[ExI] Training GPT-3 to play nice - it's difficult!
spike at rainier66.com
spike at rainier66.com
Mon Nov 2 17:09:22 UTC 2020
-----Original Message-----
...> On Behalf Of BillK via extropy-chat
Subject: [ExI] Training GPT-3 to play nice - it's difficult!
>...How to make a chatbot that isn’t racist or sexist Tools like GPT-3 are stunningly good, but they feed on the cesspits of the internet. How can we make them safe for the public to actually use?
by Will Douglas Heaven October 23, 2020
<https://www.technologyreview.com/2020/10/23/1011116/chatbot-gpt3-openai-facebook-google-safety-fix-racist-sexist-language-ai/>
...
>...BillK
_______________________________________________
Hey cool, thanks BillK, can we get a copy of that and feed it on the Exi-chat archives? That would be a hoot. Perhaps we could train it with that, then see if we can disguise it as a prole, see who catches it first. We could get have ExiMod get in cahoots with someone and use their @ so it wouldn't draw suspicion. Hey cool even better: get one of the old-timers who has written volumes, train it with just that, then use that guy's @ and see if we can fool the public.
Dang I think I just gave away the game.
Hey better idea: get archives from two opposite personalities, train it up on that. Hard telling what we would get: an average person or an apparent split personality.
(Side question: BillW and other psychology hipsters, that whole split personality business, is that real or is that Hollywood?)
We could have a lotta fun with this. We could take it over on one of my math forums and one of my motorcycle forums. People over there treat each other with respect and behave themselves more than we do, but perhaps that wouldn't be too much of a giveaway.
Another fun experiment: find old posts from Mike Lorrey, Lee Corbin, Mike Butler and Daniel whats-his-name, train it on those four, see if ExiMod boots it. {8^D That one would be cool because ExI would warn her (it?) but of course she doesn't learn from the warning, so you know she would be cruisin' fer a bruisin'. Alternative: first warning from ExiMod, tip off ExiMod that it is a joke (sorta) and the Turing Test has just been passed (sorta.)
BillK, do you know what is the best platform that we can get to? Some kind of public domain something where we give it the feedstock text?
spike
More information about the extropy-chat
mailing list