[gremlinsinc]

I don't think this is how it works, or how any of this works. It just 'inferred' based on previous input the changes. It does the same for coding, if it gets things wrong and you feed it data from docs, or tell it what it did wrong it gets better at answering related questions.

[detrites]

Could that also be a valid description of what we do?

[lordfrito]

Yeah, maybe.

My gut says there's something interesting in this interaction that's worth trying to understand.

The key here is that he got ChatGPT to ask him questions. I was like, wait, how does that happen? And his solution was ingenious. He asked it to play a game and taught it the rules. I think it's more than a passing coincidence here that human children also rely on play for learning (they're hardwired for it!).

Is it thinking, probably not. But it sure does look like it's self-directed reasoning, which is a huge step in the right direction.

It's not AGI, but maybe points us in the direction of a cognitive subsystem that we need to investigate/map/theorize/design/implement/iterate that can automate the decision process that the author was prompting the AI to follow.

Essentially, make it so the AI doesn't need a human to follow that line of reasoning, in a general sense. Find a way to let it ask it's own questions, and seek answers. Probably a later phase involves the AI asking humans for help along the way when it's reasoning gets stumped.

So much to learn. :)