the "AI Engineer" scam tech CEOs don't want you to know about

So this is Mark Zügenberg, who says that probably we'll have an AI, like a mid-level engineer in your company, who can write code. He mentioned that some other companies are working on it. But it's not only that, right? You have Mark Benioff from Salesforce, saying that they're not adding any more software engineers in the next year, 2025, because they have increased the productivity by more than 20% using AI.

Actually, I think they have about 200 openings for software engineers, but still, that's what they say. Which is why today we will talk about the big tech lie, the AI-driven engineers. Together with Bogdan, we will do a first-principal analysis of AI-driven software development, we will look at the data, and we will use the scientific method to really understand what's happening, because I think there's a lot of confusion right now, and a lot of talented engineers are questioning their career choices when they actually should not, and you will see later on why.

And the fact is, just like in the Batman movie, that you always fear what you don't understand. So let's understand this thing. Let's really dive deep and understand what is happening right now in the software development market and in AI-driven development. And here is what we are going to cover. Number one, we will do a first-principal analysis of the existing AI tooling. Then we're going to look at the myth of AI-driven layoffs and we're going to find out what's actually behind the layoffs, the tech layoffs happening.

Why AI actually means more coders and not less, also called the Jevons paradox in economics. Next, we will look at the GPT-4 hangover and what happened upon its release. And finally, we'll talk about what can you do as a software engineer? You're probably thinking, do I even have a future? Is it even worth it, just like our students? And we are here to give you the tools and to show you how you can make the most out of this.

If you don't know us, my name is Dragosh, and together with my brother Bogdan, here in this call, we founded the senior dev about four years ago, where we help JavaScript focused engineers level up to a senior level in the most effective way possible. So let's talk about AI tooling and let's make a first-principal analysis to what this AI tooling actually is based upon and what does it does. eliminate for the software engineer. First of all, it's not about AI, what we are talking about is actually an LLM, a large language model, which is behind a chat interface or a code editor. And what does that mean for software engineers, bug? So, what it means is that at the core of the AI, as they call it, um there's just a large language model, which is built on a neural network with a lot of parameters that was fed a lot of code. In the case of GPT 3 and 4, the code comes from GitHub, like all the open source code we have, basically in the world, and then stack overflow. And so you feed all that to a neural network, and then you have this capability of linguistically suggest code or solutions, and because it's built at such a scale, it really feels that it can build, you know, things like a small react application by itself. And so a lot of people think that's intelligence, right? That it's basically intelligent because of the quality of code sample it gives back. What open AI did very well and what new companies are doing is to rebrand what a large language model is, which is not very salesy and very flashy into AI, AGI, and what's the latest one? Super artificial AGI, super artificial intelligence. And they just keep pushing that into basically everywhere. And we will see why they do that later, of course, the the whole business is based on it. But the thing with LLMs is that LLMs are not limited by the data it ingest they ingest, which is they can be limited by that, and that's what they want to make you believe, but by design, right? Because we have all these people that say, oh yeah, you know, GPT can only help you and it makes tons of mistakes, but it will get better, right? There is this fundamental belief that they will get better and better because this technology, it will improve and at some point developers would be out of work. Yeah, I think everybody's assuming that the models will exponentially get better, but we'll probably hit a plateau because of the design of how is it made. And what we've seen in the last couple of months with GPTs that there was very little innovation on how the LLM works and a lot on the UI. It can now consume text and images, but again,

do that because images are literally text um and then we have these video generators that do the same but the design itself of how it's been built, it didn't change at all. And what we are actually seeing behind all these marketing schemes, it's the same LLM with marginal improvement. is like again, putting more data. Um uh actually there's a mistake there, but it's like putting kerosine into a gasoline engine.

Uh better fuel, better data does not mean better performance. Right, you can put whatever fuel in the uh the the design of the engine will become the boundary condition, right, the limiting factor there. The second point is English is not a programming language. Right, it's a human language. So people claim English became now the the programming language of the world. Well to this humans can deal with English because they share a non linguistic understanding of the world.

