←back to thread

38 points 01-_- | 1 comments | | HN request time: 0.206s | source
Show context
ssivark ◴[] No.44396070[source]
If I have decent autocomplete where I type half the characters and the AI predicts the other half that technically satisfies this metric.

Notice the loophole: there’s no qualification of how much problem context the AI started from. Most of the problem -> code “work” would still be done by a human in that situation — even if technically 50% of the code is “AI generated” [because the human did all the hard work of generating the context necessary for those tokens, including the preceding tokens of code].

As the saying goes… lies, damned lies, and statistics.

replies(3): >>44396426 #>>44397536 #>>44399676 #
1. nunez ◴[] No.44396426[source]
Funny, and ironically, enough, I turned off autocorrect on iOS after it moved to a GPT-2 model because it grew increasingly inaccurate the more I used it. (The Markov chain implementation that preceded it wasn't much better, though I remember autocorrect on iOS being significantly better many years ago.)