.107
That's correct. Because we've built up, if it's an AI agent or even a human, we've built up memories between each other over time
Viktor
00:16:00.559
E-exactly. We know how we operate. we know how much we should spend together. With some people it's more, with other people it's less. We know how to communicate. If you noticed, even people when they talk to each other, let's say good friends. Let's say you have two good friends. If you pay close attention, you will notice that you speak differently with each of those friends, Because this-- there is a different rapport between two-- One and one of them and you, and the other them and, and you, right? As long as we can adapt to those things, we can move much faster, right? We can convey things much faster. We can trust each other much faster, and that's equally s- equally valid for humans as for agents. Just it takes time. I would, for example, never recommend anybody who is just starting now with the agents to adopt Agent Deck and say, "Okay, run five, six, seven of them in parallel for one feature and then do five features in parallel overnight and let it rip." Th-that's not where you start. And that's the same thing for management. A newbie manager who just finished high school or university shouldn't be given a large team of something important
Darin
00:17:12.282
Why would you do that? Yeah. What you're also leaning into a little bit is what's been termed as AI brain fry. You felt it, I felt it. But let me ask the question this way. Actually, if you don't know, understand what brain fry is, it's context switching at the speed of light, because that's what's happening. It's like, oh this window just finished, this window just finished. If you go back and think about brain fry six months ago to where brain fry is today for you, is brain fry the same or is it different?
Viktor
00:17:45.542
Okay, let me ask you a sub-question first before I answer. For the same number of tasks being done in parallel or, or no?
Darin
00:17:53.712
Let's assume… Okay, let me give you the scenario. You've got five subtasks going in parallel right now, six months ago and today. Tell me
Viktor
00:18:02.816
Okay, then Brain Fry much bigger six months ago. The difference is that in the level of autonomy, six months ago, I would be intervening much more frequently, So I… Let me give you two extremes, and those are extremes. Don't take it seriously because the truth is somewhere in between. But I'm going to extreme now. Six months ago, I would actually review let's say, maybe not every line, but every batch of work. "Oh, this is a feature. It has 10 tasks. Do the first task." I check it, be corrected, course-corrected, do it again maybe, or just continue, and then the next one. So I would, in that feature, let's say, I would intervene. I'm throwing random number 50 times And now I'm intervening, let's say five at the beginning and then four more times at the end because the end might not be exactly how I want it, so give it a couple of tries, right? And that means that actually I spend more time on one thing at a time because I'm going to review the whole feature being done instead of reviewing parts of that feature and then switching. Oh, you've worked for five minutes on this one, and while you're working on five minutes on, let's say, 15 minutes on this one, I'm going to check for the, the next one did 15 minutes of the work of next one. Now cool or not cool, corrected, and next, next, next. So my frequency between iterating between those five would be, let's say, every 15 minutes, and your head starts hurting very fast on that, Now I would spend on that feature not 15 minutes. I would spend maybe an hour, but at a time, but it's complete it's not 10 times 15 minutes. It's not that the, the frequency of switching are, in my case, much lower now
Darin
00:19:57.251
So I guess the point that I was trying to get to is over time you've built up these muscles, just like working out physically, that it's become easier. Not so… Well, okay, easier isn't the right word. It's become manageable
Viktor
00:20:16.677
That's why I ask how many, whether the number is the same, right? It's equally hard, I think, because the number increases over time. Six months ago, I was not working more than one or two, three things in parallel on average Let's say two things
Viktor
00:20:34.557
Yeah, because if I work on two things and intervene frequently, that means that by the time I finish with the first one, the second one is already done and I can jump to the second one. And by the time second one is-- I'm done with the second one, the first one is already done in the next loop, So I cannot do more than two in parallel. I mean, I could do more than two, but if the other one is finished by the time, yeah, I'm finished with the first one, then what's the point in working with more than two? But you need to increase the amount of work that one of them is doing, then you can do more, right? ' Cause then my agents now work autonomously for hours, not minutes on a given thing
Darin
00:21:23.478
So what does a workday look like for someone once they realize they're no longer producing or being a producer, but they're being a comprehender? Com- I'm making, making up words today. Compre- like, they're just trying to understand what's going on. Again, that sounds like a manager. Most managers are not producing on a day-to-day basis. They're just trying to understand what's going on and how can I help? But the next problem to that is, how do I actually measure that? How do I make sure that, I'm not producing, but yet I'm still have to be measured
Viktor
00:21:57.363
Probably the closest of the existing roles would be product manager Right? i'm trying to figure out what the product is and should be and where it goes, and whether what's done is aligned with that vision. That's give or take product manager job.
