I'm also in software engineering. I spent some time thinking like you, then some time trying to find a middle ground between using AI like that and doing everything myself.

Now I do it all without AI and am slowly purging all of the rituals around LLM usage from my brain. I'm trying to reclaim that space for useful knowledge in the languages I like to work with.

The reasons are numerous, some ethical, some practical, and some personal.

Ethical:

• There's no way to use frontier LLMs without contributing to climate change. The products of Anthropic and OpenAI are built on raiding the planet's resources.
• They're also built on plagiarism on an unprecedented scale. Their products wouldn't exist if it weren't for stealing basically all copyrighted works ever created by people, and would be even less economically viable than they already are if they were forced to pay for the thievery.
• They're addictive to use. People become dependent on them to the point of losing the ability to make decisions on their own. The non-deterministic nature of their output means people react psychologically to them the way gambling addicts react to roulette wheels, always hoping the next prompt will be the one that hits gold.
• Small, local models help, and the ones that are both open source and open weight are far superior to the rest in terms of ethics, so for anyone addicted to LLMs, I'll recommend them as a form of methadone. But their psychological impact on their users can be just as damaging.
• The business practices of these companies is driving the largest economic bubble in history. When (not if) the bubble collapses, it will hurt millions of people. They don't care, because in the short term, number go up.

Practical

• Model-produced code isn't as good as anyone who's creating it thinks it is. I've lost count of the number of projects I've seen from other developers (who think they've mastered the art of AI wrangling) that contain rudimentary errors - including gaping security flaws. I found the problems just by being the first person in a while to actually look at the code. No, not diffs, because when you're churning out five thousand-line PRs, if the PR "looks right," you'll usually just sign off on it.
• The Supreme Court has already ruled that the output of LLMs can't be copyrighted. So from a legal perspective, you're risking poisoning your own projects with the output of coding agents. If someone steals that output and markets it as their own, what are you going to do? It was already derived from stolen work itself, so it's not like it's yours.
• Becoming dependent on businesses that have no path to profitability is a terrible idea. If you have built your workflow around Claude Code and Anthropic folds, what will you do? These companies cannot offer their services in a way that brings them sustainable profit. When the venture capital runs out, the clock starts ticking on their actual viability.

Personal

• Both of my kids are artists. I want to be able to look them in the eye and tell them I'm not using tools that are designed to make creativity obsolete and replace it with statistically generated pixels.
• I'm also a writer. I don't want to become dependent on tools that deprived my fellow writers of their copyrights so that slop manufacturers can try to replace them with statistically generated words.
• I hate the way I feel when I'm using an LLM. The brief highs of getting the output I wanted (more or less) is offset by the impact on the way my brain works. I find myself thinking less and asking the LLM for its "opinions" more. I stop trusting my own judgment. And I start feeling addicted.

On that last point: I knew I had to stop entirely - not just limit my usage or restrict myself to local models - when I realized part of my brain really didn't want me to. I had to have an argument with myself in order to do anything by hand, and I like coding by hand. I enjoy the puzzle-solving. AI was replacing my main source of enjoyment in my job.

And it wasn't even doing that part well, it was just doing it faster. I am a better coder than any LLM, and that's not a brag, because I know there are tons of developers out there better than me. Faster doesn't equal better.

So that's where I'm at today.

I expect to be downvoted.

I mean, you're not wrong. You're only saying facts. I'm a strong AI opponent, but these things you're saying are simply true. 🤷‍♂️ That said, I like coding and doing that bit myself, so I'll keep coding. 😅

My concern is that the current models have been trained on current technology by stealing knowledge from the existing pool of senior and junior engineers as well as a whole load of hobbyists.

If everyone moves up to senior now using LLMs, that's all well and good now, but then there'll be no juniors to steal the grunt work from to train AI and no-one to move up to senior in a decade or two.

Do you really think that the hobbyists are going to be enough to steal from to perpetuate the system in the near future?

I like the chainsaw analogy but I think you need to apply it more like, you're trying to carve something out of a tree stump and you only have a chainsaw nothing else. So you can make it look good. But the fine little detail work isn't as possible. You have you go back in with other tools to do fine detail work.

And like a chainsaw, without practice, you have no idea what you're doing at first. If you vibe code something without coding experience, a lot can go wrong but if you don't understand how it works, you can't fix it.