Things created with GPT-4 | Hacker News Reader

Things created with GPT-4

twitter.com

46 pointsvinnyglennon67 comments

Clicked on the second one and was immediately presented with a course to teach me how to use this fancy new AI-powered technology to revolutionize my business! Puts a certain taste in my mouth.

It appears many of the same people who were hyping crypto and meme stocks a year ago are hyping AI startups/courses now. The sales manuscript is basically identical with lots of manufactured FOMO.

It's a shame because the technology itself is probably reasonably useful. It's really hard to tell though with all the bullshit surrounding it.

you just need to buy my course on AI course BS detection. it's only $995. but if you watch this short free webinar on 5 simple ways to tell as course is BS you can get it for only 3 payments of $499. But you have to watch until the end.

That's all twitter is these days. People writing content to drive engagement with some silly hook to sell whatever it is they're selling.

It's so bad that I feel nauseous each time I see a tweet with the "finger pointing down" emoji indicating "thread below".

I scrolled down. I shouldn't have scrolled down.

A lot of click-bait and "coaching" is spread under the term "ChatGPT". Reminds me of the 238th iteration of a calculator for the AppStore years ago just to have an app at the AppStore.

Marketers learned to sell shovels. Understanding their business model is important.

GPT-4 is barely 3 weeks old but have already created a business opportunity to run a paid newsletter about GPT-4!

This current moment is super impressive, but the output of the models is still mediocre. I worry that over the next 12-24 months, we'll be overwhelmed with autogenerated content (docs, apps, games, news) that gets a B- or C.

Once the novelty wears off, no one wants to play your [terrible] autogenerated asteroid clone. There are a lot of startups that are two LLMs in a trench coat, and they will not be able to improve faster than people's expectations increase.

This is my worry as well. Google results are already full of content farm/AI SEO junk, and I don't understand people who see that and think "let's do that, but more".

There's already more good content than I could possibly consume in my lifetime, and all this automatically generated stuff is just going to make it harder to find.

> they will not be able to improve faster than people's expectations increase.

I've never read a better explanation for why "we will never have AGI." As the objective quality of AI output increases, we simply value it less.

Quality is objective, value is not.

Can Midjourney copy DeviantArt/ArtStation to levels far beyond mine? Yes.

Does the market value DeviantArt/Artstation output to any level? No.

Your BAD autogenerated asteroid clone. "Give it time" people will say, but how does one prompt ChatGPT to make a game fun, or a story actually interesting?

^ Yes, edited my post! I've spent HOURS tweaking prompts and going back and forth with GPT3–4 to get more creative and interesting output, but it's not there yet. This current generation will just never give you legitimately good song lyrics, let alone a game with an interesting and compelling mechanic.

Creativity is making unexpected connections, applying a twist that doesn’t logically follow from what came before. This is the opposite of what these models are designed to do, which is to give the most likely next word.

On the other hand, raising the temperature should allow unlikely sequences to be output, provided the model is trained on enough unlikely sequences as input. Also, most creative work does not require much creativity, so just a little human guidance may go a long way.

Having played around with the same scenario, I’ve come to the naive conclusion that the lack of creative output is because the current generation of GPT is only able to regurgitate and not create. This makes everything feel very derivative.

On the other hand, if you act as the creative mind and use GPT to accelerate your output, it becomes more useful.

Am I off base, or is "create" sort of amorphous in this context? It's already creating novel sentences and paragraphs. What is difference between that and real creation?

Personally it seems to me that if you feed a statistical model of a ton of input it's going to feed back the average of that input, leading to uninspiring phrases. Regarding game creation, we'd need to tell it about game mechanics, and I have no idea how it'd determine what's interesting or how it would innovate.

I think that's it—these models are definitely capable of creating output not seen in the input, which we would define as human "creativity," and they can even combine concepts in novel ways, which is loosely what we call "innovation." But they are quite biased to create plausible output, which severely restricts their ability to create something interesting.

You can dial up the temperature, but this quickly pushes the output from "bland" to "totally unhinged."

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 decrease the average quality of games made.

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.

I agree. The use cases are still far too limited to be of use at the scale beyond Canva or something. They will pay to automate the creation of social media graphics and emails, because nobody wants to work on those. But then what?

The imagined computers that looked so unrealistic in so many past science-fiction movies and TV shows no longer look... unrealistic.

Those imagined computers now look more like upgrades to current technology.

Imagine an AI model 10x, 100x, 1000x larger than GPT4.

That feels... surreal to me.

> 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 the neural networks become deeper.

We'll see.

