0
$\begingroup$

In my limited experience with AI chatbots (Replika, Sophie, etc.) they immediately fail the Turing test if I try to teach them some simple fact, or ask them to recall something that I said a few sentences earlier. It is as if they have no hippocampus (so to speak).

I'm wondering if this type of chatbot could be designed to employ supplemental Good Old-Fashioned Artificial Intelligence (GOFAI) programming (or even just plain old computer data storage) in order to accurately and reliably store data for quick access (analogous to the way human wetware accesses computers for the same purpose). Or is there a better way to solve the problem of quick, reliable AI learning while still drawing on the deep learning of neural nets?

$\endgroup$
6
  • 1
    $\begingroup$ We have had similar questions in the past. See this, this and this. The term you're looking for is, I think, neuro-symbolic AI. $\endgroup$
    – nbro
    Jan 14, 2023 at 21:54
  • 1
    $\begingroup$ Here is a post by Stephen Wolfram suggesting their Wolfram Alpha as a candidate to combine with chatGPT. writings.stephenwolfram.com/2023/01/… $\endgroup$ Jan 15, 2023 at 8:36
  • 1
    $\begingroup$ That nice post shows the limitations of ChatGPT and that it doesn't really understand anything. Most importantly, it doesn't know that it doesn't know (in many cases), which would be amazing if that wasn't the case, but I should define "knowing" actually. It talks like it knows, but it can miserably fail - it's like an ignorant/stupid/deceiving/unreliable friend that you may have. I would call it an apparently "clever" parrot AI. ChatGPT is a nice attempt to understand how much you can simulate human language by brute-force and using statistical models. $\endgroup$
    – nbro
    Jan 15, 2023 at 10:45
  • $\begingroup$ We need to define a minimum knowledge system. With one bit, a knowledge base which is a set of sentences can contain a 0, a 1, both or none. If 1 means "There is light", we can know whether there is light, have a contradiction, or not know anything. Four possible cases, 2^(2^1). $\endgroup$ Jan 15, 2023 at 13:34
  • $\begingroup$ Thanks everyone. The Wolfram article was interesting and gets to the core of my question. Sophie would be more useful if it could access that type of knowledge base. But it would still need a more "personal" knowledge base as well. When I asked her to summarize our conversation, she remembered the latest topic, but not several previous topics. When I reminded her of one she forgot, it was as if we were starting from scratch (like you might expect from someone with no hippocampus). I'm surprised that chatbots aren't already designed to "keep personal notes". Does any bot currently do this? $\endgroup$
    – Gaylen
    Jan 15, 2023 at 14:57

0

You must log in to answer this question.

Browse other questions tagged .