On AI

Summary

In 2019, I started watching Andreas Kling's programming videos, many of which were live-coding sessions writing his new operating system SerenityOS. He was always pretty fast at writing code, especially since his IDE, Qt Creator, had knowledge of all his C++ code and could auto-complete classes, variable names, and function signatures. Back then it seemed like he had much of the code written in his head and was limited just by his typing speed. Four years later, he switched to CLion which had early GitHub Copilot support for AI-assisted code completion. While initially skeptical of Copilot, he experimented with it and within months, Copilot was writing code blocks, "reading [his] mind", and writing whole functions. His IDE's auto-completion through AI seemed like it had finally caught up with his brain and he could get ideas out into code faster. While I found it fascinating watching Andreas speed up his development, none of it appealed to me for my work for years after. I didn't even use an IDE or an LSP in my editor. I was skeptical of AI and probably wrote it off too soon due to its early failures. I eventually tinkered with Claude in a web browser, asking it questions that I would have otherwise asked a search engine and been directed to a Stack Overflow answer. Once Claude Code came out, I was amazed at the power of giving it access to a directory of code on my laptop and letting it directly read and modify files. Its ability to completely (and quickly) reverse engineer a binary blob with zero context is still pretty awesome. Once local models got good and small enough, I started using Ollama on my M4 Mac Mini to do local offline work. Like many things, it's possible to like and use a technology, language, or tool while rejecting all of the hype, fraud, and negative communities that build up around them (like Rails, or Rust). Of course, you can also choose not to like or use any of those things. You can prefer to use a hand saw instead of a circular saw, or to ...