I use it constantly for both Python and Rust projects, and it works great. But it needs to be treated like a pair programmer with oversight and checking. Still, the over productivity boost is enormous, especially for automating routine stuff.
It does struggle with Rust sometimes, certainly more often than with Python. I think that partly comes down to there being more Python code in its training set than Rust, and, Rust won't silently pass buggy code like Python often will.
Nice! I haven’t been using it quite as much lately actually, since I’ve generally found that it leads to an explosion of complexity unless I heavily refactor what it generates. But the newer o3 and o4-mini-high models from OpenAI are incredible for riffing on system design stuff. Ironing out SQL table designs and overall program structure. I’m sure it won’t be long until the models are much better at managing code complexity though. We need to factor code size into RL training.