Darin
00:22:15.887
Which was a role that didn't exist, as we found out on the episode when we had Shuba on. Didn't exist until what? The mid-'20s, 2010, somewhere in
Viktor
00:22:25.851
Yeah, something like that. Yeah. I mean, I know that when I was young and stupid, I did not hear for the word product manager. It didn't exist back then. It was a manager, period. But, you know, it's like, what was the Palantir word we come up with new names? Because man- it's not that managers were not doing product work before, just to be clear.
Viktor
00:22:53.046
maybe it's not, I don't know. But it's not that role did not exist. It didn't, did not exist in that narrower scope or with that name
Darin
00:23:04.229
Let's look at it a little bit differently. We talked about, a couple episodes ago, about hiring and trying to get people in, especially the juniors. And most junior hiring is down, at least as of 2026. For the juniors that are there, they may be just being filled by seniors because they needed a job. So the company's able to get a more experienced person for less money. If you were hiring a junior developer today, like you, you had an open requisition, would you spend the money, A, on hiring a person that is truly junior level? Or B, would you use that same amount of money for agent usage for your existing senior?
Viktor
00:23:50.025
the latter without doubt. And that's not thinking about the future, just to be clear. 'Cause if I think about futures, I need future seniors, which if you follow my advice, you would not have, right? But if I ignore the future, yeah the latter is the choice
Darin
00:24:07.336
how would we plan for the future then? 'Cause if I'm just trying to solve a short-term six-month, 12-month problem, get the credits. But if I'm talking 36 months because the guy, the senior that we're doing this for, I know he's on his way out in three years, right? 'Cause he's, he's at retirement age, so he's going to be gone. What am I gonna do?
Viktor
00:24:27.936
didn't we have exactly the same problem before? I, I, I constantly get confused how we label something as new and… I was quite a few times in conversations where people, "Yeah, we need 50 more people." And I would very often say no, we don't. We need three. They need to really know what they're doing." Well, at the same time, recognizing that somehow those three that really know what they're doing were juniors at some point as well. But that's what we really need
Darin
00:24:57.169
I don't think in the past we could have had this same argument to where I don't need a junior anymore, I just need to augment my mid and seniors with AI, and that'll be more valuable to me as a business owner than hiring a junior
Viktor
00:25:13.350
okay, let's put it this way. We need people who… I prefer, without a doubt, with or without AI, it doesn't matter. I prefer fewer people who really know what they're doing than a larger number of people who don't, any time of the day. Now, there are things that is the waste of time of those people, and that's where juniors come in. This might sound disrespectful, but that's how you learn. You, You learn the… You go through, through the drill when you're young. So on larger organizations, we always needed, and I still think we need juniors, It's just that they don't count. First, we figure out what are the selected few that can really solve this problem And then how do we bring John? He's a kid from Bronx who will do the things that those three shouldn't be doing Because their skills are more valuable as well, elsewhere, right? They need to be concentrated, and they don't like doing it hard, right? So John, take a shovel. There is a pile of you know what waiting for you. this might sound disrespectful, but I think that's the cycle through which we go. We all, We all went through a pile of stuff we didn't want at some point
Darin
00:26:31.453
here's the problem though. That pile of stuff is getting smaller and smaller in some places
Viktor
00:26:36.113
No, it's not. It's not getting smaller and smaller. It's absolutely not because when you say this pile is getting smaller and smaller is, is that you assume without even thinking, maybe subconsciously, that we do the same volume that we were doing before, and that we are also more focused than we were before on cleaning up the mess, which we are not. The pile is equally big or even bigger because now we do much more The scope just increased tenfold So the pile is there
Darin
00:27:08.203
So if I'm having to pay money for the juniors, I'm losing senior time because they're mentoring the juniors, and I'm spending money on agent usage, that seems rational locally, like within my team, but apply that at a enterprise scale and that's catastrophic cash-wise, potentially I don't know how you would solve that. We're pushing on the boundaries on every different axis to see how it works, and eventually that bubble is going to pop. I'm not saying the AI bubble's gonna pop. It will, but this is a different level of problem because we have a new tool now that we haven't had in the past
Viktor
00:27:52.909
Because our competitors are having new tool that we haven't-- that v- there wa- it wasn't there in the past. Most companies would be very happy to employ AI to solve much more efficiently the problems that they were solving in the past. And you can do it with fewer people and much cheaper, if your goals are the same as they were five years ago, I can guarantee, unless you're really doing something really bad, with AI you can do it with fewer people and cheaper. The problem is that your competition is not thinking that way. Your competition is now releasing hourly. look at Claude Code. The-- Are you still reading release notes for Claude Code?
