It is easily a 6x performance increase in the amount of code I can write. I either have to use it, or get laid off.
It is easily a 6x performance increase in the amount of code I can write. I either have to use it, or get laid off.
I’m a huge fan of Sonnet but I can’t reconcile 6x. It’s a faster keyboard. Stuff like aider architect chat or whatever seems relatively useless.
What’s the secret for 6x?
However, writing that kind of code maybe makes up 5% of my work. Analysis, trial-and-error, discussions etc. make up the other 95%, and AI only seldomly helps with that. It can sometimes be useful for research and spec ingestion, but it quickly becomes dangerous in those cases because as soon as you enter any kind of niche area (and unfortunately my work has a lot of those) LLMs tend to hallucinate and present made-up "knowledge" with enviable certainty.
It’s great at boilerplate like filling in configs, but the results of iteratively coding with AI are first draft quality.
Yesterday I used copilot to kick out a really quick REST api for a flask app. A lot of the code was boilerplate parsing args and DB lookups, so it was fine.
When I changed one of my db models it broke all the hard coded error handling in frustrating ways.
After writing suitable helpers and rewriting the module (with AI) it cleaned up into some good code.
Definitely speed up the exploration phase, but required a top to bottom rewrite for the final product. Overall I’d say a little faster than me doing this myself (but I don’t use flask often)