alt Hacker NewsI have been a nay sayer on LLMs/GPTs in general having tried many, but recently Ive been shepherding a fairly complex code build through the latest opus model and its quite impressive.
It still gets things wrong occasionally but the time its saved me has been substantial. Im starting to enjoy it.
I recently built a reasonably-complex embedded controls project using codex and an esp32.
Starting with systems stuff like "Set up vscode with whatever it needs to work with codex and talk to an esp32," and ending with "Now add a web interface with persistent tunables that always runs in both AP and station modes," my prompt inputs were very terse.
And it'd just kind of go forth and just do it. It'd even design and run its own tests.
I never once looked at the code. For all I know, the code doesn't even exist.
And it works. I'll be using it in the field (in the proverbial middle of nowhere) all next week. I have every expectation that it will behave itself.
(I did spend a lot of time defining and refining some ground rules with AGENTS.md, but in theory I get to re-use that effort for the next go.)