Hacker News Reader

Comment by johnfn | Hacker News Reader

I don't think I buy this argument. ChatGPT seems to understand how to reason about a large multiplication the same that a 6 or 7 year old might, and I would expect a 6 or 7 year old to make similarly large errors. No one claims that 6 or 7 year olds are unable to reason.

View on HN

I took the opportunity of seasonal proximity to family and tried this question on a small selection of 6--8 year olds, and I received one correct response (accompanied by proudly showing the correct application of long multiplication on a big piece of paper) along with two "I don't know, what is it?"

None invented their own algorithm and confidently claimed it to be the way.

Interesting, but I wouldn't draw the conclusion that children don't sometimes make up plausible sounding but totally wrong explanations for things because they definitely do. Anyone with children knows this.

Though they are definitely better at saying I don't know than ChatGPT et al, which have basically been trained to never admit they don't know and always bullshit instead.

Yeah, in the original gpt-3 paper one of the more interesting bits was that it made similar off by one errors a human would make when doing arithmetic (and they controlled for memorized test data).

Seems pretty clear that ChatGPT is parroting a primary school multiplication course here, not actually explaining anything it understands.

> seems to understand how to reason about a large multiplication

How so?

It is a very impressive accomplishment what a large language model can do. It can piece together coherent text from a Google-sized corpus. But I don't thinking describing the process as "reason" is a useful description.

For a subject as purely logical as AI is, it sure seems to draw a in a lot of wishful thinking.

By that logic a parrot should be doing calculus within 10-20 years of its first word. Unlikely...

It's a fallacy to anthropomorphize a pile of linear algebra, relate it to a scale of human development, and extrapolate that AI is on a similar trajectory of progress/potential.

Relevant xkcd: https://xkcd.com/605/

> By that logic a parrot should be doing calculus within 10-20 years of its first word. Unlikely...

If a parrot's brain grew exponentially every year for 20 years, then I might expect it to. :)

> It's a fallacy to anthropomorphize a pile of linear algebra, relate it to a scale of human development, and extrapolate that AI is on a similar trajectory of progress/potential.

I don't believe I said any of those things.

By that logic a parrot should be doing calculus within 10-20 years of its first word. Unlikely...

I don't know about parrots, but if you could take a corvid and grant it the same rate of improvement in brain size and training input that these models are receiving, I wouldn't be the least bit surprised to see it doing calculus within 10-20 years.

Size isn't everything. It's not about the brain growing - it's about the teaching methods. Right now, I think it's still an open question whether current architectures and objectives are for sure capable of reasoning/AGI/etc or not. If they're not, then scale itself isn't the answer.

> ChatGPT seems to understand how to reason about a large multiplication the same that a 6 or 7 year old might

This is purely ignorant magical thinking.

You're ignoring the very real facts of how ChatGPT works and fabricating a fantasy that you prefer to believe.

Hardly. Go ask a 6 or 7 year old to write down their "reasoning" at multiplying 203948029384029384 by 7834928734982374982374 and, assuming you can even get them to do it, I argue that the words they write will be very similar to what ChatGPT wrote. Perhaps the vocabulary will be less refined, but the essence will be similar.

In any case, the point is that it's an objective thing anyone can observe: compare the two outputs. See if they differ.

> This is purely ignorant magical thinking.

I counter that you ascribe "magical" thinking to the human condition.

A 7 year old would know that you would want to use a calculator, and indeed, would be capable of multiplying those numbers with a calculator.

But by that standard, I could just ask GPT to write code to multiply the numbers. It's not even trying you be an embodied agent with a hand able to type onto a calculator.

GPT is remarkable to me not for its failings, but because of how far it can get when it's only trying to win the game of "predict what token comes next".

To be more accurate, the 7 year old will tell you they can't do it and suggest the use of a calculator.

GPT simply gives you a complete hallucination. This can be gotten around imperfectly with prompt engineering to get GPT to admit it doesn't know. But still, hallucinations are pretty dangerous failure modes in many applications, and definitely something we don't want in a future iteration. I don't know how much of a change it is to create a GPT without hallucinations, but I suspect that it is non-trival.

Reply on news.ycombinator.com