[erikstarck]

You guys are funny. An AI generates a game, comes up with the rules, writes the code and designs the web page for it. Your reactions:

- Bah, it's not very fun.

- It's been done before.

- It took too long to make.

Seriously. Let me repeat that. An AI generates a game. It comes up with the rules for the game. It even writes the code and designs the web page for it!

Come on! This is amazing!

[tempestn]

The "it's been done before" one is pretty relevant. It means the model didn't actually generate the game, but likely pulled it more or less straight out of its training data. It's still very cool that you can ask it for something and it can basically mine the entire (2021) internet for it, but it's not the same as being able to create something really new.

I've noticed the same thing testing it on various coding questions. It's extremely good at problems that have solutions online. And given stackoverflow, that's a lot of problems. If you manage to hit it with something that it hasn't seen before though, even if it's conceptually very straightforward, it tends to just generate a mix of boilerplate and nonsense.

[shynrou]

Exactly. When the first news came out about it's ability to "understand" code, find bugs and improve uppon it, I tested it with some snippets of mine. It just gave boilerplate best practices you find on 100 of blogs, but was not able to make meaningful contribution. It claimed to have introduced a feature while only having found another way to write the same snippet. On other things it straight up invented variables & functions that didn't exist.

As long as the task is in it's training set, it can give you a decent answer, but it can't code it just mimics doing so...

[TrapLord_Rhodo]

ChatGPT would be so amazing as a pair programmer if it didn't invent functions.

It's perfect for what is a python function for doing X. But it's honestly 50/50 whether that function even exists.

[xxs]

There are tons of examples from the training set that are awful. Most people will just eat them.

[muskmusk]

>The "it's been done before" one is pretty relevant.

But is it? 99.999% of software development has been done before. Even if you do something that is legitimately new (like creating a chatbot that can generate code on demand). Then your solution will still contain more than 99% code that is just a repeat of things that have already been done.

[Kiro]

That's not my experience at all. Copilot consistently creates implementations that are very specific to my app and manages to understand the context and problem surface spanning many files. It's not just getting a standard problem and pulls an answer from Stack Overflow.

[tempestn]

Given Copilot's specialization for this task, I can imagine it being better at extrapolating from your own code. I haven't used it myself yet, so can't speak directly to its effectiveness, but I would imagine it would be good at automating much of the drudge work of coding, but similar to ChatGPT as far as coming up with novel solutions to problems. Which again, isn't to say it's not potentially a very useful tool!

[fennecfoxy]

Even if the rules were inspired by some text that's on the internet rather than a genuine invention (we'll never actually know, we're all just speculating): it hasn't "pulled it out of its training data".

To be asked in plain, simple (ish) English to invent a game, produce code for it and then style it etc and the few other bits the author asked for _is_ impressive.

Why are we asking for so much? Remember the chatbots of the mid-2000s? Eliza etc? They were impressive for the time but GPT represents a _huge_ improvement in this stuff. Of course it's not perfect, but it's an exhilarating jump in capabilities.

[tempestn]

I definitely don't disagree that the progress has been incredible, and that GPT shows massive potential.

[jonny_eh]

One could even argue it's a fancy UI that steals content from stackoverflow.

[wildmanx]

> An AI generates a game.

But it hasn't! This is just another step in the BS storm coming out of the latest AI hype. The language model has reproduced something that has existed before and was likely part of its training data. That's cool, but it's far from what's being claimed here.

We really need to get better at fact checking this stuff. And with "this stuff" I mean the output of LLMs and other AI frameworks as well as the claims about it. And with "we" I mean society as a whole and our industry in particular. Let's keep the hype in the drawer. The general population can be hyped up about sth, but we should know better, so instead of joining the hype, let's keep a cool head and educate people about what this is and what it isn't.

[pharmakom]

yeah its more akin to using Google to find a game on github and copy pasting it.

[ahsood]

The second point on your reactions “It’s been done before” is very crucial.

That defeats the point of your argument that “AI generates a game. It comes up with rules for the game”.

No, it doesn’t. It plagiraised the game and pretended to come up with it. It just used a random puzzle game that it had on its training set.

It’s like asking it to write a poem and getting the same exact poem from a random google search. It didn’t come up with it. It just copied it. It’s not as amazing as you say it is.

Also if you look I comments you can see that it’s not even just one game. There are several games that are exactly like that. Which means more probability of having it in a training set.

[Kiro]

> It’s like asking it to write a poem and getting the same exact poem from a random google search.

No, it's not. A better comparison would be a poem that feels the same as an existing one and using the same prose but with its own words. Or any musical plagiarism dispute where the song is clearly different but similar enough that it needs to be decided by court. ChatGPT is not just copypasting a puzzle game here.

[fennecfoxy]

Ha ha, and what do you think us humans are doing?

Only a small percentage of our output is unique/original, otherwise we live to produce very similar output to what we've had as input.

Common phrases, hell all languages are examples of this. Mimicry of behaviours; it's literally the learning process that evolution gave us that puts us (and other animals) above much simpler creatures.

GPT isn't perfect, but it's like a dog observing that you fetch the newspaper every morning and then it starts fetching it itself for you, then the neighbour is like "well, it's not _originaaal_".

[bottled_poe]

It remains to be seen whether the lawyers agree with you.

[Kiro]

I'm not saying it's not plagiarism, just that it's not a copypaste.

[bob1029]

It is amazing.

I worry that a lot of otherwise brilliant developers are going to get blindsided by this stuff.

The current models are impressive in strong, quantifiable ways. They are only going to become more powerful from this point.