Viktor
00:28:35.592
I give up as well because it's-- nobody can follow anymore. That's how fast it is, and that's my competition now. I mean, Not my competition, but somebody's competition, right? And that's the crux of the problem, right? That's why we don't see savings. That's why we he- see higher rate of everything, because the scope just changed drastically
Darin
00:28:55.332
And understanding that, if y- 'cause you mentioned if you as a company still have your same goals as you did five years ago, it's game over for you. Because the competition is thinking, okay, you said 10X, even if they're doing 2X what they were doing before. Let's say that classically, an application coming out of an enterprise was released once a year, right? That was a real thing. You can laugh at it, but that was a real thing. Sometimes multiple years before you get a release in between, but let's say it was once a year. And now they have moved and they're able to provide f- let's say it's a bank, ' cause we've used banks before. We'll use a bank here. A bank, and they typically update their website once a year, because that's when they get a new CMO. That's an inside joke that's not really an inside joke. But now, because of the use of AI that they have within their company, they're able to update and add new features for their customers on a monthly basis. So they have moved from once a year to 12 times a year And you, bank uh, that is, that would be a miracle, let's put it that way. But you at bank B are still working on your once-a-year update cycle because that's, th- that's all you can do, and your people have decided that, no, we're not gonna use AI to help us out. We're just gonna keep throwing bodies at it because it'll be, In fact, we're not just gonna throw bodies at it, we're gonna th- we're gonna go get some high-paid consulting companies to come in and help us, where they back up the bus and dump people out. and charge us 300, 400, $500 an hour, e- even though the people actually doing the work are only getting paid $30 an hour. you know how this is gonna turn out Both will have problems. I could argue both will fail, but the chances of me shipping new features every month to keep my customers happy, guess how many customers may be moving over from bank B to bank A I'd say could be pretty large
Viktor
00:30:46.875
it's case of banking is not that much of a question how many people move because m- not many people change banks. It's very sticky. But you're not getting younglings anymore In Europe, I don't know in US, but 18-year-old is going to get Revolut. It's our fintech type of bank. It's not getting your traditional bank. They are not getting ne-new kids anymore.
Darin
00:31:11.852
Yeah, because do kids even know what a check is? Most people don't. My daughter doesn't. She knows what it is, but she doesn't have any. She doesn't need them. She does everything online Once we understand that understanding is the game now, by the way, understanding was always the game. We just thought we were coders that needed to understand the specs that were… being handed to us
Viktor
00:31:34.944
That's the part that always annoyed me with our industry. It's kind of I expect somebody to write requirements for me and I'm going to follow them to the letter And it annoyed the heck out of me in the before, and it annoys me today, and it's the same situation. You need to start thinking, man. Use the brain cells, please
Darin
00:31:56.194
But what you just said there means, ' cause I could, I could draw that string through that you should not be using AI agents at all, period. Because,
Darin
00:32:08.848
b- because you… It's like, I want somebody to write the spec, and if you're not gonna be the one writing the spec for the people, Viktor, for the agents, then who's gonna write the spec?