Not because English is a great programming language. Exactly, and this is where I believe a lot of the tech influencers are wrong. A lot of people have come out and said, well, AI is already better than an mid-level engineer. Or you can use AI as a junior engineer, I create a 20 a lot like, hey, AI it's your little junior engineer. I'm like, no, it's not, it's the total opposite. Because the things that AI does really well or the things AI knows or the things you can retrieve with this language, large language model, which would be more correct language, right?

You can retrieve code samples and manipulate them, they are the total opposite. What the junior engineer can do. Junior engineer and mid-level engineers usually have a very good understanding of the world, of intention, of where the team, you know, what the team needs, and they sometimes have trouble understanding the lower-level APIs. uh the browser, how JavaScript really works. So I believe that both mid-level engineer engineers have huge complementary skills, but they are exactly on opposite when it comes to replacing each other.

Basically, as he says here, the worst junior dev, it's much better than the best AI tool. Junior devs can actually solve problems. And, you added this quote from uh Yann LeCun which says, LLM errors are not fixable without a major redesign. short and and the point of this first point it's, you know, don't freak out. These things they they might seem very intelligent, right? Because they have a lot of tweaks and things behind, but they are not. The only intelligence we know so far, it's human intelligence. And if you want to dive deep into this topic, this is one of the slides from Jan. Talking about how auto-regressive LLMs, the current technology, it's it's doomed because they are not controllable. And the probability that they produce a token that's outside of the set of correct answers is very high and it diverges exponentially and it's not fixable without a major redesign. The technology is not there despite what, you know, Mark Zuckerberg or Sam Alman might tell you. Exactly. Remember that LLMs are all about predicting the next token. That's how they were designed. And there's been a huge rebranding on hey, this is artificial intelligence, but it's really just predicting the next text token. So you can imagine if you ingest, and I think they did it over the whole Wikipedia, the Wall Street Journal, they ingested pretty much every text database you had. And they're very good at predicting the next token. That's it. So the whole AI discussion, it's not about AI. It's about auto regressive LLMs, which are, you know, an neural network that can predict the next token digesting text. to the second part, which is the meat of AI driven layers. And I chose this picture from the movie Margin Call, because in the end, who's pulling the strings up there? It's a bunch of MBAs, they might be great at financial things, but they're not that good at technology. So you see companies claiming that the tech layoffs that we've seen were driven by an increase of productivity due to AI, right? You've seen the CEO of Salesforce saying that, and actually the reality it's the opposite. They were not driven by any kind of improvement, technological improvement. They were driven by overhiring in the past, most companies, most tech companies overhired, they thought, hey, you know, COVID means everybody's going is going to be remote, they're spending more time on the platform, but then, you know, COVID went away and they hired all these people to have resources to deal with that. But when restrictions were removed, people went out to their normal life. Then they you had like huge product fakes and the metaverse, for example, in the

is of meta wants one of them. These companies they have to always sell the next big thing, that's the game, to raise money. A lot of times in the last years, those products they don't meet the expectations of investors. To do, you have to get rid of people. The third part is the lack of innovation, right, which is beyond acquisitions and monopolies. The tech sector like the traditional tech companies, including Alphabet and meta.

They don't have a lot to show, like if you look at their most successful products, they were all built maybe half a decade or a decade ago. 2025, this is just a perception game that NBAs and the people at the top are using to keep the programming class, which is developers, software engineers. from asking for more money, and I will show you why. Because the reality is the stock market, like we have the Nasdaq here, we have the S&P, these companies are doing very well.

They talk about layoffs, they talk about efficiency, but they are making more money than ever, check the stock price of Alphabet, check the stock price of meta and you will see they're doing pretty fine, they have the cash. But what's happening with all that that cash? And people talked a lot about inflation, right? But if you look at the the day data from 2023, right, corporate profits, they have more than half of inflation.

Right? What does that mean? In the last 40 years prior to the pandemic, profits, right, like the money that corporate pocket, were just 11% of the price growth. Right? And we saw um an increase in the price for consumers by 3.4% in the last year. But the input costs, like the cost for companies to generate those products has reason by just 1%. Which means that for many commodities and actually for many commodities and services, producer prices actually decreased.

