Hacker News Reader

Comment by dvt | Hacker News Reader

Totally agree with this. Thinking of doing a deep dive to see if there's any actual real uses cases of these toys, but I honestly don't really think so. Art generation is probably the most interesting application, but just having a chatbot to talk to seems like a weird value prop.

View on HN

I'd hope so. If not then you probably shouldn't be in charge of delivering a summary.

I used ChatGPT to give me a list of test areas for a certain type of scenario. And it actually pointed out one that I had inadvertently missed. Now if it said something off the rails I would've known that too.

I've probably had 100 conversations with ChatGPT and I can only recall one blatent lie that it told, which I think was P != NP. Which maybe wasn't a lie, but it didn't have sufficient evidence to make that claim.

I've actually so far gotten more accurate information from ChatGPT than I did in Wikipedia's earlier days -- where I discovered that Matt Damon dropped out of Harvard because he was too stupid to make it there.

Here's ChatGPT being wrong about me lol[1]. First, it confuses me with another person (also wrote a Meteor book), straight up telling me that I'm the one that's confused. It also makes up a book which doesn't exist. Um, ok. Then, it tells me I'm based in Romania, which I am not. I was born there, but I'm based in Los Angeles (every one of my bios has this info).

[1] https://imgur.com/a/kaYyRqW

Where does it say that you're the one who is confused?

Third prompt

Oh, the "not David..." part I assume. I thought it was just wrong -- not saying you were confused. I think sometimes we pull in too much "human intention" into AI chatbots.

I'm not the person that posted the prompt, I just read it and easily identified what they were referring to.

The fact that you don't, or have managed to pass that as anthropomorphizing is curious. I'll have to think about that, but on its face seems strange alongside your comment about if someone can't find a legitimate use case for AI that they shouldn't be a person who summarizes it.

What's a blatant lie, really? ChatGPT has routinely told me about APIs and books that don't exist. And just the other day it told me that paper sticks to my hand because the roughness of my skin creates suction, which I'm almost certain is wrong (I believe it's due to weak chemical bonding, but who knows).

I know I'm being pedantic here, but a "lie" is a deliberate thing. It requires intent to deceive. LLMs have no intent, and therefore they can't lie.

They can certainly be wrong, though.

Several people I know are using it for sales and customer service writing. As a user, that's not really what I want, but they are finding it to be a useful tool in generating text for them to review and modify.

Sounds like another race to the bottom to me, as with email spam and SEO spam.

And the reason ai art is interesting is precisely because art is about mistakes and hallucination and wierdness, and unexpected connections.

So the flaws of chat AI currently are exactly what make art great. And art came before science, so perhaps scientific thinking is much harder on an evolutionarily timescale than we think.

It’s unfortunate that art generation, something humans genuinely enjoy doing and can be a source of personal fulfillment, is capable of being replaced, but customer service and answering stupid questions are apparently hard problems.

Answering stupid questions has already been thoroughly solved, and you don't need an AI trained on 500 GTX4090s to index a stupid question database. The problem with customer service is that the questions that aren't stupid are genuinely long-tail edge case issues, where an AI is very unlikely to solve the problem.

For sure. There's a lot of bad customer service out there that is functionally equivalent to modern AI in that it's poor pattern recognition plus no real understanding plus cobbling output together from existing text of varying relevance.

But sadly, that may be evidence that there's a market here. If Amazon can cut support costs by 80% and only moderately worsen their already bad quality, some execs would consider it a success, not a failure.

Reply on news.ycombinator.com