Will AI Replace Software Developers? · Oleg Dubovoi's Blog<br>A Markdown version of this page is available at /publications/will-ai-replace-software-developers.md.
← All publications Lately, the question “Will AI replace us?” has worried many people. We can see how LLMs handle programming tasks very well and write code at a middle to senior level. This makes many software developers concerned about their future.
Introduction
To be honest, I rewrote this article several times and spent more time on it than usual. I didn’t want to take the side of people who are against AI, that’s not how I see it. I’ve been using LLMs in my daily work for several years, and it’s hard to imagine working without them. Not because I wouldn’t be able to code or solve complex problems, but because my efficiency would definitely be lower.
AI is evolving faster than most developers can adapt, and we’re seeing major changes in the IT industry. Because of that, many people feel stress, denial, or even hostility toward AI. But most of these feelings are driven not by real threats, but by hype and strong marketing from large AI providers.
The goal of this article is not to show that AI is weak or useless, or that we shouldn’t use it. Not at all. I want to highlight the other side, the one that people don’t talk about enough. LLMs are powerful tools, but they come with limitations and require skilled professionals who understand what they are doing.
Artificial Intelligence in Software Development
Modern LLMs have truly become powerful tools for software development. Claude Code or Codex can write high-quality, well-structured, and quite complex code. It can work with large codebases and understand the project context.
To understand whether AI can replace software engineers in the future, let’s first look at the main question: does an LLM really understand why this code is needed?
As you know, an LLM works by predicting the most likely continuation of a sequence of tokens based on a huge amount of training data. In simple words, modern AI does not “think” and does not “understand” the goal of the system. It statistically decides what is most logical to write next.
That is why LLMs show excellent results in typical and well-defined tasks:
CRUD applications, standard REST APIs, simple SPAs built with Angular or React, and template-based business logic. All of this appeared many times in the training data, so the model can confidently reproduce familiar patterns.
Problems begin when deep understanding of the domain and execution context is required. For example, when designing a distributed system with complex requirements for fault tolerance, data consistency, and business constraints. In such tasks, AI may generate code that looks “clean” and correct, but:
does not consider real load scenarios,
breaks important business logic rules,
or suggests architectural solutions that cannot work in the given environment.
The more complex the system, the wider the context, and the less formal the request, the higher the chance that the model will get confused, hallucinate, or move toward wrong solutions.
Why Scaling LLMs Is Not Enough
One of the biggest challenges in building more powerful LLMs is the quality of the data they are trained on. Even if we keep scaling models, issues like model collapse can limit progress. When models are trained on data that already contains AI-generated or low-quality content, they can start amplifying errors, repeating mistakes, or learning unrealistic patterns. Simply making models bigger won’t solve the underlying problem, the foundation itself needs to be clean and reliable.
Yann LeCun, a Turing Award winner and one of the founders of modern AI, and former Chief AI Scientist at Meta, believes that simply increasing the size and power of LLMs will not help. According to him, this is not the path to real artificial general intelligence (AGI).
He argues that real intelligence needs a model of the real world, including physics, cause and effect, and goals. Language alone is not enough:
“We need systems that understand the physical world, not just systems that generate plausible text.”
Programming requires planning, reasoning, and understanding long-term consequences. LLMs can help write code, but they do not truly design systems or understand why solutions work . That is why, no matter how powerful new models become, the same fundamental problem remains.
At the same time, Yann LeCun is working on a new AI architecture called VL-JEPA (Vision-Language Joint Embedding Predictive Architecture). This is not a classic generative approach like GPT models. Instead of predicting text token by token, the model works at the level of semantic representations. It does not generate answers word by word. It predicts a semantic representation of the answer, a kind of “meaning fingerprint.” If needed, this representation can later be decoded into text.
VL-JEPA may be more efficient than traditional multimodal...