Eighty percent. Then fifty. Then zero.
That's how Andrej Karpathy - former head of AI at Tesla, OpenAI founding member, and one of the most respected names in machine learning - described his coding trajectory in a recent podcast. He went from writing most of his own code to writing none of it, living in what he calls a state of "perpetual AI psychosis" where the possibilities feel so infinite that he tries everything.
Karpathy isn't some middle manager who just discovered ChatGPT. He literally helped build these systems. And if the person who understands transformer architecture better than almost anyone alive has completely handed over his coding to AI, that tells you something about where the tools actually are right now.
The "Opportunistic" Workflow Problem
The pattern Karpathy describes is one that a lot of developers and power users will recognize. Once you realize AI can handle most implementation work, your brain shifts from "how do I build this" to "what should I build." The bottleneck moves from execution to judgment.
That sounds like pure upside, but there's a real cost. Several developers have described the same phenomenon: you start five projects instead of finishing one. You prototype constantly but ship less. The friction that used to force you to commit to one approach is gone, and without it, you bounce between ideas like a pinball.
This is especially relevant for people using tools like Cursor, Claude Code, or GitHub Copilot as their primary development environment. The tools are good enough that the old discipline of "think carefully, then write code" gets replaced by "generate it, see if it works, regenerate if it doesn't." That's faster for known problems. It can be paralyzing for open-ended ones.
What Actually Helps
The developers who seem to be navigating this well share a few habits. They treat AI-generated code with the same scrutiny they'd give a junior developer's pull request - useful output that still needs review. They set hard boundaries on exploration time before committing to a direction. And they keep writing some code by hand, not out of nostalgia, but because the act of writing forces you to understand the problem at a level that prompting doesn't.
Karpathy's honesty about this shift is valuable precisely because he's not selling anything. He's describing a genuine psychological change in how skilled practitioners relate to their craft. The tools didn't just get faster. They changed what it feels like to work.
For anyone deep into AI-assisted development, the question isn't whether you'll experience some version of this. It's whether you'll recognize it when you do.