1. AI Coding vs. Human Coding
AI does not truly code but instead predicts and guesses based on prompts, lacking the mental models and deep understanding that human developers develop.
Hear it at 0:00 →Engineering video
9:34 · 27 January 2026
AI Can’t Code: 7 Myths Debunked
Chapters 7 sections
Key concepts in this video:
2. The Fallacy of Mass Code Generation
Instant massive code generation is a weakness rather than a strength, as lines of code do not measure progress and debugging almost-right AI code consumes more productivity.
Hear it at 0:29 →3. Mathematical Limits and Plateauing of LLMs
AI performance faces mathematical limits rather than mere engineering hurdles, as evidenced by a 10x parameter increase yielding only a 1% accuracy increase in models like Qwen 2.5.
Hear it at 1:07 →4. Context and Tool Overload Failures
Adding context engineering, scaffolding, or more tools often increases failure rates, with longer conversations causing LLMs to get lost and driving up costs.
Hear it at 1:59 →5. The Myth of AI Prompting Skills
AI tools are extremely intuitive and lack deep skill requirements, but prompt experts with zero domain expertise cannot properly evaluate the quality of technical output.
Hear it at 2:53 →6. AI Reliance and Learning
The speaker discusses research on AI use, learning and recall. He uses it to argue that developers should keep thinking through problems and practicing their skills instead of simply accepting generated answers.
Hear it at 3:31 →7. Junior Hiring and Macroeconomics
The junior developer hiring crash was driven by macroeconomic shifts like the 2022 Federal Reserve rate hikes rather than the introduction of generative AI.
Hear it at 4:36 →8. Senior Engineers’ Concerns About AI
The speaker rejects the stereotype of senior engineers as unwilling to adapt. He discusses their use of AI alongside concerns about reliability, technical debt and added review work.
Hear it at 5:48 →9. Circular Financing and the AI Bubble
The AI boom relies on circular financing models, massive projected debts, and short GPU lifespans of one to three years, pointing toward a market bubble.
Hear it at 7:50 →Transcript
Meet number one, AI it's great at coding. False, AI it's not great at coding because AI doesn't really code, it guesses. When a human developer codes, they develop certain mental models, certain understanding. In exchange, what an LLM does, it just looks at whatever you send it in terms of a prompt or a context and it tries to come up with the best prediction. And that's the first thing we have to understand, the value of hiring a programmer, it's not only the output itself.
The next thing that we have to understand is that instant code generation at massive scale, it's a weakness. Not a strength, veteran programmers like Bill Gates, for example, understand that measuring programming progress by lines of code is like measuring aircraft building progress by weight. It's completely backwards. Because what happens with AI coding is, hey, you can create and ship a lot of code, but that productivity vanishes when you actually have to debug that stuff and when you have to review it.
66% of developers say that they are spending more time fixing almost right generated AI code. AI meat number two, AI is getting exponentially better and replacement it's inevitable. Technology doesn't always get better, just look at your calculator. But again, AI boosters will come and say, hey, AI performance, it's at this point in time, just an engineering problem. We had some negative surprises where meta surprise Llama 4 drop exposes the gap between AI ambition and reality, and it showed us that LLMs are reaching a certain plateau.
Here we have Queen 2.5 mat 7 billion instructed versus Queen 2.5 mat 72 billion. So we had 10x more parameters in the model and we've got a 1% accuracy increase. AI performance, it's not an engineering problem, it's a mathematical one, and you cannot solve that by simply adding more parameters. Mr. Sam Altman will tell us that, hey, we can add context engineering, we can add agent harness, we can add scaffolding, we can add MCP or whatever the hell is next.
But there's a problem with that. Long context actually increases failure rate. And we have several studies about this which show that less is actually more, that actually reducing the numbers of tools available to LLMs significantly improves their function calling. Or that, for example, LLMs get lost in multi-turn conversation and that the smaller your context, the smaller your conversation, the easier it is for LLM to come up with a good solution.
Even people who are building AI agents are betting against them, because they've seen firsthand that conversations, the longer the conversation gets, the more you pay for the AI agent and the worst the results, an over-engineered unstable system, it's still an unstable system. AI meat number three, it's a skill issue. You know, you just don't know how to use it, let's gaslight people instead of accepting the tool is broken.
When in fact, the truth is, there are no such things as AI skills. AI tools are incredibly intuitive and simple to use. Most AI skills can be learned in a few minutes and an expert at AI prompting with zero domain expertise won't be able to check if the quality of the AI output is actually good. Whether be we're talking about coding or cooking or medicine. Which is why people say, hey, the best developers working with AI are the ones that are already senior or experts at what they are doing.
AI meat number four, AI won't replace you. But another developer using AI will. There was an experiment, and there was this university teacher in the MIT, and he taught a fourth hand class, and he made an experiment. He said, hey, let's split the class into a groups. One group will use chat GPT to solve the questions. The other group will be able to use the the Lama Yama model from from Meta, and then the others can only use Google.
