The label AI has always tracked hype, not capability. The research field has been an active discipline since the Dartmouth workshop in 1956 — McCarthy, Minsky, Shannon. The target was general intelligence — and hasn't moved.
The label is industry shorthand. Nobody who builds game AI thinks they're building intelligence. They're shipping a system that looks alive for the few minutes the player is paying attention to it. The same applies to LLMs.
LLMs are statistical autocomplete
LLMs are built on top of neural nets. The job is simple: given a chunk of text, guess the next token. Train one long enough — on most of the internet, on every book in the library, on a few decades of code — and the guesses start to look smart. None of it is intelligence — it's a very thorough study of what token usually follows what.
The model doesn't know. It predicts. No experience, no emotion, no knowledge, no critical thinking, no wisdom — none of the human factors that produce understanding. There's a model of what comes next, one token at a time.
What an LLM can produce, once harnessed, is absolutely remarkable.
Vibes all the way down — what a freaking nonsense
The term vibe coding makes me puke 🤮 "What you vibing?" — one may ask. Just button it... Andrej Karpathy came up with the term in early 2025. Everyone has been parroting it since, without thinking.
I don't want to have anything to do with vibe coders. Vibe coding is letting an LLM run loose, building black boxes with the attitude of "ahh whatever, if it breaks I ask clanker to fix it." That's asking for trouble. It may build you a simple thing — but a demo has nothing to do with a production-ready system.
The outcome is a tangled mess. Spaghetti code looks organised next to it. No structure. No architecture. Everything dumped in one file. Unnecessary code everywhere. Security holes bigger than the ones in Swiss cheese. Performance issues. Just a random mess assembled from random snippets that almost works.
AI assisted development
This is the term I like to use. I've been an incredibly heavy LLM user since day one — I hate it as much as I love it. My favourite joke about all this: if AI ever gets sentient, I'm at the top of the hit list. The amount of abuse I offload onto the models to find the right workflow has made sure of it.
There's no golden bullet. Everyone is different. Everyone works differently. Everyone has to find their own way to make it work. It works for me — and believe me, I put that thing through its paces daily.
One thing is undeniable: LLMs do better in some areas than others. Code generation tracks the training data — the more there is for a language, the better the output. For popular ones, decent. For niche ones, worse. For Odin, the models absolutely suck 🤷
When pointing LLM at the codebase, the quality of that codebase matters a lot. Garbage in, garbage out. The model imitates what it reads. The same applies when generating from nothing — the training data quality is questionable, the output mediocre at best. This is where moderation comes in.
How I incorporate LLMs into my daily engineering
AI is just another automation tool. Nothing else. All it does is save me time — sometimes a lot of it. Most of programming work is boring anyway: simple, repetitive code. In the grand scheme of things, the internet is basically CRUD. Generating code speeds my workflow — I don't have to type it.
The difference is I'm in full control. I own the architecture. I decide the code organisation. I follow every line. I push back when needed. I ask for changes. I give directions. I ask for specific implementations. I verify what the model claims. I refuse output I don't understand. I iterate. I adjust manually. I refactor what doesn't fit. Small iterations work. "Build me a Facebook" doesn't 😂 The model is the tool. I'm still the engineer. If I'm not going to end up with code I'd write by hand, it has no place to exist.
The code that ships under my name is mine — accelerated.
I use the LLM to research things I don't know — cross-referencing sources, parsing specs, drafting documentation. Faster than the manual route.
LLMs are invaluable when I'm learning new concepts and new languages. The rate at which I can pick up new ground is intoxicating — weeks of reading compressed into hours. Moderation, again. Never trust an LLM. Never. If it tells me something I don't know, I cross-reference. They hallucinate — that's how they work.
And of course, writing — which LLMs excel at. But there's a big but...
Ghost in the prose
I'm sure it doesn't surprise you: this very article was written with an LLM's help. The distinction matters — with help, not generated. There is no point hiding it. I'm allergic to slop. The kind of AI writing that delves into every paragraph, opens every post with in an era of. I won't ship it.
Yeah, I could prompt it. Give it a title. Ask it to fill in some generic SEO-friendly text. No voice. No personality. Nothing interesting. Just slop. The internet is already drowning in it. I won't be adding to it.
Working on a RAG system recently, I discovered something fun: AI detectors are worthless. I fed one the full text of The Last Wish by Andrzej Sapkowski — published in 1993. The verdict: AI-generated, 60% confidence. (Yes, kids — The Witcher is from the 80s.)
Here's how I write. Hours of thinking. Collecting notes and anecdotes. Research. A draft written by hand. Then the LLM comes in — proofreading, grammar, awkward phrasing. More reworking. Shuffling lines around. Reading the thing a hundred times to make sure the words flow, the rhythm holds, the whole thing is coherent. Writing is a long, hard process of constant refactoring. The same applies to code. Neither will ever be perfect — it just has to be good enough. Make your own judgment based on that.
The model isn't in git blame
Unless it's vibe coding, then it is. There is absolutely nothing wrong with AI-generated code. None of it bothers me. What bothers me is companies bragging about the percentage of AI-generated code in their codebase. 80% of our code is AI-generated. It doesn't matter. It's all hype-driven. It's not an achievement. Code is a means to an end — shipping a product. The end user doesn't care if the app was written in Rust, Ruby, or PHP. They don't care if it was written by AI. They want to get their stuff done. They want to use the product. If they don't, you're in trouble — but the code isn't the problem, if it's good enough. The product is.
Shipping a reliable product is a choice. A decision. Engineering is a craft — care, knowledge, judgment, experience. How you get there doesn't matter. A Sunday driver and an F1 driver both have a car. The car needs a driver.
What matters is the end result: software that's reliable, performant and secure.