Eh, has there been any update since the last posting? https://news.ycombinator.com/item?id=27443528
I will admit, I'm especially impressed that it caught the "as BS" part in its code generation -- then again, it's entirely possible that a lot of code examples do that, so maybe a more stunning result would be "as BSOUP" and see if it still generates plausible code
what i understand from a completely armchair perspective, is that this kind of context-holding is THE feature of this wave of ai models, which is what similarly allows it to generate paragraphs of prose seemingly out of whole cloth, tracking what it's talking about as it rambles along. the output seems (with occasional glaring issues) remarkably cohesive and realistic, given that it has nothing original to say. I picture it as taking sentence structure it's seen before, and narrative structure it's seen before, and topics it's seen before, mixing them all up, and putting them back together like multidimensional legos- from this perspective, processing the 'import as' lego seems to fit right in.
The whole concept is fascinating, and i'm immensely relieved that we seem to have such a competitive open-source implementation of this nightmare