Companies need to be aware of the long-term affects of relying on AI. It causes atrophy and, when it introduces a bug, it takes more time to understand and fix than if you had written it yourself.
I just spent a week fixing a concurrency bug in generated code. Yes, there were tests; I uncovered the bug when I realized the test was incorrect...
My strong advice, is to digest every line of generated code; don't let it run ahead of you.
It is absolutely terrifying to watch tools like Cursor generate so much code. Maybe not a great analogy, but it feels like driving with Tesla FSD in New Delhi in the middle of rush hour. If you let it run ahead of you, the amount of code to review will be overwhelming. I've also encountered situations where it is unable to pass tests for code it wrote.
Like TikTok AI Coding breaks human psychology. It is engrained in us that if we have a tool that looks right enough and highly productive we will over-apply it to our work. Even diligent programmers will be lured to accepting giant commits without diligent review and they will pay for it.
Of course yeeting bad code into production with a poor review process is already a thing. But this will scale that bad code as now you have developers who will have grown up on it.
When have companies ever cared about the long-term effects of anything, and why would they suddenly start now?
> It causes atrophy and, when it introduces a bug, it takes more time to understand and fix than if you had written it yourself.
I think this is the biggest risk. You sometimes get stuck in a cycle in which you hope the AI can fix its own mistake, because you don’t want to expend the effort to understand what it wrote.
It’s pure laziness that occurs only because you didn’t write the code yourself in the first place.
At the same time, I find myself incredibly bored when typing out boilerplate code these days. It was one thing with Copilot, but tools like Cursor completely obviate the need.
AI can get you to "something that runs" frighteningly fast, but understanding why it works (or doesn't) is where the real time cost creeps in
100% agree with you, my sentiment is the same. Some time ago I considered making the LLM create tests for me, but decided against it. If I don't understand what needs to be tested, how can I write the code that satisfies this test?
We humans have way more context and intuition to rely on to implement business requirements in software than a machine does.