Hacker News Reader

Comment by theptip | Hacker News Reader

This is a terrible eval. Do not update your beliefs on whether LLMs have Theory of Mind based on this paper.

The eval is a weird, noisy visual task (picture of astronaut with “care packages”). Their results are hopelessly narrow.

A better eval is to use actual scientifically tested psychology test on text (the native and strongest domain for LLMs), for example the sort of scenarios used to gauge when children develop theory of mind (“Alice puts her keys on the table then leaves the room. Bob moves the keys to the drawer. Alice returns. Where does she think the keys are?”) which GPT-4 can handle easily; it is very clear from this that GPT has a theory of mind.

A negative result doesn’t disprove capabilities; it could easily show your eval is garbage. Showing a robust positive capability is a more robust result.

View on HN

> A better eval is to use actual scientifically tested psychology test on text (the native and strongest domain for LLMs), for example the sort of scenarios used to gauge when children develop theory of mind (“Alice puts her keys on the table then leaves the room. Bob moves the keys to the drawer. Alice returns. Where does she think the keys are?”) which GPT-4 can handle easily; it is very clear from this that GPT has a theory of mind.

Aren't you confusing having a theory of mind with being able to output the right answer to a test? Isn't your proposed evaluation especially problematic because an "actual scientifically tested psychology test" is likely in the training data along with a lot of discussion and analysis of that test and the correct and incorrect answers that can be given?

How do I know you have theory of mind or are concious if not the "right" response to a test ?

As far as I'm concerned, the only person I can be certain is concious is me.

It doesn't have to be a "scientifically tested psychology test"

Construct your own story with multiple characters of varying knowledge and beliefs and see how it does.

You're missing the point.

With prior knowledge of the test, even something that verifiably lacks the cognitive capability to legitimately pass the test (e.g. FizzBuzz level stuff) can pass by cheating.

Yes, but if you keep shoving tests at it, including brand new tests, and in total more than it could've ever memorized, and it keeps passing, then perhaps it's no longer cheating.

We are not incapable of creating new tests if that's the issue

Of course when testing an LLM you do not use the exact text from studies in the training set. There are an infinite number of structural permutations that preserve the methodology sufficiently to be comparable to the existing literature, while being distinct enough to be outside the training distribution.

Of course, another approach you can take if you suspect contamination is to ask follow-up questions that are less likely to be in any training data; if the LLM is a mere stochastic parrot without a ToM it will not be able to give well-formed answers to follow-up questions.

Perhaps if I'd said "established psychology research methodology" the intent would have been clearer?

Or there are enough of those examples in the training set that it can guess well. Not sure how such an example would prove anything when we know an LLM is just guessing the best words.

Nothing I’ve seen shows evidence of any sort of abstract concepts in there.

Wouldn't this also be the same for humans?

If you introspect and decide it is so, I won't disagree with you.

Almost every discussion about consciousness or human level intelligence eventually devolves into questioning whether everyone apart from you is just a robot

Negative, I am a p-zombie.

Yes, in normal conversation and written correspondence. Sure, humans have deeper wells of reason, emotion, and recall to draw from, but much of small talk is pure reflex.

I'm wary of what happens when AIs can interface directly with the real world. LLMs only read about the world second-hand. An embodied neural net that could observe cause and effect, push pokers into fires and witness the metal glowing red, then learn from the experience. Spooky.

Reply on news.ycombinator.com