Apropos of AlphaGo (BetaGo?), why reinforcement learning AIs are here to stay: agent AIs are smarter & fasterhttps://www.gwern.net/Tool%20AI
-
-
Replying to @gwern
"google maps is not an agent" and yet it does things like this http://www.bldgblog.com/2017/01/the-season-of-burning-trucks/ …pic.twitter.com/8QimN3k436
2 replies 0 retweets 1 like -
Replying to @allgebrah @gwern
maybe I'm just rephrasing you here, but tool AI and a company using it combine into an agent AI (albeit with a longer feedback loop)
2 replies 0 retweets 1 like -
Replying to @allgebrah
But you need to close the loop for it to be a secret agent. Just people getting into trouble doesn't close it. Need A/B tests etc
1 reply 0 retweets 0 likes -
Replying to @gwern
should the assemblage of AI and company become large enough to be incomprehensible/illegible though, you get the same problem again
1 reply 0 retweets 0 likes -
Replying to @allgebrah @gwern
humans in the loop make the feedback loop manageably long, which is good, but they'll miss things that look innocuous on their own
1 reply 0 retweets 0 likes -
Replying to @allgebrah @gwern
to spot those, you'd need an about equally powerful AI (maybe a sibling?) - is that what you mean with A/B tests?
1 reply 0 retweets 0 likes -
Replying to @allgebrah
No, I mean it needs to optimize for something based on observing things affected by its classification/predictions.
1 reply 0 retweets 1 like
and how do you plug other less-explicit info leaks into the box like, say, grad student descent? (love that term)
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.