Viktor
00:32:17.618
that's the thing. Whomever is the one writing the spec before agents is the one who should continue writing the spec, and that's the person who should be using agents, and you just got fired Right? Because that's the game now. The game is, can you define what you want and can you validate that you got what you want? That's the job. If you're incapable of defining what should be done, and when I say defining, I do not necessarily mean somebody writes the spec and I then rewrite that, copy and paste that spec into Claude. If you're not the one writing the spec, and if you're not validating, so you're not the product manager in what you do, not necessarily role, and if you're not the, let's say, tech lead, the one that approves PRs, if you're not one of those two, you have no job here
Darin
00:33:12.029
That makes all the sense in the world, except not everybody understands the business. Are you saying now that we no longer need people just to write code?
Viktor
00:33:22.669
Oh, I'm definitely saying that writing code is becoming irrelevant. I would argue that it was never the high value The high value was in all other roles. Architects, product managers testers. Rather than uh, a glorified typist
Darin
00:33:39.998
With that glorified typist, are you stating that as a business owner, I get more value out of the cleaning service that comes through at night and cleans the office than I do out of my person that's writing code?
Darin
00:33:56.054
Okay. But even so, I'm saying that a human coder is valued less than a cleaning service within the company
Viktor
00:34:04.526
Yeah, correct. Now, coder is more than qualified to become a product manager Best product managers are people who actually know the stuff. I really don't like working with managers of any kind that don't know what's going on. Those are bad managers. You have potential to become a good one, but you need to become a good one ' Let's put it this way, you need to be a very technical product manager. And the question is, who is more qualified to become that? A manager like, you know, uh, I know how to do things in Excel, or a coder?
Viktor
00:34:46.646
It's a coder. It's a coder. You have all the opportunity. when developers complain kind of, "AI is going to take away our jobs," that could only be true if you don't take away managers' jobs
Viktor
00:35:01.038
developers are under threat with AI if they don't put under threat management jobs So one of the two is not welcome, And I would feel much more scared for myself right now if I would be a typical stereotyped manager than developer Much more. Because I think the developers are more equipped for this new management style than managers are. They need to learn new skills. They need to stop being nerds. I am one, just to be clear. So it's not free. But if anybody is, is, should be afraid, that's traditional managers
Darin
00:35:43.435
And by traditional manager, you mean they fill out the annual reports or the annual reviews and blah, blah, blah, blah, blah, and that's all that they really do
Viktor
00:35:52.202
the one that gives you a ser- uh, what's the n- what's the word? Uh, The one that gives you specs that you follow to the letter, and the one who wrote the specs has no idea what you're doing, and you have no idea what the product really is or should be. those are the two roles, and they need to combine, and the question is who is better fitted for that new combined role?
Darin
00:36:17.362
You're talking about collapsing of jobs. If I can have a technical person that understands the business, that's gonna be worth more to me than a person that doesn't really understand the business, but they sort of know and understand how to manage people
Darin
00:36:37.192
Therefore, if you're just a people manager and don't care about the business, this is why mid-manag- mid-level managers are starting to disappear in certain places
Viktor
00:36:46.063
exactly. So let's say those managers that of teams of people of five who don't have people below them, they're in trouble, right? Because the teams of five is now my, my, are my, my agents. So I am that manager of a team of five. But actually that's old news. That was half a year ago. Now I'm the manager of those managers because now I work on, with five teams in parallel, 25 of them in total. So that's two layers of management that I'm performing right now I just replaced six managers plus 25 coders all those are me now
Darin
00:37:30.007
And that's what businesses are going to expect. As soon as they start paying for your tokens, that's what they're going to expect because that's the bill of goods they've been sold about how AI is going to help with productivity
Darin
00:37:46.564
let's go back to Joe from the Bon- Joe from the Bronx that you said. He's got the shovel to deal with the manure. But let's not look at it as manure because, sure, there's the manure there, but let's make sure in the manure that there's some hay in there, too, meaning there's some validation work to be done as well.
