I came across GPT-3 today.
An AI being able to "create" a pastiche of Jerome K. Jerome's writing is very impressive:
[
technologyreview.com...]
But
then I read that Sharif Shameem experimented with GPT-3 to see how well it coped with
a high-level programming language, prompting it with some examples of JSX. And then, when he told it to make:
a button for every colour of the rainbow
It did so. In JSX. No, really:
[
twitter.com...]
Did anyone believe that we would already have AI
this capable in 2020?
N.B. It's not all positive. Some people are shocked (and some aren't) that GPT-3 can easily be prompted to produce xenophobic or sexist output.
Well of course it can - if the AI has been trained using English language discourse from the last four centuries, that's surely a given? If you seed an AI by telling it consistently and repeatedly that multiple human races exist and that Eve was made from Adam's rib and never tell it that this is pseudo-scientific (or mythical) claptrap to justify the politics of colonisation and patriarchy, it's probably going to take you at your word. It doesn't know better unless you tell it.
