Hacker News Reader

Comment by dr_dshiv | Hacker News Reader

I do! So long as we don’t appeal to consciousness as the basis for understanding but measure it with benchmarks, LLMs understand many things.

Maetschke, S., Iraola, D. M., Barnard, P., Shafiei Bavani, E., Zhong, P., Xu, Y., & Yepes, A. J. (2021). Understanding in artificial intelligence. arXiv preprint arXiv:2101.06573.

View on HN

I saw this example on Twitter today which shows ChatGPT doesn't understand:

Mike's mother had 4 kids; 3 of them are named Luis, Drake and Matilda. What is the name of the 4th kid?

I'm sorry, but I don't have information about the fourth kid. The information you provided is that Mike's mother has 4 kids, and three of them are named Luis, Drake, and Matilda. However, the name of the fourth child is not specified in the information you gave.

But thanks for the paper. I'm reading as much as I can.

I perceive that as: misunderstood a riddle, but understood the language

I had to read that 3-4 times before I understood that it was a riddle. So I put it in GPT3–and it got the answer right away:

“Mike's mother had 4 kids; 3 of them are named Luis, Drake and Matilda. What is the name of the 4th kid?

The fourth kid is Mike.”

So, maybe they just needed to try again.

It's not a riddle because the answer is plain to anyone who understands English. I'm not saying you don't but the answer is right there.

I asked it a followup:

The name of the 4th child is in the question

Ah, I see! My apologies for missing that. In that case, the name of the fourth child is Mike.

I had to read it multiple times before I understood why it was a test of understanding. It is a riddle.

even if you classify it as riddle, you still need intelligence to "solve it", which is what is missing in LLM. the good answer you later got was probably prompt created by someone manually, to further feed the model.

did you need more data to solve the riddle ? rather more attention to what you read than more data, and that's the sole issue with AI/ML in current form.

Q: A week back I had a weak back. What should I do?

A: Something went wrong, please try reloading the conversation

I couldn't get ChatGPT to answer despite multiple tries. It answers simpler questions just fine

I've tried a bunch of riddles, and it often cops out and refuses to try, but you can nudge it by telling it is a riddle and trying to nudge it. Sometimes you need to lead it a bit the way you might with a human who can't find a strategy to work their way through.

This doesn't seem too different from dealing with people to me.

It might not be all that smart, but if you try leading it through riddles step by step it's if anything harder to dismiss its ability to reason and understand - spitting out an answer wholesale might just demonstrate recall. Being able to talk through the steps, and getting it when you suggest it answer a sub-problem first, is of anything more impressive to me.

Where’s the article about this example ?

It doesn't show it doesn't understand. It shows that if it is capable of understanding, it's reasoning skills in some areas lag its language, and that is confusing to us. Plenty of people would fail to answer that riddle without prompting, but they'd likely answer in simpler language, and so we would evaluate them accordingly.

I asked it a bunch of riddles yesterday, and it failed to get many at first, until I led it through parts of the reasoning the way you'd lead a child.

E.g "you leave your camp, go 3 miles south, 3 miles east, 3 miles north, and find a bear in your tent. What colour is the bear?" At first it said it couldn't tell. I then asked it where the tent must be for the description to be true, and it correctly answered the North pole. I then asked what colour the bear must then be, and it correctly got the answer.

If I had to lead a child through it like that, my reaction would not be that the child can't understand. But when ChatGPT responds in formal, adult language it makes it hard to give it the same benefit of the doubt.

I don't know how much we can say that it actually understand, but I also think it's too easy to dismiss it like that. It is also too easy to assume based on other examples that it is "smarter" than it is.

Accurately evaluating understanding is hard especially because it's ability to solve problems is "lopsided" - it's unusual to be this dumb while at the same time being so eloquent and "well read".

I asked ChatGPT this same question and it said:

> Mike is the fourth child.

Is it training itself continuously?

Reply on news.ycombinator.com