Category: AI, Computation, and Technology

RT @curtdoolittle: @RealLouisPrince @NektariosAI @GaryMarcus Hmm… If I do a few searches, and find that no one else is discussing the sub…

I’m pretty impressed that it can write in the ‘style of curt doolittle’ and it’s pretty accurate. Now the logic I put in the text isn’t there. But the summary of the work it produces IS logical.
So that’s the interesting generalization that has promise.

I’m not ‘fooled’. Given I understand the technology (and dropped out of it in the 80s because of the then hardware limitations), and I understand rather deeply the human brain. I’m asking a technical question of whether a world model that allows logical testing of sequences of operations etc, and predictions from them, can emerge from enough self reinforcement. You wouldn’t think so. I woldn’t think so. But I can see how it might be possible.

We are in the first stages of considering how to put our ethical model into it, and I can see how that might work … eventually.

So I’m just reserving judgement.

Reply addressees: @j_q_balter @NektariosAI @GaryMarcus

I’m not ‘fooled’. Given I understand the technology (and dropped out of it in the 80s because of the then hardware limitations), and I understand rather deeply the human brain. I’m asking a technical question of whether a world model that allows logical testing of sequences of operations etc, and predictions from them, can emerge from enough self reinforcement. You wouldn’t think so. I woldn’t think so. But I can see how it might be possible.

We are in the first stages of considering how to put our ethical model into it, and I can see how that might work … eventually.

So I’m just reserving judgement.

Correct. But it sure speeds up the process.
If you can feed it a list of definitions and rules (Microsoft is about to release it) then it’s over.
GPT can do that. Just not deeply enough.

Yeah. Imagine daily life when these things can be anywhere at any time.

I’ve been working on these issues since the 80s. And until I saw the emergent behavior in GPT-4, I would have agreed with you. I’d thought we’d need a world model, a prediction (autoassociation) manifold, an evaluative model (preference, ethics) and a language model. But I’m a little concerned that I might have been wrong, and that it’s possible to construct all of them from the language model – at least incrementally.
I dislike this evolutionary pathway because the systems aren’t discrete and controllable(auditable). So we aren’t overcoming the problem of the human mind (and correcting it).
And so I’m not disagreeing. And while I’m skeptical as you are. I’m back to reserving judgment.

Reply addressees: @NektariosAI @GaryMarcus