Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Things created with GPT-4

twitter.com

Post reply on HN
Print view View original

61–69 of 69 posts

Re: Things created with GPT-4

#61

Post by ketzu »

cahoot_bird wrote: post #42
ketzu wrote: post #28

It is so interesting to me seeing how everyone utilizes those large language models. I often wonder how they do it. I tried to get something usable out of ChatGPT now and then, but the code is littered with misunderstandings of the data format and errors and more problems. Maybe I am bad at explaining what I want to the system?

It can make mistakes and hallucinate. I've found GPT-4 to be slightly better than earlier versions. Ideally tell it one step at a time you need done, I've found that it struggles with multiple overlapping steps for some things but not others. It can make tic-tac-toe in one prompt in HTML5 and Javascript. But maybe if you ask it to make it rain in the background or something like that in the same prompt it might mess…

Thank you for the advice. I would like to make use of those tools properly, so I'll give it a try.

Re: Things created with GPT-4

#62

Post by fiso64 »

lm28469 wrote: post #48
letmevoteplease wrote: post #41

Earlier quoted context omitted.

He was responding to viral tweets[1] that said GPT-4 had "100 trillion parameters" and was "500x more powerful than GPT-3," which is indeed absurd hype. On the other hand, two weeks ago a team of Microsoft researchers put out a paper claiming that GPT-4 "could reasonably be viewed as an early (yet still incomplete) version of an artificial general intelligence (AGI) system." [1] https://twitter.com/andrewsteinwold/st…

You're missing the end of the quote, a bit further in the next sentence: > including the possible need for pursuing a new paradigm that moves beyond next-word prediction. So all in all "it's not AGI and we might need a paradigm shift to achieve it" A paradigm shift being: "a fundamental change in approach or underlying assumptions."

Exactly... In other words, they say that they don't know whether this will get us to AGI.

Re: Things created with GPT-4

#63

Post by abi »

senko wrote: post #34
abi wrote: post #24

I built a web app that lets you generate simple, shareable web apps just by prompting with GPT4. https://picoapps.xyz Here are some cool apps people have built with it: read aloud a wikipedia article: https://backend-pico.onrender.com/northeast-architect matrix animation: https://backend-pico.onrender.com/chair-northeast face distorter: https://backend-pico.onrender.com/drench-virtual simple excel: https://backend-pi…

This is awesome! I'm doing something similar but for backend. Is picoapps frontend-only for now or do you generate both the front and enough of a backend to eg save what the user creates?

Thanks :) Right now, all the code it generates is in a single index.html so no backend.

Would love to try/test your backend thing.

Re: Things created with GPT-4

#64

Post by marcosdumay »

Name_Chawps wrote: post #21
lm28469 wrote: post #8

Earlier quoted context omitted.

> Those imagined computers now look more like upgrades to current technology. I still don't get that argument. You can train a parrot for a week and he'll talk a few words, training it for a year won't make him become Shakespeare, LLMs will never ever automagically evolve into these AGI tier computers, not even their creators argue that, no one besides techbros argue that in fact.

I can, right now, have GPT 4 write me sonnets on any topic. We are already beyond mere parroting. People say "it's just parroting what it's heard"! But did GPT4's training data really include a sonnet about Minecraft creepers? So they're not parroting. They're parroting PLUS something else. Plus some kind of analysis and transformation. Your claim is that this transformation is superficial and won't become deeper as…

> But did GPT4's training data really include a sonnet about Minecraft creepers?

The entire internet is one hell of a large knowledge base.

It does surely look like it's doing some transformation. But I don't know how to get certainty on this.

Re: Things created with GPT-4

#65

Post by marcosdumay »

jmull wrote: post #32

On one hand this is stunning. On the other hand, almost everything shown here either generic or crap (or both). I'm happy that chatgpt writes boilerplate code for me so that I don't have to but I'm not seeing the revolution yet.

> I'm happy that chatgpt writes boilerplate code for me

I'm really concerned that Copilot and similar stuff will make boilerplate-heavy code even more popular than it is now.

Re: Things created with GPT-4

#66

Post by dandellion »

lm28469 wrote: post #52
dandellion wrote: post #50

Earlier quoted context omitted.

I can't rhyme in French either. A lot of these comments just remind me of that Alan Kay quote: “Artificial intelligence is what we don’t know how to do yet.” In the past year we've got models that can draw, write poetry and code better than a significant percentage of people, and now the goal posts are moving so fast they're even hard to see. I'd probably agree that we're still a few breakthroughs away from something…

I have a really really really hard time thinking we're even remotely close to AGI when I see stuff like: https://imgur.com/eukW66W You must have a really convoluted definition of AGI to come to the conclusion gpt is AGI when it proves times and times again it quite literally have no understanding of what it writes

I said that "I think we are a few breakthroughs away from something could pass as AGI". That means, first, that I don't think we're there and so I don't consider chatgpt AGI, and second that even something that "passes" as AGI at first glance might not necessarily be so under close inspection.

But at the same time, I won't call what chatgpt is doing parroting. It's not just repeating input data as-is, at least it's capable of applying some transformations on that data. Most of the time the transformation is useful, occasionally it's pretty borked. Parroting to me would be just repeating the training data without much transformation.

> it quite literally have no understanding of what it writes

Even though I DON'T think gpt AGI, I don't consider that a deciding factor. Humans can also show no understanding of what we write or read, and I consider us "GI".

Re: Things created with GPT-4

#67

Post by anukin »

kozikow wrote: post #39

I work in property insurtech and had lots of fun. - Customer support chatbot. Convert knowledge base and other materials to q&a form with gpt4. Ada2 embeddings, search for questions in repository with faiss, ask gpt4 to answer. Works a lot better than standard "search with embeddings in KnowledgeBase". - Parsing out and standardizing insurance descriptions, Excels, schedules of values - Parsing out and standardizing…

How does this work? Are there some tutorials etc to learn how to do this?

Re: Things created with GPT-4

#68

Post by kozikow »

anukin wrote: post #67
kozikow wrote: post #39

I work in property insurtech and had lots of fun. - Customer support chatbot. Convert knowledge base and other materials to q&a form with gpt4. Ada2 embeddings, search for questions in repository with faiss, ask gpt4 to answer. Works a lot better than standard "search with embeddings in KnowledgeBase". - Parsing out and standardizing insurance descriptions, Excels, schedules of values - Parsing out and standardizing…

How does this work? Are there some tutorials etc to learn how to do this?

https://www.youtube.com/watch?v=ih9PBGVVOO4 is quite close to what I did

Re: Things created with GPT-4

#69

Post by flangola7 »

SketchySeaBeast wrote: post #59
wongarsu wrote: post #53

Earlier quoted context omitted.

There are a lot of bad hobby games out there already. Telling human programmers "make this fun" doesn't reliably work, because being able to design good games is a skill with very little overlap with programming. If anything it has overlap with UX design, and we all know the stereotypes about programmer-built UIs. It's certainly possible to learn both skills, but taking away the need to be good at programming won't d…

I'm not saying humans consistently make good games, but we do know that humans can make good games. The question is can ChatGPT. If it's supposed to reach a the ranks of a general intelligence, it should be able to. I'm not sure how it can.

Ask it to list verbosely step by step essential elements of what makes a game fun, then iterate upon each one of those.
Post reply on HN