Think of GPT like a document simulator. If 90% of documents on the web talk about an old version of a library, then GPT-4 is likely to use that old version.
When I asked it write me a Chrome extension, it used Manifest V2, which was already slated for deprecation. When I asked for a V3 extension specifically, it was happy to comply. So even if a new release came out before Sep 2021, if it hasn't had time to become the dominant version in code examples around the web, GPT-4 may still use older versions.
My point (which may not have been clear) was that when I explicitly told it that the versions of SQLite I'm using supports "returning," it still returned incorrect code. It did not use "returning", and repeatedly, even after being corrected, grabbed the row count instead of the column value.
I understand that it's a language model - my point was that in my experience, the depth of training data just isn't there for alternative implementations. It can do the things you ask one way, maybe two or three, but if you know how you actually want something done you're probably better off writing it yourself (for now).
Its dataset cuts off around 2021. There's a little footer message warning you not to expect knowledge of recent events.