-
the AI code generation stuff i've seen doesn't seem like it's much beyond what github copilot could do? just with a different interface. and i tried using copilot for a bit but turned it off because it had the same keybindings as regular autocompletePermalink Look on archive.org
❤️ 9 Favorites Mood +2 🙂
-
and regular autocomplete was more limited but also told me things that were almost always true, rather than 80% true. and it turns out the problem isn't coding, but instead understanding the problem & understanding how the current code works.Permalink Look on archive.org
-
and it is easier to understand code you just wrote, rather than code a ML model just hallucinated. if it was a coworker's code you could trust some aspects, you could build a mental model of what they were going for, you could believe that a plausible variable name would resolve.Look on archive.org
-
anyway, it is cool to see this stuff, but it is still the case that ML stuff is interesting because it gives us a way to understand large datasets from a new angle, not because it has some new access on truth or beauty.Permalink Look on archive.org
-
this thread inspired by this one @freezydorito/1599507538604457984Permalink Look on archive.org