On coding with LLMs

18 Jan 2026

It looks like we are near the peak of the AI mania. The usefulness and the limitations of this technology are clear now. What we see now are attempts to overstretch the idea and apply it where it will never work, much like with blockchain and cryptocurrencies several years ago.

The ubiquitous AI startups will fail and we will see a decline in stock indexes, as well as a re-evaluation of AI strategies in many companies. I'm not excited about new AI agents, nor scared of becoming irrelevant; I'm bored now. Users are sick of tasteless drawings, trying-to-be-helpful AI summaries, and chat windows popping up from every toaster.

Please don't get me wrong: there are a lot of areas where LLMs are truly useful. I'm glad that I don't have to write tedious unit tests fully by hand anymore. It's great that you can quickly toss together a throw-away script, translate a function from one programming language to another, or generate something as a starting point when coding in a language that you barely know. The problems arise when you perceive this "something" as a finished product.

An old book (Code Complete by Steve McConnell) says that writing and reviewing code takes around 30% of an average developer's time. The rest is spent in discussions, replying to emails, reading documentation, troubleshooting, learning, or even personal stuff. The data come from a 1964 research and the proportion certainly depends on the person and the company, but any developer can confirm that they are not working with source code 100% of their time. If we assume the 30% estimate is correct and you can speed up coding by 20%, you can only expect a 5% reduction in the project duration (see Amdahl's law).

So the promises of creating a complete product in a weekend are unrealistic. Programming is hard. I spend hours debugging UI issues or checking the assembler output and this is my own project coded by hand, so I understand it quite well. If I were vibecoding, it would take more time to find the root cause.

In my career, I never worked on a project where I would need to generate a lot of mediocre code quickly with an LLM. The usual story goes like this: two Californian guys, working after their day jobs, create a product that suddenly becomes popular. The code is a mess; they hit scalability limits, so they hire a Central European like me because a developer from the Bay Area with the same skills would cost them more; at the same, the salary is higher than what I could earn in my region. So we both tolerate the cultural differences, wild accents, and crazy time zones.

I gradually debug the issues, make their code better organized and faster. For one company, I managed to make it six times faster. This does not mean that the founders were bad developers; they just cut some corners when creating the first version. But at least, the code was written by a human.

Many new founders today vibecode, so I think that I will be able to do such consulting work until I die or lose interest. We will see a lot of weird class hierarchies, logical mistakes patched with special case handling, and buggy snippets that the LLM ripped off from GitHub or StackOverflow.

Less young people will be able to solve technical problems because they will rely on AI when studying. And if you got used to asking ChatGPT for solutions to simple problems, you will never learn how to troubleshoot complex or novel issues that AI cannot help you with. Cognitive skills require training and practice just like physical ones.

Besides that, prompting is a skill on its own. Sometimes it's easier to write code than to explain verbally in every detail how it should work, especially if English is not your first language.

I would say that vibecoding only adds another layer of abstraction above high-level programming languages. If your project contains a lot of repetitive AI-generated code, it's a sign that you may benefit from designing a library, a framework, or a DSL (a domain-specific language). Metaprogramming techniques can help in some cases, too.

As Noam Chomsky noted almost three years ago, LLMs are not a replacement for human mind; they are useful, but cannot think in the way that we do. The mental health impact of AI is under active study now. IMHO, chatting with a video card chip is not something natural for a human being and it certainly creates some information overload. Overall, LLMs don't make software engineering much simpler; they even add complexity when used incorrectly.

I may add some AI features to Aba Search and Replace in future, but it will never become a privacy-invasive AI agent and your files will always stay on your computer. When developing the app, I only use a locally-running LLM for proofreading texts. No vibecoding, no AI slop. As an experiment, this blog post was written without any help from AI.

Stop jumping between browser tabs and random online tools. Aba Search and Replace is your Swiss army knife for fast, safe text updates across multiple files and data conversions, with all your data staying on your computer. Built for developers, testers, and analysts.