Re:Building a Virtual Machine Inside ChatGPT
by mxmlnkn on Saturday December 03, 2022 @06:26PM
(#33848601)
I'm not sure that this is an insurmountable problem.
There are two kinds of "loops" with this. First, it never learns in inference mode. It's basically "dying" / being reset after closing each chat session because it cannot reference any of the chat sessions it did. But, this could be remedied with online training on all dialogues that it did.
Then there might be the problem that it keeps spewing out the same thing over and other if there is no further input from a human. I'm not sure this is the case. But, if it is, this sounds like something that could be solved by combining it with a training algorithm incentivizing curiosity, i.e., promote replying with and seeking out new unknown data. This might be a tradeoff with efforts to reduce hallucinations though. But it might even lead to it asking the user questions instead of just answering like Wikipedia.
There are two kinds of "loops" with this. First, it never learns in inference mode. It's basically "dying" / being reset after closing each chat session because it cannot reference any of the chat sessions it did. But, this could be remedied with online training on all dialogues that it did.
Then there might be the problem that it keeps spewing out the same thing over and other if there is no further input from a human. I'm not sure this is the case. But, if it is, this sounds like something that could be solved by combining it with a training algorithm incentivizing curiosity, i.e., promote replying with and seeking out new unknown data. This might be a tradeoff with efforts to reduce hallucinations though. But it might even lead to it asking the user questions instead of just answering like Wikipedia.