IMO GPT-2 is immensely superior to Eliza in that it does embed some idea of context, it probably transcends Parry's scripted model of schizophrenia, but I think it teaches the same lesson as Eliza: A Clever Hans understanding of linguistic structure will fool way too many people.
Exactly: ELiza and GPT-2 are cut from the same cloth; they leverage human gullibility in part by combining phrases that were prewritten by human beings, and parroting user input, but lack any genuine conceptual understanding of things like human relationships that they discuss.https://twitter.com/yoavgo/status/1199393534786310146 …
-
-
-
maybe on the first; definitely agree on Clever Hans part
End of conversation
New conversation -
-
-
I think, in essence, GPT-2 is a more sophisticated mechanical Turk: https://en.m.wikipedia.org/wiki/The_Turk
-
yeah, there is no "intelligence" in any of the language models thus far and as a result they don't really show intelligent behavior. I guess i would define intelligence in language models as converting natural language to a world model and then back again (which would be useful)
- 1 more reply
New conversation -
-
-
And more than that! (as you pointed before): there's selection (and possibly, editing) of outputs by humans. It's like making photographs of smoke or fire (see two of mine as an example :) Is it the smoke who is the artist? Of course not.pic.twitter.com/MptEQU03of
-
That's exactly what inventors of photography claimed. Henry Fox Talbot took photos of his house and said it had *drawn its own picture*. The AI industry is aping ancient sales techniques atm.
End of conversation
New conversation -
-
-
True. The difference here though is that when Weinzenbaum saw the amount of misinformation around what ELIZA was and wasn't, he actively spoke against it. OpenAI seems to be doing the opposite.
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
1.) Play chess against a human online 2.) Fire up chess game on my end and set the AI to hardcore 3.) Repeat every move my human opponent makes to the chess AI 4.) Repeat every move the chess AI makes to the human opponent
-
This successfully demonstrates that not a single entity in this scenario understands chess. It shows everyone is able to mimic the behaviors that allow us to signal we understand chess. Signalling is enough for most people. Deep understanding is rarely required.
- 1 more reply
New conversation -
-
-
You're right. It's unclear to me why anyone averagely skilled in machine learning would think otherwise. For me, the big question is - why is this is an even slightly controversial view? Are people being disingenuous? Or is this stuff harder to understand than I think it is?
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
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.