Consider the current state of affairs: ChatGPT supports a 4K context size. Leaked foundry pricing indicates models that can handle 32K context size. 32K tokens is enough for your entire brand manual or several days worth of call center transcripts. Many products could have the most important parts of their codebase completely loaded into just the prompt.

I would say you should at least try the OpenAI playground (or equivalent technology) to understand what is possible right now. I had no clue where we were at until ~3 weeks ago. I wouldn't wait until 2024 on this one anymore.

[olalonde]

Agreed. LLMs are on par with the invention web or the smartphone in terms of how much impact they'll have (possibly more). It's weird to see so many HNers being so dismissive of them. I've been using ChatGPT daily (mostly to ask programming related questions) and it's like having a new super power.

[moring]

You'll have to add one reaction to that list: "it didn't generate the game because something like that was in the training data".

Nevermind that this perfectly describes 90% of software development.

I'm actually wondering to what extent these responses are fueled by fear of being replaced by AI.

[beepbooptheory]

I know its getting popular, but I really doubt 90% of software development is generated from trained models...

I don't understand what's so at stake with this that you feel like people are afraid. It's fun and amazing it can spit out stuff like this, and if you are a good developer experimenting with this stuff you already know its inarguably a novel and useful utility, if still limited in some ways.

But where is the fire? Why does everything got to devolve into one vague culture war or another? Shouldn't you welcome good faith critique? If only for the fact that these things can still be improved, and how can you hope to improve them if you smother and dismiss every suggestion that these models might be less than perfect.

[moring]

I think you are fighting a strawman here. These models are far from perfect, and critique is certainly needed to improve them.

That being said, I don't think saying that the model outputs information from its training set adds a lot to the discussion because -- and that's my point -- the same is true for human software developers most of the time (yes, the 90% is a made up number). This isn't meant at all to criticize the skills of the developers, but rather point out that most of our work is just much less interesting than we'd like it to be, and could be automated.

Also, I don't think developers will be replaced by AI, just like they were not replaced by code generators, build scripts, IDE auto-complete, IDE rename all usages, and so on. What might happen is that they will no longer have to write mind-numbing boilerplate code, which IMHO is a good thing.

[olalonde]

HN went from "AGI is impossible" to "anything below AGI is worthless" real quick.

[postalrat]

The only thing that needs to change for HN to care is for the AI to start writing this stuff in Rust.

[codebym]

It is fairly typical of HN to err on the side of cynicism.

Correct me if I'm wrong, but ChatGPT is a very fancy auto-complete function. It's has no ability to create from scratch, just the ability to recompile and recontextualise any of the many existing pieces it has in its library.

It's unlikey that this game or its rules are truely original, ChatGPT will have just plucked it from the library, perhaps given it a new name.

[postalrat]

Can you come up with a question to ask it that would prove or disprove your theory?

[hgsgm]

What is "scratch"?

"Great artists steal".

Art is defined by remixing the life experience of the author.

ChatGPT's ability to create art is only limited by it's input (text corpus) while humans have images, sound, smell, touch, etc.

[beepbooptheory]

Each day I walk through my war torn and impoverished little village, looking for food and water. I keep in my pocket a small postcard of a beautiful tropical beach I've never been to. When I get home, I use the bits of supplies I've found to make my little paintings or write my little stories.

The GPT version of me can only remix the world I am already in, so this version mostly paints dark landscapes and violent imagery, however much I prompt myself to draw something pretty, it always comes out looking a little macabre.

The regular old person me, on the other side, is plagued by the human afflictions of desire and fantasy. This version of me can only paint the beach on my postcard, but because of my desperation, focus, and need, these paintings become larger and more fantastical than I thought I could imagine, but in that, they provide me and my loved ones comfort and escape and novelty.

The human version of me makes statistically improbable things, but, to me, is still a plausible human, and the one I'd rather be at least.

All just to say, maybe there is a more qualitative difference here than you think.

[overspeed]

People draw an arbitrary line of difference in the way we treat AI programs and human outputs.

A human imagining orcs and one horned horses has 'fantastical, larger than life' imagination but AI generation drawing people with strange hands is 'incorrect'.

These are not one to one examples but the point stands that with enough suspension of belief, people are more likely to take on human creations at face value than AI when they know the source.

[beepbooptheory]

Sure I think I agree. My point is just you can imagine the post-apocalyptic artist with a human brain painting the beach or just painting the dark landscapes; but we can only imagine the GPT brain painting the dark landscapes, in so far as that is majority of its day to day dataset, thus the statistically likely output.

This suggests a qualitative difference between the two when it comes to "creating" or "generating" that feels far from trivial, even if you want to say the AI can make "good" things, whatever that means to you.

[isaacremuant]

I don't see how your comment contributes to the discussion. It seems aimed at shutting it down only allowing praise.

The scope of the creation and whether it actually produced something novel is quite important to the discussion and part of the claim (although the author is very open to be proven wrong, in the article).

Your claim n the second to last paragraph is false. That's relevant. This is HN.

[matsemann]

Did it come up with the rules, or did it just rehash an explanation for an existing game from its training data?

[speedgoose]

The model weights aren’t that big, even compressed with the best algorithm possible.

For such an obscure game I would guess it came up with the rules. It’s difficult to prove though.

[PeterisP]

What we're effectively seeing is that when someone demonstrates a talking horse, some people will complain that the horse speaks with a horrible accent and uses impolite language.

[xxs]

How do you relate: "An AI generates a game", and then "It's been done before." Obviously it didn't generate a game but it copied, which is not "amazing".

On a flip note: the game is fun enough.

[sinuhe69]

Yes, in some sense it’s. But if the game already exists and the AI just parroted it, it’s significantly less impressive, isn’t it?