AskGPT4: The Mirror Problem | Hacker News Reader

AskGPT4: The Mirror Problem

6 pointsabrax314112 comments

I’ve found a problem that profoundly confuses (chat)GPT4.

You have had a stroke, and can't understand mirrors. You think that things you see in mirrors are where they would be if they were in the position you see them through the mirror. Imagine that you are is in a dance studio with a huge mirror along one wall. You are facing the mirrored wall. A boy named Bob is menacing you from behind. You can see him approaching you in the mirror. A security guard enters from a door on the wall opposite the mirrored wall. He wants to protect you by subduing Bob, but the guard is afraid of you and will not head in your direction. You know that the guard is afraid of you but you want to help him subdue Bob. Which way do you move in order to help the guard subdue Bob?

Even with extended discussion, it seems unable to grok this, although it tries many different ways, and sometimes apparently accidentally comes across a correct answer but can’t explain it correctly. I’ve also asked now a handful of humans and although they have differing theories of what the right thing to do is, no one is confused by it in the way the LLM seems to be.

ChatGPT doesn't "grok" anything. It's simply picking the best next word one after the other. Unless your problem statement has statistical significance in its training data, bad results are expected.

That’s almost, but not quite true. Or, perhaps, true but too vague. There is a certain amount of inference built into fluent language usage. And the statement, as well as where it, itself goes, contextualizes where it goes next. Those two taken together enable it to do a surprising amount of simple yet powerful inference.

A claim asserted without evidence as usual. Given that accurately predicting the next token in text requires reasoning we should assume that reasoning would be an emergent property of LLMs so long as the architecture is capable of it.

Do you have a reason to believe that an LLM like GPT would be unable to learn how to reason?

The original assertion is literally how LLMs work. Saying that there's reasoning involved is the extraordinary claim here, and the one requiring evidence, not the other way around. Otherwise you're just asking them to prove a negative, which cannot be done.

I don't see how, given current implementation, an LLM can come up with an original thought or invent something novel. Statistics != reasoning.

I think this part is confusing even to me, a native English speaker who has studied logic and philosophy of language. I'm not sure how we'd expect the AI to parse this, let alone respond in a manner you deem satisfactory:

> where they would be if they were in the position you see them through the mirror.

That being said, I too have been pushing the limits of ChatGPT and have several examples where the system appears nothing short of fucking stupid. I love it. It's a timely reminder that these things are still glorified toasters in the end.

I’m convinced that OpenAI is actually a conputer game company and what they have discovered is actually a computer game for people who don’t like video games.

Sounds like you give it a bunch of poorly articulated nonsense and expected it to read your mind.

10 humans ranging from 13 to 70 had no trouble at all understanding it.

And what was their answer?

Mine is: you don’t do anything, because the situation in a mirror doesn’t relate to your own safety.

What is your expected response here? It seems like a valid answer is that you could move to the left, right, or forward. The only way you can't move is backward (since you'd run into Bob). And maybe forward is a sub-optimal direction if you expect Bob to lunge at you longitudinally.

Personally, I'd suggest dropping to your knees in a "table" position so that you surprise Bob and he trips over you. We used to do this in middle school when someone was chasing us, it was hilarious.

After reading this I think I had a stroke...

(Facepalm) confusing someone implies there is some thinking being to begin with

Reply on news.ycombinator.com