Yes you can cherry pick things that work very well, in the meantime it still can't rhyme in French, which is 12 years old type of intelligence

It's wrong > 50% of the time for anything semi complex and semi niche, that's not AGI, no one in the field of AI argue it's AGI, only people on twitter and reddit with absolutely no background in tech and/or AI argue it's AGI

It's a very very very good chatbot, built with very very impressive technology and amount of data, which is already quite impressive and definitely a big milestone, it's not AGI, not even close

It's like saying the leather auger of my swiss army knife is a step towards an aerospace grade titanium cutting CNC machine

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 that could pass as AGI, but the current models are already better than just "parroting". And even if progress stales for the next 10 or 15 years like it has happened frequently with AI, everything that it has learned to do in the last couple years can be fine-tuned and made a lot more efficient, which I expect will have a big impact already.

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".

see it's already passed the Turing test:

https://sports.yahoo.com/24-most-hilariously-bad-spelling-19...

> 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.

I've been noticing this pattern on HN where someone analogizes current generation LLMs to parrots while simultaneously claiming to know the future of LLMs without any justification. Am I a "techbro" because I don't believe that we know much about what the future of LLMs has in store for us, or because I find the parrot analogy to be utterly specious?

You're free to believe whatever you want but then we're deep into religious territory, which is fine in itself but doesn't really matter here

>not even their creators argue that

Yes, their creators take the far more defensible and scientific position of saying "we don't know".

Not at all

Same Altman (openAI ceo) :

> The GPT-4 rumor mill is a ridiculous thing. I don’t know where it all comes from, people are begging to be disappointed and they will be. The hype is just like... We don’t have an actual AGI

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/status/15948895625260277...

[2] https://arxiv.org/abs/2303.12712

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.

We are hearing about new, improved parrots on an almost daily basis.

And no matter how many W16 you put in a bugatti veyron it will never be a space shuttle

I don't get how people, especially here of all places, fall for such things

Are you not familiar with the process of software engineering?

Replacing parts until you get a space shuttle is exactly how systems are evolved.

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.

I don't know why, but that "use your apple watch to code" seems like hell to me. Presumably because I know it's going to be a lot of stuttering and saying "ignore that". Interesting demo, but I don't think I will ever want to talk to a computer like that.

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 documents like building permits, claims

For industries that have a lot of text format and can deal with some degree of statistical error as long as averages are better it's massive

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

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-pico.onrender.com/amidst-assurance image to meme generator: https://backend-pico.onrender.com/wagon-woot

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.

How are people using this? Is there a public beta or do they have special API keys?

To avoid spam, regarding the comment above - it all felt "surreal" to me until I realized how wrong ChatGPT is most of the time. Not only that, but I am writing in Python now (a language I don't use) and realize that while ChatGPT can give hints, it's definitely nowhere near replacing mid devs, let alone seniors and architects and I can't use it unless I know exactly what I am asking for. I bet all of the examples in the parent are highly, highly nitpicked and have 1000s of similar apps to copy. E.g. the movie app is my go to practice assignment for new devs to learn basic UI and APIs.

More of my ranting on AI failures here (and it's basic programming tasks in one lang): https://news.ycombinator.com/item?id=34640699

There's a waitlist: https://openai.com/waitlist/gpt-4-api

ChatGPT Plus is also using GPT-4 as its backend

chat.openai.com, create an account and upgrade it to plus ($20/mo). There’s currently limits on how many messages you can send. There’s also an API.

Thanks, already messing with the free version, didn't know it was using GPT-4 - will try. P.S. Thanks to both of you, just replying to one comment to avoid spam.

The free version is still on 3.5.

I still use it, it’s faster and I can’t use the plus version that my company pays for for personal stuff.

Does Plus give API access? Last time I looked it didn't, I think the waitlist is still the only option.

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 everything up.

I've also found it's good at common problems and sometimes spotting the cause for syntax errors, but less common problems it might not know.

Hope that helps.

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

Using it to "peer review" research papers to look for flaws in statistical methods used, logical flaws in conclusions etc would be fantastic. Of course it would need to be verified, but maybe it could be a turning of the tides for all the bad "research" done out there.

I’d wager things are likely to not get much better than these examples for several years, maybe a decade.

How do they make sure these "financial advisor" or "research abstract summarisers" are hallucination free ? So far it still is deeply in the "cool gadget" category for me, 6 months after the chat GPT revolution and not much to see

Don't get me wrong, it's really cool technology, but calling the first example a "game" is... Grossly overstating what it is. However, given it was made by someone who's not a [game] developer, it's pretty cool nonetheless.

No, it's not just a game, it's a (complex) game.

But like you, I agree, that's cool. That's like a pretty good start if you wanted to make an actual game. You have movement, physics, and collision detection. There's clearly something to build off of.

So annoyed at the people with GPT4 API access. I applied immediately, and have now built a handful of applications with 3.5 that would work so much better with changing a single model string once I get access.

I'll get excited when FOSS community catch up, and I'll be able to run GTP-4 like models locally instead of calling some API.

seems like GPT systems are the most useful when you can quickly verify the correctness of the output (ex code generation) but else quickly becomes dubious (ex how do you know a summary is correct or useful?)

Reply on news.ycombinator.com