Hacker News Reader

Comment by 6d6b73 | Hacker News Reader

GPT doesn't understand the questions therefore it can't answer them. It can only generate a bunch of text and make it look like it knows what's being asked. It's not AI, not even close.

View on HN

Yesterday, I gave the AI Dungeon version of the model the questions from the first round of the most recent episode of Jeopardy (almost certainly not in its training set). After I burned the first category on priming it, so it knew the kind of task we were doing, it got 72% of the remaining questions right. That's not just making it look like it knows what it's being asked.

That answer is too easy.

What does it mean to „understand it“?

How do you know how you answer a question yourself?

Presumably, what you mean is that you can stop and _reflect_ on what was asked. You‘re aware of it. And if it‘s complicated you might be able to reason and break it down.

GPT-3 certainly doesn‘t do that, but all that means it‘s not conscious.

Next step might be two/more GPT instances having internal dialogue, trying to achieve a consensus before giving its answer. Would also give us opportunity to peek inside its inner workings.

Or a layered system. The surface instance takes the question and produce an internal "thought" out of it, a statement or question itself, and submits it to the layer underneath. This second instance works out a response / text extension to the statement/query from the surface, submits it up, and the surface layer generates the final answer by answering or extending the response of the sub-layer.

This can be extended down or wide, with several instances representing different thinking modalities and the final answer being generated from all of them. Basically reproducing a neural net architecture but at the higher abstraction level.

I think we're close to the point where this sort of attitude becomes a form of theology.

Reply on news.ycombinator.com