January 28, 2025
Hot take: LLM-powered programming tools are going to drive down average code quality for a while, before we turn the corner and code quality goes up.
For the last few months, I've been watching developers at hackathons, and reading code samples people post to the @pipecat_ai discord when they need help, and paying attention to the workflows of people making YouTube tutorials ...
Programming is changing. New standard workflow:
1. Type a prompt into <LLM front end>.
2. Watch some code get generated.
3. Copy-paste or click <accept>.
4. Run the code.
5. See if it maybe worked.
6. Repeat.
Notice that there's no "read the code" step. Of course, sometimes people read the generated code before running it. But often, not so much.
And it's easy to see why, when you start programming like this. Fast iteration is such a super-power. And LLM-powered programming is a new kind of fast iteration.
But the generated code is often pretty bad, even when it mostly works.
Which is ... fine, I think? It's early days, the models are only going to get better, we're starting to build some pretty good tooling to support this workflow, etc.
We have to use these new tools in order to build new tools to support the new tools.