Hacker News Reader

Comment by PaulHoule | Hacker News Reader

People have just fallen out of love with it and are realizing how it really isn't that good.

Sorry "prompt engineers" but papers on arXiv show that when you give it fairly sampled problems it struggles to get the right answer more than 70-80% of the time. When you are under its spell you will keep making excuses but when you are looking at it objectively you'll realize the emperor is naked.

If you give it very conventional problems it seems to do better than that because it is a mass of biases and shortcuts and of course it will sentence Tyrone to life in prison because it's a running gag that "Tyrone is a thug"... That's how neurotypicals think and no wonder why many of them think ChatGPT is so smart... It mirrors them perfectly.

View on HN

>When you are under its spell you will keep making excuses but when you are looking at it objectively you'll realize the emperor is naked.

It's not perfect, but you have to admit the emperor is at least wearing a thong. It is the second most intelligent thing on the planet at creating text, even with its flaws. Putting that accomplishment in league with naked emperors is astoundingly biased.

Yeah, maybe it is "wearing a thong" since it's developers were so concerned about social acceptability.

I would say though that it is no mean feat for the Emperor to get away with being naked in public, it takes power and the ability to wield it.

Similarly ChatGPT has a few competences that add up to people perceiving it is able, one of which is the ability to come up with plausible and satisfying answers whether they are right or wrong (often using shortcuts) and another one is getting people to engage with these answers.

By this, do you mean 'the second most intelligent thing' after humans?

A dog understands body language and subtle cues, participates in social hierarchy and community, has self-reflection, dreams, etc.

> you have to admit the emperor is at least wearing a thong

And a lot of people are getting turned on by it.

>It is the second most intelligent thing on the planet at creating text, even with its flaws.

This also puts it in last place.

I don’t know, I think there is certainly a lot of hype but it is still pretty damn good. I used it every day for a wide variety of coding tasks and more often than not it is correct. As in compiles, produces correct output for inputs, and mostly writes reasonable code. Has it made software dev obsolete like some maximalists said it would? definitely not, but it seems equally hyperbolic to say the emperor has no clothes, it is undeniably a useful tool to me.

"More often than not" is compatible with "correct 70% to 80% of the time".

Your previous comment was a bit confusingly worded, I, and presumably the commenter you just replied to, read it as "more than 70-80% of the time" it struggles; rather than your intended meaning of it struggling to do better than correct 70-80% of the time.

I used it to help me write an email begging for my job back and it was pretty helpful

Just ignore this fella, GPT-4 is massively useful. But it takes intelligence to get intelligence out of it.

I think this is our path - become good at working with AI, for jobs or if there are no jobs, to directly support ourselves. Supposing we will be able to use tools and AI to build things and make our own means.

This is our path yes. All paths end, eventually, but oh well.

I very much doubt the tweet in question is of someone who loved GPT-4 yesterday, and has "fallen out of love" with it starting today.

For someone so dismissive of "neurotypicals" you did the neurotypical thing and drop a hot take before clicking the link.

It's more that the balance in usefulness vs waste of time shifts into the other direction after a few times spending long time getting it right and just doing it yourself anyway.

I've spent 20 years explaining my code to a rubber duck and getting great use out of that duck. No one will ever convince me that it's less useful now when the duck actually has read the Internet and has useful suggestions and ideas back to share. Even if it makes shit up on an occasion.

I don't know. The value of rubber-ducking is that you have to break the problem down into very simple terms to explain it. It's the explaining it part that is magic. That value goes away (or is greatly diminished) if the rubber duck responds with anything other than a request for further explanation.

You have to break down a problem to GPT just like you would to a human, and if it misunderstands you or asks you, you'd have to elaborate just the same. It's modeled after us, after all. The rubber duck is a stand-in for a human. GPT also is. But just a vastly better one.

Of course.

The part I was responding to was this:

> when the duck actually has read the Internet and has useful suggestions and ideas back to share

I think that if it does that, it's legitimately less useful as a rubber duck. I'm not saying it isn't useful -- it's just not useful as a rubber duck anymore.

Similarly, but not quite, I often find browsing new submits on HN more useful than simply going through the front page. Just because, uh, front page is decided by statistics, that is, the community's neurotypicals

The tweet is an OpenAI employee who is responding to people who have fallen out of love.

I've seen transcripts of people interacting with ChatGPT who were obviously seduced by it and in a very giddy state, having so much fun because ChatGPT was playing an extended "game" with them that it didn't bother them at all that ChatGPT was spouting wrong answers.

A major complaint I've had about the social sphere is that I seem to get the same result if I am 20% right or 50% right or 80% right or 95% right or 99.8%, it is just exhausting and I can never be good enough and I'm frankly envious that people see more of a glimmer of light behind that thing's "eyes" than they do behind mine.

The core thing about neurotypicality isn't so much that they get the wrong answers but that they get the same answer whether it is right or wrong. For a long time I thought the basis of the "language instinct" is a derangement about reasoning with uncertainty that causes the grammar representation to collapse into a low-dimensioned subspace which is learnable with a limited amount of data. I wouldn't be surprised at all if other animals could beat us at rock-scissors-paper or poker if they could understand the rules of the game. The success of LLMs might give us some insight in this area although they are working with so much more data that Chomsky's old "poverty of the stimulus" argument might not apply.

Reply on news.ycombinator.com