Hacker News Reader

Comment by p-e-w | Hacker News Reader

Reading ChatGPT's responses from the post made me genuinely afraid.

Yes, if you pay very close attention it's clear that ChatGPT doesn't fully understand some concepts. But damn, it's close. Some of the inference it does is fucking unbelievable. It's beyond my comprehension how this can be achieved by just training on text inputs, even if there's Terabytes of them.

We have to carefully define the Turing test now when talking about it, because this system would obviously pass many versions of it with many people.

View on HN

> if you pay very close attention it's clear that ChatGPT doesn't fully understand some concepts

> We have to carefully define the Turing test now when talking about it, because this system would obviously pass many versions of it with many people.

I would guess many children, who we would define as 'passing the Turing test' wouldn't explain the joke nearly as well, or fully understand all the concepts.

Agreed.

I've been thinking of GPT as an incredibly well-read 8-year-old. Smart, but sometimes wildly silly and not yet developed in a sound understanding of the world (but getting there!)

I don't agree. ChatGPT explains all jokes as a play on words that defies expectations. It's as if it's learned "A joke is a play on words that defies expectations".

I'd like to try this game, but OpenAI insists on having both my email address and my mobile number. Sorry guys, that won't wash, all I want is to have a lark. I'm not dropping my pants for that privilege.

As far as Turing tests are concerned: I notice that ChatGPT didn't "get" any of the jokes. If I don't get a joke, it's "Huh?" or "I don't get it". So you can tell ChatGPT from a human by asking it to explain a joke that simply isn't funny.

Apparently the neural network has converged onto a model that can to some extent model the concepts that language talks about.

It is also simply good at making sentences, which can help it bullshit its way through a conversation, but you may note that the output stays really close to the models limited understanding of a concept (e.g. joke = subverting expectations) without straying much even if the situation demands it.

I had the exact opposite response. I was relieved that it could make guesses that sounded sophisticated, but missed the mark entirely.

It made it sound like the guy in the room who thinks he's smart but has no grasp of subtext, and I think those guys are more likely to be mistaken for robots than robots like that are to be mistaken for humans.

Fully agree. ChatGPT sounds like a pretentious savant in comedies from the 17th century. It "knows" many things by name, but doesn't actually understand anything.

Reply on news.ycombinator.com