Darin
00:38:06.114
Because then that's building our pipeline of getting Joe the junior built up towards Joe the mid-level to Joe the senior. Because if they are just doing the grunt work, sure, that's helpful, but that's not helpful for Joe in growing and giving me so I don't have-- so Joe doesn't get ticked off and leave after only dealing with grunt work. It's like give him three days of grunt and give him two days of validation work,
Darin
00:38:48.851
that's what we want. We still need to hire… Again, people are going to leave the company, they're going to retire, something. They're just no longer working for you as a business owner, and you're going to have to make sure that you've got somebody to keep your business going. Otherwise, it's you. Now, unless you're that person to where you can do everything yourself
Viktor
00:39:07.831
It's all about choices. They are just I might be correct, I might be wrong, but that's the description of the jobs that are in front of us. You want the job, you don't want the job You can keep your old job as long as it exists It's just that with the note that your old job, if it keeps existing, that's fine, you can keep it, but in the future when it stops existing, you might realize that you're not needed anywhere anywhere anymore. That's the DB2 expert story, right? if you retire before your company gets rid of DB2, power to you But if it turns out that the company stops using it before you retire, you're in trouble
Darin
00:39:49.479
What we're seeing now is throughput is going up, but understanding is going down as a ratio, We were able to produce X amount before, now we're able to produce 10X, but the wetware, our brains, can only still deal with the 1X.
Viktor
00:40:04.951
Again, that was always, you know, when I was young we were not working on this scale, man. If I had a server, proper server of a single, as a single thing, I was happy, right? Back in the day, I knew literally every single line of code that, of an application because there weren't many Those things were h- are happening all the time. understanding of, the systems that I've been working on 10 years ago is much lower than 20 years ago, which is lower than 30 years ago
Darin
00:40:37.334
I'll give you a human example of this. We're talking about br- the, the throughput is a lot higher. We're getting a lot more code. I remember the first time I dealt with an offshore solution for producing software. up to that point, it's like, okay we're, we're in a good balance of reviewing. This offshore team was great, and they implemented what was written, which was fine, but they were implementing it at speed. This was back at the turn of the century. Makes me feel old there. You know, Around 2000. And they were producing code, I mean, just produc- they were cranking it out like a factory. Until finally management had to come back and say, "Hey, look, you guys need to go take a break for a couple of weeks 'cause we're behind now. We can't keep up." Now it's just automation that's getting us to that point So we've got to figure out how to deal with this, A, if we're going to stay employed, and B, how we're going to help the juniors come along that we do get so that they want to stay and continue on in building their career with your company. I don't know how we're going to do that other than offering them a lot of money, and making sure that they're happy and getting to do what they want, understanding, and this is going to be a communication part. Look, because most kids, especially… Okay, let's put it this way. Guys under 25, our brains aren't fully formed. I could even argue that guys under 60, our brains aren't fully formed So there's pride that jumps in. It's like, "Oh, I, I know better than anybody else." Let me tell you, son, you don't Very simple. You don't. ' Cause I've seen things over 40 years in business that you will never see. Thankfully, you will never see. But you're gonna see things that I will have never thought I would have seen in my lifetime, much less my daughter's lifetime or my grandkids' lifetime So how are we going to keep track of that? How are we going to measure these things? for the IC right now you're the senior IC, you've got to dial up your comprehension muscle. Viktor said, going back to Brain Fry, six months ago Brain Fry was fast and furious. Today it's not as bad, just different
Viktor
00:42:47.312
Just to be 100% clear, I think that the, amount of information that goes through my head is higher now than half a year, which is higher than two years ago. So it depends how you define comprehension. If anything, I understand more now than I ever did Now, as a percentage of the system, less, right? Because I'm building five features a day. So in absolute numbers, my understanding is higher. As a percentage of what was built, lower If that makes sense because the only thing I'm doing is processing information now. That's the only thing I'm doing
Darin
00:43:30.646
You're, you are coming up with the Boolean flag. Is it right or is it wrong? And do I continue on from there if it's correct?
