No. of Recommendations: 3
Manlobbi:
The AI systems are basically dealing with section 2 only, and not involved with the ludicrously more rich semantic/syntactic conversions systems of 1 or 3. They don't need to be, as it is ourselves that are then using our conversion systems (particularly 3) to interpret what the AI system is producing as having meaning, even though the AI system never dealt with 1 or 3, and has no understanding of anything it is producing.
You seem to have completely ignored the Neural Net. It is said that GPT is a text predicter, but what is it intended to predict? Its not made super explicit, but what it seems to be aimed to predict is what a human would respond to the text input... IF the human was incredibly well informed, which seems to be approximated by having read everything on the net.
I am a predictive text machine, you are too, so is Stephen King. I read a post then create text as a response. You hear someone say I'm going down the street to get some milk, do you want anything, then you create text as a response. Stephen King's editor says to him I'll give you another $20 million if you give me another bestselling book and then Stephen King creates text.
What we have found is we (well they really, I guess, I didn't help) can build a predictive text program that behaves a lot closer to what Stephen King would do than we were thinking we could. The outputs from GPT are spectacular, even as they are flawed. Much like a human's outputs might be, actually.
And what is going on inside that neural net? Is some sort of essence-of-stephen-king created after every prompt, which then, at GHz clock speeds, spends the equivalent of a year working on a response, and then gives that to you seconds later? WE DON'T KNOW!
But whatever it is that is going on inside the neural net, it seems like it must be similarly complex as what is going on inside our heads when we are acting as predictive text creators. It would seem to us that you NEED that complexity in order to get the meaningful responses we get from each other. And so we wonder, if GPT is giving us meaningful responses, what is going on inside its neural net?
The good news is GPT's are still pretty young so it will still be many years before they get the vote.
R:
