LLMs are a weird thing. You can love and hate them at the same time. I’ve come to realize that LLMs can indeed be valuable tools, but they also disrupt programming as a craft in ways that aren’t immediately obvious. The real insight isn’t whether LLMs are good or bad. It’s that their value is highly circumstantial and depends entirely on the expertise of the person wielding them.

The Vibe Coding Illusion #

LinkedIn is full of AI advocates and project managers celebrating their vibe-coding successes. One is almost inclined to believe their stories. But very often these people are judging a book by its cover without knowing what is written inside. If software does what you want and you created it via vibe-coding, what’s the complaint?

The complaint is this: software that works today is not the same as software that will work tomorrow. LLMs can easily get you 80% of the way there, similar to the Pareto principle. The code works for a while, until it doesn’t [4]. The remaining 20% contains the architectural decisions, edge cases, and hidden assumptions that determine whether your codebase survives contact with reality.

My Journey from Skeptic to Power User #

For a while I stood on the sidelines. Back in 2022, it was clear that these models were not yet good enough. But model quality progressed rapidly, and by 2024/2025, frontier models had reached a level of capability I could no longer ignore.

I was struck by fear. Would the career I had relentlessly built, a path full of tough learnings across a broad skill set, soon be invalidated? But I was also annoyed by people praising AI without understanding how the technology works or its fundamental limitations. So I decided to dive into the unknown.

Today I consider myself a power user. I subscribe to Claude Code, Mistral, and ChatGPT, using them extensively for brainstorming and coding at work and in my private life. What follows are the lessons from that journey.

The Hidden Costs of AI-Generated Code #

From the start, it was tempting to outsource boring tasks. Create a method here, generate some tests there. It gives you a feeling of progress. And some of the code was actually good, especially when starting new projects from scratch, where LLMs excel.

But other code silently introduced hidden debt. On existing codebases, models would sometimes ignore clear guardrails (even explicit instructions in CLAUDE.md) and produce mediocre solutions that became real problems later in the product lifecycle. And sometimes LLMs completely failed, even on tasks where I didn’t expect it, like Azure pipelines with Docker-in-Docker configurations. These failures cost me significantly more time than working without AI, because the models kept producing wrong solutions with compelling confidence.

This is the fundamental nature of the technology: LLMs are best described as “whatever machines” [1][2]. They generate plausible output regardless of correctness. Knowing when something is off is the key skill. Intriguingly, research suggests that senior developers can actually be slowed down by AI, even while their perception of pace is the opposite [3].

Skills Still Matter (Perhaps More Than Ever) #

This is when I realized how valuable my skills still are. In bounded contexts, frontier models are extremely helpful. But they can easily introduce flaws that a non-programmer would never identify as critical. These issues accumulate under the surface and will ultimately threaten a maintainable codebase, leading to data leaks, performance issues, malfunctions, or at worst, business failure.

The knowledge and experience of a programmer is a strong asset here. In the right hands, these models will make you undeniably faster. Even semi-autonomous development with multi-agent orchestration might be sustainable with proper planning and supervision, though evidence suggests caution is warranted [5][6]. It is amazing how fast we can build things with LLMs. Maintaining, scaling, and expanding those things is another matter entirely. For that, you need to know what you are doing in the first place.

Beyond the Hype Cycle #

From how model performance is progressing, we should not expect major leaps on the LLM frontier in the near future. The current approach is plateauing, and it does not lead to AGI. But if applied wisely, it remains a revolutionary tool regardless.

The topic has darker dimensions too: the transhumanist movement in Silicon Valley, the significant climate impact, massive copyright infringement, and a potential economic bubble. Is the push toward AGI really something that benefits humanity? These questions deserve their own treatment.

What’s certain is that the technology will stay. The question is whether we use it as a crutch that diminishes our abilities, or as a lever that amplifies them (to some extent).

References #

[1] https://eev.ee/blog/2025/07/03/the-rise-of-whatever/

[2] https://tomrenner.com/posts/400-year-confidence-trick/

[3] https://arxiv.org/abs/2507.09089

[4] https://spectrum.ieee.org/amp/ai-coding-degrades-2674835455

[5] https://matthiasnehlsen.substack.com/p/13671-lines-deleted

[6] https://embedding-shapes.github.io/cursor-implied-success-without-evidence/