Viktor
00:43:41.787
That's the thing, right? It's-- the quantity is not smaller. That's what I want to stress. As a percentage of the system, maybe
Darin
00:43:49.927
But also, are you concerned as much about retention? We're talking about comprehension, that's one thing, but do you care as much about retention anymore? Because if I've worked with my agent long enough and I've got good memories built up and I've got, enough to where I'm not starting from scratch every time, I don't care whether I retain that information or not because I know where it is
Viktor
00:44:13.167
that's where AI is very g- in Agentic AI, to be more precise, is very good. Finally, I can do something that I could do only with Michael before, and that's kind of, "Okay, explain to me this in detail." Anything but trivial code base, you needed a person, the one that started the project in the first place and hopefully never left the company to explain it to you. Anything but a trivial one. Now I can talk, "Okay, I don't understand this. Ex- explain it to me." And I get the answer in a minute. I could not do that. So if you want to know this, the, w- what I'm, where I'm going, if you do want to know how the system works, this is now easier than it ever was, unless you're the, literally the person who wrote the whole system line by line yourself
Darin
00:45:01.772
Yeah, but if you wrote it line by line 10 years ago, what's the chance of you remembering every little nuance?
Viktor
00:45:07.990
and anything but trivial projects is not you alone anymore. That's the point. So I don't understand the complaints. Oh, I don't understand anymore what, what's going on. That's only if you don't want to
Darin
00:45:19.048
as that senior IC, you've got to build up being able to, what we were just talking about here, building up your comprehension, but understanding how to comprehend it fast. think about it. Used to, you'd be hired into a new place as a senior because you've had time, and now you're assigned to this new system and you've got to get up to speed. What did that look like back in the '90s, early 2000s? You look at some Word documents, you might poke at the code, and in about three years you might feel okay with it
Darin
00:45:52.730
Yes. That was already six months out of date because there's been releases since then
Viktor
00:45:57.126
Actually, the people will think that we're joking, but I remember reading it on paper, like physical paper, you know, made from wood and all that stuff
Darin
00:46:14.747
An 80 by 24 LCD terminal That eight-- That 80 by 24, by the way, is characters, not pixels. Characters. For managers, if you're deciding that you're gonna go ahead and just gut your junior pipeline this quarter because you gotta save some money and you're gonna throw it all to AI agent spend, that's fine. But what are you thinking about five years in the future when you've gotta have somebody maintain these systems? Or is that just not your problem? Either way, could be fine So what are you going to do? I would say pick a system that you're trying to learn. I heard somebody talking about this today. What's the first thing that you do when you get a new model? Or what's, what's the first thing you should do now when you get access to a new model? Point it at whatever your app is and say, " Find the problems," and walk away
Viktor
00:47:05.292
One thing I do with new models is I pick a task that I'm about to do that day and give it to whichever model I was using yesterday and the new one in parallel Then you measure the time, the error rate, all the stuff.
Darin
00:47:26.062
These are the things you've got to start doing now. So as a senior IC, you've got to start figuring out, okay, what am I doing? How can I use these tools to become better? Again, we talked about this on a couple episodes ago. If you're choosing to, it's like, uh, this AI stuff's a fad I'm not gonna do it, I have a, I'm against it morally, great. I'm happy for you. You're probably not gonna be working at that place for much longer. Now, of course, we know everything's gonna go into a cycle, and people are gonna freak out, and we'll be overspending, so we have to … what is that next thing that's going to happen, right? We're talking about getting rid of spend when things are bad. gonna happen when we've overspent our token and now we're in trouble? We gotta pay the token bill. It's sort of like, I guess, a drug. It's like, well, this is how we're getting things done now, so we have to come up with more money, but we're not making more sales, so how are we gonna deal with this? Figure out your balance. Manager, your job is hard. Senior, your job is hard. Juniors, finding a job is hard So what do you think? Head over to the Slack workspace, go to the podcast channel, and leave your comments there