The Poor Man's Claude Code

A dispatch from the pathologically value oriented

[Edit: As I note in one of the comments, I wrote this post right before Google announced new models during OI day. I didn’t mention in the post, but I figured they would have new/better models that would also be comparably cheap. It’s very possible that the opposite of that has happened. Gemini Flash 3.1 is no longer available in Anti-gravity and all online discourse says that the replacement Flash 3.5 will be substantially more expensive. I didn’t even get a full day to be “not 100% wrong yet anyway.”]

One aspect of AI discourse that I find fascinating is that it ports in our historical assumptions about the value of human time. So even though the machines are replacing incremental amounts of human work each day, we still pretend as if our time scarcity should be an important consideration.

For instance, it is widely accepted that Claude writes much better code than most other models - with OpenAI’s Codex perhaps being neck and neck. The other models are also-rans for now.

However, I have been using Gemini inside Anti-gravity and while it does lots of stupid stuff, it will eventually get the job done. It just takes more total iterations and more total frustration to get there. The most common frustration is that it will make a change and then report back that the code should run flawlessly… except it didn’t test shit. It has no idea whether the code runs at all. Let alone flawlessly.