1
$\begingroup$

Alpha Zero is good at looking into the future to plan it's next move. GTP-4 is good at generating language from previous text.

It seems like combining these two systems would create a general conversation and problem solving AI.

For example, in response to a question, the AI could generate a thousand candidate replies using GTP-4, then generate a thousand suggestions of how the human might reply to each of those responses. This chess-like thinking is exactly what Alpha Zero is good at.

By evaluating it's own responses and predicting which ones would lead to a successful outcome, this is more or less what human's do.

(The only difference might be that humans have a longer memory than a GTP-4 algorithm).

Improve this question
$\endgroup$
1
  • 1
    $\begingroup$ I would argue that AGI is much more than handling generic conversations. Specifically, I think such model falls into the category of models discarded as AGI by the Chinese room thought experiment $\endgroup$
    – Edoardo Guerriero
    Commented Jul 14, 2022 at 7:35

1 Answer 1

Reset to default
1
$\begingroup$

The main thing that I can spot that is missing from your outline of "a general conversation and problem solving AI" is that there is no assessment of direction or outcome.

In AlphaZero, that is provided by a game rules engine which tells the AI when it has won. There is no such game rules engine for open-ended conversation, or in general for a goal-based conversation such as technical support or sales. Without this, an engine like GPT-3 has nothing to ground it to any particular text other than statistics - it could pick the most (or least) likely progression of a conversation, but not use any concept of utility for itself or the other participant in the conversation.

The grounding problem is also still a major issue independently. A conversation "solving" AI may have no concept of the subject of discussion, or of the state of the person it is talking to. It models the language around these things - e.g. it may adjust to a human expressing fear or excitement. But it has no concept of what these mean, beyond finding matching text in its reply. Importantly it has no state model for the other side of the conversation beyond what is written.

You may find you need to solve the grounding problem though in order to perform any assessment of a conversation in a general context.

By evaluating it's own responses

How, on what basis? That's the problem that still needs to be solved after combining the two technologies.

Improve this answer
$\endgroup$
5
  • $\begingroup$ Well there are many "sentiment" analysis AI's, so the "goal" of such an AI might be to elicit good "happy, engaged and non-angry" replies from the human. As such the primary focus of such an AI would be to make a human laugh or be happy and not bored. A human might become angry and bored if the AI started talking gibberish for example or get the human's name wrong. You are right that a main thing it lacks is the ability to see a human's face and judge it's emotional response. But then again, humans can communicate through text too. $\endgroup$
    – zooby
    Commented Jul 15, 2022 at 17:14
  • 1
    $\begingroup$ @zooby That might work to aid the search for responses that try to guide the conversation as opposed to simply fit it, although I'd hestitate to call the resulting system an AGI by any measure (it still has no internal model of what it is doing or why), it might make for an interesting chatbot design. $\endgroup$
    – Neil Slater
    Commented Jul 15, 2022 at 18:54
  • $\begingroup$ True, but then we might say Alpha Zero has no model of how to play chess and yet it can play chess! $\endgroup$
    – zooby
    Commented Jul 16, 2022 at 0:30
  • $\begingroup$ AlphaZero, when trained, does have a model of how to play chess, using the same meanings as I am using. Although it's model of how time/turns work, and the rules of the game are hard-coded, not learned. The difference I am trying to express is that a system like GPT has a model of how words work, but the interpretation of what those words mean and how they are being used is all external to it. It is a sophisticated symbol prediction machine that only has references to the symbols and how they combine in its model. $\endgroup$
    – Neil Slater
    Commented Jul 16, 2022 at 8:49
  • 1
    $\begingroup$ You can try to extend the model to add meaning, in a similar way that AlphaZero's NN is extended to understand taking turns and the rules of chess (i.e. with additional models). By modelling a conversation with GPT-3, you would already be providing something (although a very fixed taking-turns conversation model, it is something). Same if you tried to guide to positive sentiment. What I am saying is that is nowhere near enough to claim to have an AGI. $\endgroup$
    – Neil Slater
    Commented Jul 16, 2022 at 8:53

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .