Engineering video
12:25 · 5 August 2026
Engineers report using AI for 90 to 95 percent of their daily coding tasks, treating it as an assembler that builds features from pre-existing components while manually reviewing critical system parts.
Watch section from 0:09 →Developers heavily rely on tweaking their own custom skill repositories extended from open-source tools and are exploring self-hosted local models to prevent sharing proprietary codebases with commercial providers.
Watch section from 1:11 →AI excels at style transfer problems, refactoring legacy services, migration tasks, and writing tests, but struggles with system design and feature planning because it is over-optimized for the next token.
Watch section from 1:56 →To prevent skills from degrading when AI writes most of the code, engineers recommend dedicating 30 minutes to writing critical sections by hand or focusing on areas where they lack confidence.
Watch section from 3:09 →AI output improves dramatically when applied to existing codebases with established standards because the model can infer patterns, whereas starting completely from scratch tends to accumulate technical debt.
Watch section from 4:00 →The distinction between juniors and seniors now lies in the scale of features deployed across multiple systems, and technical interviews are shifting toward reviewing AI-generated code and handling larger scopes.
Watch section from 4:37 →Commercial and open-source models are becoming commodities, meaning the quality of software output depends heavily on the harness and combination of tools used rather than any single model.
Watch section from 8:38 →AI is expanding software creation into new verticals, shifting hiring demand away from traditional front-end roles toward full-stack product engineers who can leverage AI effectively.
Watch section from 10:15 →Hi folks, we here in London with Bogdan. Yesterday we had our in-person event with the software engineers we mentioned here. And of course, a lot of questions popped out. Today together with Bogdan, we're going to answer a bunch of them. Question number one, how much AI are software engineers actually using on a daily basis? So, from everybody we talked to, I would say pretty much 90 to 95%. Which means a lot of them are writing almost all the code with AI.
But we heard people being quite skeptical. I mean, they use it a lot, but they are still skeptical about the output, right? So, when it comes to using AI in production systems. We talked about workflows, we talked about reliability, what can we say about that? So, the thing is in a lot of the use cases in the enterprise, you don't need perfect code. You just need good enough code. And AI it's a great assembler, so what a lot of them do is once you have the building pieces, you have your components, you have your functions, then AI is pretty good at putting everything together and actually building the feature.
And so, yeah, a lot of them still review certain critical parts of the system manually, a lot of them spend a lot of time defining how to write code. But then AI is actually scaling there and doing the actual implementation. But what's the best AI work so? You know, what's the best way to to use it? And I guess a lot of people are asking themselves that. Yeah, so there's there's a lot of hype around uh skills and the best agents and the best model.
There's two basically trends that I saw with most of the Menties working in the industry. Pretty much all of them had their own custom skills, so they have their own skill repositories. And they were kind of tweaking them to their own use case. So a lot of them extended like an open source skill and then they take that and they extended. Uh, that's number one and number two, talking about workflows.
I think it's important to mention local AI. A lot of them are using commercial models, mostly on trapping. Fable and Up for .8, but it's also um. A lot of the companies are kind of questioning, you know, sharing their whole code base with. Private companies is not the best thing to do. So they are looking into kind of hosting their own model. The question is, hey, when it comes to software engineering, you know, what is AI?
good question, what we saw is that whenever it comes to style transfer problem, basically whenever you want to refactor a legacy service into a new language, or you want port something to a different technology, AI is extremely good. You know, it's it's basically a transformer, right? It's a language transformer. So it's really good for migration tasks, refactoring, yeah, even writing test. What is really bad at, it's, you know, kind of designing features and and understanding the system design because it's over optimized to the next thing.
Uh the one thing that we were just talking with with one of the most experienced engineers is that for the first time it's possible to change the engine while you fly. So, things that took two weeks or three weeks, now you can do on the spot and you can go much faster. But whenever you're building features and you're redesigning things, then it's always takes much longer to just think about how things will look like than go in and have your models implementing.
The other thing we saw is that there are very few people using agent loop. And the reason was they all agree that when you use agents for a long time, the output just decreases exponentially. It was literally unusable. I think one of them had fable running for five days and it was unusable and it spent way too much budget. So, a lot of them are using it supervised. And talking about that, one question that popped up was, okay, AI is writing 95% of my code, what can I do for my skills not to atrophy?
