Also, I suspect most "production" Swift –the type of stuff written by seasoned experts– (I just had to add em-dashes ;) is behind closed-source walls.
replyNo doubt in my mind, a future Apple model will be the best to use for this purpose. They likely have more swift to train on than anyone else, and would benefit directly from more quality apps, rather than the slop flowing into the App Store (>1k app submissions per hour; they claim)
reply> which is only going to get worse as it "learns" from garbage outputted by other LLMs moving forward
replyYou seem to assume that autoregressive pretraining (and unfiltered behavior cloning, maybe) are the only ways to improve LLM performance.