Source: https://geohot.github.io/blog/jekyll/update/2025/09/12/ai-coding.html
I spent my Sunday morning revisiting a September 2025 blog post from George Hotz on AI coding, and his framing is brutally clarifying. Hotz argues that the best mental model for a programming AI is simply a compiler. You feed it a prompt, which acts as your source code, and it outputs a compiled version. The problem is that the input language is English. It is imprecise, highly non-deterministic, and makes for a terrible programming language. What stood out to me was his critique of the industry's reaction. We treat natural language prompting like a miracle, but if a traditional compiler behaved with the same unpredictability as an LLM, we would immediately file a bug report. Hotz suggests that a company replacing developers with LLMs is really just demonstrating how bad that company's codebase and hiring bar is. Instead of chasing what he dismisses as vibe coding crap, he thinks we should be doing the hard work of building better foundational tools. He points to a study claiming AI makes developers feel 20 percent more productive while actually making them 19 percent slower. It is a sobering reminder that we are not eliminating programming, we are just settling for a fundamentally worse language to do it.