So, what we recommend and what the most experienced ones are doing, it's, you always take 30 minutes to write things by hand and usually you reserve that for the most critical parts, or the parts where you feel like you're lacking. So, let's say you're a front engineer and you're really unsure of how to do CSS transitions, then whenever it comes to that, you use AI. But when it comes to things that you're really good at or you're really confident, just use AI, get it out of the way, get it done and move on to the next day.
And I think one of the, one of the rules that people mention is the 70 30 where uh the software engineer does like 30% of the work and uh the model can do like 70%. You think that's uh that's a good rule of thumb? We're not I think it's hard to put a number on it, um but again, it's it's more about the what than how much. So if you're doing something, well, very comfy end with the standard in the code base.
And this is something a lot of them mention. It's whenever you put a model on a code base that you already have standards, the output is so good. Because it can infer from the rest of your code base. Got you. So when you start from scratch, it's when you accumulate the most technical depth. So the more wider you have, the more structure you add to it, the better output you'll get. Got you. Got you. Got you. And and that's one of the things that, you know, talking about the job market.
Right? We are in a senior only job market, companies prefer senior engineers because they're the ones that can get the best output from so many right. But what device would you give junior devs, mid-level devs who are kind of using AI and they're not so good at kind of, you know, anticipating problems. Right? How can they build that skill? So I think before AI, being a senior was mostly about getting things done end to end.
Yeah. But nowadays, that's the expectations for everybody. So, uh, we just talked to a lot of these days and pretty much everybody was kind of full stack in their team. Everybody was delivering feature. The difference today between a junior and a senior is that the surface of the features they they deploy is much bigger. So a senior would be able to deploy a feature that's sponsored across multiple systems.
Where a junior would be able to deploy something end to end inside the scope of an application or, you know, a back and front end kind of thing. Um so that's number one number two when it comes to the interviewing process, a lot of them mention that the fundamentals are the most important things. So it doesn't matter if AI writes the code, they'll ask you to read the code and tell them what's going on with it, how would you change it.
And it doesn't matter if you're going to tell an agent to do that change or you do it manually, companies don't care. The most important thing is that you need to be able to get things done end to end. So you got to use the model and just have value and a lot of them told me, you know, a lot of cases you have to go talk straight to the customer or to the product. And then release working software.
And one thing I wanted to mention right, one of the students mentioned that in their company, when it comes to the technical interview, they kind of, you know, they scrapped all of the all of the structure they had before and right now they give you a bunch of AI generated code. And they're like, hey, let's review this code together just what we treat, right? And I think that's that can be very good news for some engineers and that can be very bad news for for other engineers because it's like, you know, how deep should I go? Uh what should I talk about it? What kind of code am I going to see, etc. You know? What are the code code smells I should mention? So, you know, if you were faced with with such an interview, right? What how would you go about it? Well, I would say, number one, you need to develop that critical thinking about code and if you're not ethical at the code. So, you need to make sure you read a lot of code. Uh that means you got to work with AI, but never trust the first output, always think, okay, how would I change this? How would I iterate on this? How would I improve performance and what are the kind of rules I need to get to my agent to improve the output next time? So, again, it's you have this very powerful machine that's going to help you go much faster, but you're responsible for the quality and the quality of the output you get from an LLM. It's proportional to the quality of your thinking. And so what we see is that these tools will end up being what we call expert tools. You have folks and uh there's a lot of talk about the future of back-end and the future of front-end. Now that's a very in-depth topic, we can cover a lot. The profession, software engineering is changing, front-end engineering is changing, full-stack engineering is changing, and we are going to make a video specifically about that. We don't have time to talk about all that now, but you can already understand some of the changes that are going on. Number one, less code writing, more a lot more thinking, number two, interviews are a bit more ambiguous. You need to go very deep, you need to signal seniority very fast, etc. Just let us know in the comments if you want to touch any of those topics. And we will make another video about it. Now, uh talking about the technical interview process, right? I mean, AI totally it kind of blew up the whole you know, in the past we used to have a very clear structure, you know, life coding, system design, maybe a bank of technical questions that you will be asked to. Uh we heard the interview process change a lot. Right how does the technical interview looks in the age of AI? So, it's very early to say, but there's two things we saw. Number one, uh you will be asked to review AI reticle. So, the last year, okay, how do you improve this? How would you test this? How would you change this? How would you work with agent to do this? And the other thing we saw, and this is more prominent in the take home test right now,
is that they will give you for example they'll give you the designs, they'll give you an agent and then they'll tell you build a feature. And the scope of the feature it's much bigger than it used to be before. So that's that things get easier with AI. It's just the expectation raise and the new criteria is that can you build it? It's how fast and how higher the quality can you achieve? Because everybody can build anything with LLMs, but not everybody can build quality into it. And quality is the new differentiator, you know, all way everybody can build software. The only thing that makes the difference, it's okay, how
good does it look like, how well does it work, how reliable is And Bogdan, I know you you know, you're very much focused on on principles and software engineering fundamentals usually. And you don't like these questions, but everybody's thinking, you know, what's the best model I can use, um what's the best tool? We you know, we are software engineers, we used to love frameworks before and now it's all models.
