Stupified by AI

Rarely used skills deteriorate. I used to be great at arithmetic in high school, but today I double check everything with calculator. But this is why we force kids in schools to not use calculators: so they learn and understand math.

I can tell when AI tries to bullshit me (and it tries a lot). My reviews take 3-4x longer than last year and I find 6-8x more bugs in “80% AI” code reviews. Which is frankly all the code reviews I do these days. But I only see bullshit in code that I have deep understanding of. I’m blind when it comes to fields outside of my expertise.

I ask LLMs to do the first review and I carefully read all of their findings. Sometimes they’re right, sometimes they’re wrong. If they find a bug, I describe it with my own words to the author. I never copy-paste, because the slop produced by LLM is incomprehensible. At the same time I do independent, 100% human review. For complex changes I use AI reviews as an anchor and start in the areas identified by them as problemmatic. This approach sits nowhere near the industry expectation that AI will make me 10x more efficient. There’s no way to speed up if I add AI on top of me and my human brain instead of replacing it. Unfortunately, we count efficiency in hours, not in bugs found.

This makes me anxious, because I’m becoming a bottleneck: one easily replaceable with AI. AI bots reviewing AI code 24/7 must be a wet dream of “our industry”.

I learned my skills in the battlefield: by spending hundreds of hours analyzing requirements, architecture and, especially, full of hacks implementations. I learned by reading lots of code, both good and bad (and by identifying which is which), by making mistakes and by creating beautiful future-proof programs (and by identifying which is which). I learned by talking to my seniors and specification engineers, trying to understand why something’s done some way and not another. By thinking what automated tests we need instead of asking AI to write tests that confirm current implementation.

AI takes this away from the newcomers. Just as you can’t learn math by only using calculator, you can’t get deep knowledge by asking AI to architect and write a solution for you, and without deep knowledge you can’t judge the quality of AI-generated code. Vicious circle. What’s the future when old farts like me go away? When we either die, burn out, are replaced by AI, or when our skills deteriorate? Where will we be in 2/3/5/10 years if we don’t train deep code understanding in the next generation of programmers? Who’s going to call the AI bullshit?

Vibe-coding is a short-term strategy focused on immediate profit, not long-term sustainability. AI can be a great tool, but it can’t be the substitute for knowledge gap of its operator.

Or maybe I’m delusional and my work means nothing.