So corporations have kept those money and they are not passing those savings to the consumers. And if you look at the corporate profits as a share of national income in the US, they skyrocket. by 30%, right, from the start of of the pandemic. The economy is back or it's a past, it's pre pandemic level, but workers share of the corporate income has still not yet recovered. Why? Because it's very easy as a company to raise your prices, you can do that tomorrow, right?

You can just update and raise your prices, but for employees to to also update their salaries. they have to go to an negotiation and they have to go to a review process. And software engineers are part of that. These people are bluffing. Uh these companies they have a lot of money and they just want you to believe that you are replaceable. They want you to believe that somehow the magical of AI will mean that you will replace you when actually, you know, they need you more than ever, they just don't want to pay you more and and that's the sad reality. On to the third point, why AI actually means more coders. This is a very well known uh phenomena in economy, called the Jevons paradox. Uh and why do we have the picture of uh of coal factories here? Is because people thought, you know, uh when coal will decrease price, it will not be used that much. But in fact, it just more and more companies started using coal. So why does AI means actually more coders? Well, an increase in developer productivity, it means that building software gets cheaper. So people will build more software. Right, more companies will build custom solutions or ramp up projects that were on the shelf. Like basically, a lot of companies were not migrating uh legacy code, a lot of companies were not building certain things that they knew they had to build because they were like, "Oh, yeah, but we need to hire more devs and it's going to take a lot and we don't have the budget." And again, even if, you know, coders will get faster, the main challenge for those tech companies that you saw, it's not how fast your developer can developers can code. Definitely, if they can code faster, that's awesome. But are you building the right thing? And most of them, as you'll see later, fail on exactly these things. Fail on building the right thing, not churning more code. And big tech, what big tech is hiding, it's product failures behind this AI improvements, as I mentioned before. And going back to the to to Mark, who said, "Hey, we saw an increase of productivity of of 30%. Engineering is doing very well. We don't need software engineers. But if we look a bit further and we look at how his company is actually doing, we see that Salesforce has a product problem, right? They need engineers. People don't like the product and they are saying, "Hey, we don't, we're not going to hire more engineers. We we will hire sales people. To do what? To push a expensive product that feels like technical death down the customers' throat?" No, what Salesforce actually needs, it's engineering talent to simply

fy the product itself, okay? The UI, it's so clanky and frustrating. I can even begin to express how non-intuitive it is. They have UI issues. Like this is a screenshot from you know, this this this was July 1st, 2024. They need a better designer, product testing team, or something. We tried it out, got stuck, we done for a year because of contact, and have seen it moved on to a different CRM. You know, Mike, you don't have an AI problem, you have an engineering problem, and you're going to solve that with more software.

And it's the same for Amazon, it's the same for Meta, it's the same for Google itself. These products are not, you know, delivering and the solution will not be AI, the solution will be as it was in the past. Good product and great coders and great engineers that will actually simplify those products and get them back to be competitive in the market. Well, they didn't worry too much because they have a Monopoly. That's for the for another video.

On to point number four, the GPT4 Hangover. Yeah, so literally when GPT4 came out, you already had the whole YouTube and internet saying, okay, it's over, AJ is here, and not only developers will be out of a job, but marketers, you know, product people, copyright, and so on. It's been eight months, right? And as far as I know, there are no none of the stuff we write the the senior dev it's using AI.

We tried to use AI several times and it would create, you know, this one step forward to step backwards kind of thing where you had to just start from scratch because again, um, many times what happens with this this large language models is that they'll just give you like faster horses, but you need to think about the car. You need to stop and you need to really think about the customer problem and they're pretty bad. They actually stole you down when you do that.

We've seen GPT already ingested most of the stuff it could. The only data that GPT 5 will have that GPT 4 didn't have is the conversation data it had with the 500 billion people that are using it every day. However, you got to understand that that data is just a subset inferred from the original set of data. So the quality is going to be very low. And so GPT 5 will most likely be trained on the same data because the humanity did not produce the evaluation of the previous day, because valuable data was produced by humanity getting better at specific problems, right? Not by just duplicating the same text and rewriting it with an LLM.