The other thing about this question is that a lot of people will tell you, well, I use this one and uh I I had a good experience with it. But everything is very subjective. Yes. What I can tell you is we do see open source models catching up. So, I think Anthropics was leading the race with Ob. Yeah, open. Yeah. But came after, to be very honest, it's not so far away from. It's just like the no, it can work a lot longer.
And so we iterate, it's a lot more and it's different agents talking to each other and then you get a better output. But what we saw is that a lot of people also considering models like GML 5.42, which is pretty much you know, you use it, you're going to get pretty much the same output as Opers 4.8. Not label, but hey, it's open source because you can self-host it. So, I think models are getting to this point where a good model, it's it's more or less a commodity.
Sure, you can get the best best model, but would you pay, you know, whatever it takes to get that? Probably not. So I see a lot of value in hosting your own AI. A lot of value in unopinionated uh harnesses. Because the difference right now, I think it's all about the hardness. Like if you look at Cloud Code and the output, um it usually gives you a very good output with the model that's not an undropping model.
So it's more about the harness and the combination of model plus harness, then a specific model. we will know the answer because you know, some people say, hey, you know, American models are very expensive, other people say, hey, pricing doesn't matter that much. It's more about data integrity, you know, some people don't like to use so many Chinese models because like security concerns and the data they're handling.
Well, I guess it's an open question, right? Which brings us to the final question, right? Which is, okay, what about the future? You know, what about the next six months, next next 18 months? I know a lot of software engineers are not even like planning their career anymore, because, you know, it's kind of pointless. Like their things changing so fast, I disagree, but what would you tell those people? So we were just talking about this a couple of minutes with one of the engineers.
And I think the agreement was that on the aggregate, AI will probably create more software and more software engineer jobs. Yeah, right now there's certain local minima where it's kind of changed, right? A lot of people might lose their job, but other people are getting hired because of AI. Yeah, and number one reason is because now you can go much faster and you can build software in verticals where you wouldn't have software.
So one question I ask everybody was, do you feel like your team is not hiring or they fire people because of AI? And they were like, no, not at all. We're actually hiring, but we're hiring for a different position. We want we're not hiring front-end engineers as much as we used to, we want product engineer. And if you look at the requirements as quickly the old the old stack, yeah, the old front end engineer, but now they have the full stack and they need to work with AI.
So the job is definitely changing, it's not disappearing. And for the people displaying it, if you consume a lot of the marketing from Antropic, from all those companies, then yes, you know, you might think they are, you know, you're being replaced. But the technologist, as I mentioned, it's a tool for an expert, so you've got to become that expert, then you'll be in a very good mark. And by the way, folks, what's the best way to become that expert and to become that full stack product engineer? Check the LinkedIn comments, book a call with us.
If you qualify, we might get a chance to work together and Bogden personally will mentor you and help you become that AI proof. Senior engineer, with that being said, before we go, how was how was London so far? What did you enjoy the most? I love the city, I love the food, like it's one of those cities where any kind of food is good, you know, it's Italian food is good, Indian is great. Pretty much everything.
Basin was good. And, Lama, it's a big city. You get that metropolis filling, or you know, this is a big place. It's a place where things happen, and I think it's pretty exciting. Yeah, I love my time, guy. I'm really looking to go back again. Yeah, folks, for the people based in London, you guys are living in an amazing city, lots of things to do, lots of interesting people. And, uh, Bogdan and I, we're going to have, uh, we still have a bunch of people to meet, so we will see you, folks, in the next one.