Which group will be able to solve the problem faster? Well, as you probably guessed, Chat GPT users were the fastest, Yama was second place, and the slowest were the ones using Google. But now comes the interesting part, which group do you think passed the final exam about the task? Surprise, Chat GPT users passed at 0%. Everybody failed the real exam. Yama or or Llama users passed at 50%. And of course, the students using Google passed at 100%.
Developers who are using AI excessively are actually slowly making themselves obsolete. AI myth number five, AI it's already replacing junior and mid-level engineers. This myth started with this amazing study from Howard University, which claims that generative AI as seniority bias technological change. Then, of course, the media jumps on it and is like, hey, top software engineering students can't get a job because of AI.
And everyone is so panicked, entry-level tech workers describe the AI fueled job apocalypse. Actually, the junior hiring crash started in 2022 before many people knew what prompt engineering was, a few months after the Fed jabb the rates from 0% to 5.25%. And as you can see in this graph, I mean there was a chat GPT launch in Q1 2023. But the preference for hiring senior engineers started way before GPT has launched.
This is a known phenomena and it happened ever since 1980 when the economy slows down, junior positions are first to take a hit. Nothing new, it's not because of AI, it has to do with macroeconomics. A tip for you, if you are a recent graduate, look for jobs around software dev, you know, maybe Salesforce admin, maybe Shopify. Find shelter until the good times are back. AI myth number six, senior engineers oppose AI because they are too attached to the old way of doing things.
Now, we're just a bunch of dinosaurs and we are screaming against the asteroid. adapt or die and it feels like senior dev side are unwilling to adapt because they love typing code by hand. It feels like the industry really picked on them, you know? The AI shift is real, but senior engineers are slow to adapt. Senior engineers struggle to build AI agents because they are used to thinking in a deterministic way, and AI agents and the new way of coding, it's probabilistic.
Senior devs are the ones that are using AI the most, they are shipping more code, and they are the ones that also are the least afraid, they are more confident in their expertise, right? So, this idea of senior devs being some kind of dinosaur that doesn't want to adapt, it's fabricated, they know that AI generates technical depth. Because imagine being in a team with three juniors who are just pushing hundreds of lines of code that they they claim to have reviewed, but they don't review.
And you are the one that has to stay late at night and review all that stuff because if the application stops working, it's on you. Seniors also know that AI creates too much work, again, debugging, code reviewing, etc. They know that context context engineering sometimes takes more work than just fixing the problem. Again, when you have a good knowledge of the codebase, when you have a good understanding of the the technologies and when you when you are a senior, you know, it's a lot easier to just think a bit about the problem, go there and fix it, then, you know, querying the freaking chat bot or or or whatever, the agent and and waiting for it.
And then maybe you waited 5 minutes for nothing and I've seen this a lot. In all the demos that that that AI influencers are posting is like, oh, yeah, let's wait 20 minutes, oh, it kind of made a mistake and, you know, I I kind of fixed it, but for the sake of this video, let's suppose it it's getting it right. What what what the heck? You just wasted 20 minutes, you could have fixed the problem yourself in in probably three to four minutes.
Four, seniors know that AI is creating a generation of developers who can code. And fifth, the technology cannot be relied on and makes you less productive. That's why senior devs are against AI. AI mean number seven, and we're going to finish with this one, AI is not the bubble, because AI companies have revenue and they are building infrastructure for the future. There is some revenue in AI, but what you find a lot, it's what we call the infinite money glitch, you know, like like we have this circular financing.
Open AI buys chips from Nvidia and Nvidia invests in Open AI. Open AI partners are carrying 96 billions in debt. Actually, it's estimated that Open AI needs to waste about $500 billion in the next four years to stay alive and serve its commitments. That's more than the projected available supply of US venture capital at the end of 2025 and the combined available capital in top in the top 10 private equity funds. But again, AI boosters will say, hey, we are building infrastructure, right?
It's like the long-term investment into the future, you know, just like the internet. Well, it turns out that the average useful lifespan of an Nvidia chip GPU in a hyper scale data center, it's generally estimated to be 1 to 3 years and that's why there is no long-term infrastructure investment. This is a bubble and my advice for you as a software engineer is to avoid working at companies with the huge exposure to AI. You know, avoid AI startups.
They're going to be the first one that are going to take the cut, even if you understand that this narrative is false. You got to play the music, right? So if this is the music, if this is the song they're singing, just sing along. And if your company and your CTO, CEO, whatever the heck is telling you, hey, use AI, use AI, just say, yeah, I'm using AI and that's it and just go go and fix the problem.
However you you see fit. With that being said, if you are a JavaScript engineer and you like to get to senior and beyond, become the best software engineer you can be.