And so it's important to understand what's the semantic value? What is the value of an actual text? And if it's derived from another text, then it has almost a zero value. It's just saying the same, but in a different way. And so, it will probably have marginal improvements, but they will market it as AG I, and they'll probably use some UI tricks like trying to add voice or trying to make it great itself to make it look very smart.

Or you have the standardized programming benchmarks that they come up with like hey, you know, 03 is like 95% better than 02. But that's all made up, that's all in the lab. Those are standardized tests that AI can easily pass. It has nothing to do with the real world. And what we've seen is that if you look at the data, GPT usage kind of tank. So this is the data in terms of unique visitors to GPT, as we see it's kind of stable now.

And the average visit duration. Because that's actually what's telling us that people using it. Do do people get value out of it? And we see it's slightly decreasing. And if you use GPT a lot, I use it heavily for coding. Uh, Darrow also has been using it for anything that we do a senior dev that's not strictly coding. I use it inside Cursor, which is the code editor. And you'll see that it gets to a point where you just get tired of the back and forth and you start, you know, looking at the code and trying to improve it yourself.

It doesn't mean you're not going to use it as an auto complete, but it will get to a point where it makes you less productive. And I think people will get smarter and slowly you'll see a decrease in the average visit, that'll be like kind of less visits. And people will just get to the point where you're not really using it anymore for important things. And so going back to the garden hype cycle, which is how we analyze new technologies.

There's a lot, we are right now on this peak of inflated expectations. Where is there's a lot of over promising. It happened back in the days with Bitcoin where people were saying well it's going to replace in a year the normal currency. It didn't. Like the US dollar, it's still the main currency. The euro it's still there and yes, there's a lot. Bitcoin is growing and so on, but it's far away from what they predicted it to be.

And so, after the peak of inflated expectations, you have the trough of disillusionment. That's when investors that invested in Devin, AI, or all those kind of companies that came out will use money. And finally, people see where this does this technology actually fits in our productive cycle and then it gets added to our everyday life. And that's the plateau of productivity. We are right now still at the peak of inflated expectations, and I think the US just announced 500 billion investment.

That they will do with open AI, so things will still inflate, inflate, inflate and then we'll see what actually from there, what's actually useful. That will just add more gasoline to the fire. Let's see where that takes us. Whenever we have a huge deregulation in place, there was a recession like the big bubble coming afterwards. We'll see if that's that's the case now. And finally, the most important question for you, the people watching us is, what can you do as a software engineer to prepare for this?

There's a bunch of things you can do. Number one, use AI as much as you can. Add it to your workflow, play with it. Of course, use your brain in the end, do not, you know, do not put your fate or your git commit in fully 100% in the hands of cursor AI or Devin if you have access to that or chat GPT, okay? Use it because again, what you you fear what you don't understand. The more you understand it, the more you play with it, the more comfortable you will be and the more you will see it for what it is. The second point is think in first principles.

Look behind what people say, perception versus reality. And you've seen, right? And you've seen these tech CEOs talking so much about innovation in AI and how AI will solve all the problems, when in fact, if you look at the people using the product, you'll see they have an engineering problem, they need more engineers. Right? So we need in this in this world where everybody is saying this and that and, we're telling you all these things.

But I do invite you to start thinking by yourself and start looking at things in first principles and build your own opinion. Yeah, don't just follow the crowd because it's a very confused world and it's a very noisy world we're living. The third point it's, if you feel underpaid, most developers are, you've seen the inflation data. Companies have increased their prices. out any justification and they pocketed most of those money. They have money, even if they're trying to negotiate you on the phone, do not listen to that, right?

If your boss is telling you, hey, we cannot give you a promotion because times are tough, it's not true. They have the money. Okay, so go out there, interview and get a better paying job if you feel underpaid. Most developers are underpaid right now. And of course, for that you'll have to do some tech interviews and you will discover you do have some technical gaps. If you want to identify, to find out those gaps and find out what would be the fastest way, the most efficient way for you to fix them, then check out the free angle assessment that I put together in the comment.