Oh wow, this is just a straight ripoff of that last one isn’t it?
I feel like this is a recurring theme. Someone posts about one shotting something and then it’s revealed that you could also have one-shot it buy git cloning the repo it’s based off of
The real point of these posts is karma farming. Posting a good one shot is worth at least a hundred or so upvotes. People need to recognize this for what it is and ignore these kind of things.
My Belgian Tervuren and I have a basic herding title and about 4 years of herding experience.
The sheep movement is excellent. You could make it even more realistic by having them favor lusher areas and by having one occasionally bolt spastically (hard mode?)
A handler mode where you play as a human and shout commands at the dog could be cool too!
> My Belgian Tervuren and I have a basic herding title and about 4 years of herding experience.
Do you happen to have any videos of a Tervuren doing actual herding work or competing in herding trials?
I was apparently under the wrong impression that Tervurens are no longer bred for herding. My girlfriend just told me that there are still dedicated herding lines, but when I started looking into it, I could hardly find any footage. Compared to the thousands of videos of Border Collies working stock, there seems to be very little material available.
I'd also be interested to hear more about what your dogs are like. Are they from a specialized herding line?
Do you see them as being closer to the typical working Malinois - very intense, high-drive dogs with a strong prey focus - or more like a blend of the classic German Shepherd and Border Collie traits?
The latter type, at least in my experience, can make excellent everyday companions as long as their owners know what they're doing and provide enough mental and physical stimulation.
The reason I'm asking is that I've been considering getting another Malinois. I used to own a fourth-hand Belgian Shepherd who eventually turned into a reasonably good everyday companion, but it was a long and sometimes difficult road. There were also certain situations where he could never really be trusted.
He also had a habit of finishing arguments that my Pointer started in a very Malinois-like fashion, while my Pointer was generally too stupid(and proud) to learn anything from that experience.
Feel free to write a wall of text if you feel like it.
I think it’s impressive that an LLM can take you to a local maxima in one-shot.
But once you start maintaining it, improving it and fixing bugs, you’ll eventually need to rip it apart and put it back together again while understanding how it all works.
This is why I think the better approach isn’t to one-shot but to have the architecture in your head and build it up piece by piece, with the AI accelerating the code writing.
I’ve found it very easy to maintain, add features to and fix bugs in software I’ve written entirely with LLMs, and in languages and frameworks with which I’m unfamiliar. You just ask the LLM to explain the code and then work with it to come up with the fix.
How big are those projects.. I dont think this is good for your mental health or physicaly your brains health. Problem solving keeps your brain strong. The laziness in us is inclined to take shortcuts, don't do it. Its like driving your car 3 blocks instead of walking, your physical health will suffer.
Exactly this! People think the one shot gets them to 95% complete on an implementation of their vision. Issue is, it actually gets you the AIs vision adjacent to what you want, and coercing that into the actual implementation you want is now 95% of the work.
This is really no different from working with humans. A visionary founder has to spend tremendous effort to get their engineers to will a vision into existence. This will be the key skill with AI.
LLMs are good now at looking at existing project and suggesting big refactors for technical debt removal and new better architectures after the project grew organically for a while
I think this is true for projects beyond a certain complexity. I have 100% vibe coded projects with tens of thousands LOC, and haven't seen any real issues with fully automated maintenance. Will that approach work in every scenario, absolutely not, but the size and complexity of projects where it does is growing with each new model release.
Strongly agree, well said. The one-shot is sexy purely because that first demo is so impressive. Going from zero to working app in minutes.
Like you said, working and maintainable are very different things. One-shot hits a wall the moment you need to do anything non-trivial after the initial generation. Bug fixing is extremely hard, even with AI assistance. Same with feature additions. It's pretty much black box at this point on. AI that wrote it now goes in loops wasting tokens without being able to can't reliably fix it either, because it has no memory of the architectural decisions it made (or didn't make, for that matter) the first time round.
What I realized is that the failure here is the absence of a shared mental model between you and the code.
I'm a product designer with average front-end know-how, and a solid understanding on HTML/CSS and how the web works, coming from the era of hand-coding html/css files. After vibe-coding a few products early this year, purely to learn how AI works, how to design AI interaction patterns etc., I built something called Intent Model. (largely inspired by SDD / BDD.
Intent model is a structured, typed artifact (basically a JSON contract) that captures actors, entities, journeys, rules, and constraints before I write (or make the AI write) any code. It sits upstream of everything. Think of it like a condensed, strict distillation of your PRD / BRD / requiremnt doc.
When you hand the AI a well-defined intent file instead of a vague brief, this one-shot becomes structured and bound by rules. Now you're giving it an architecture and to conform to. You define (or make the AI define) the precise variable names, their types, lifecycle, user roles, responsibilities, business rules and constraints in the file. Every generated artifact can trace back to a decision you made deliberately, reviewd and signed-off.
In the design world, we already do this by using design tokens. We can tell the AI that it needs to strictly use design tokens and not use stray properties like a hex color value or raw values not defined in the token contract. This is easily auditable by AI as well.
The result is you can still move absurdly fast and still maintain the understanding, which the one-shot approach throws away. This way, you know why every piece exists because you defined the intent before the AI implementated it.
AI is the accelerant, and you're the architect. The intent is the blueprint you generate to guide/harness the AI.
The best part is, once you have an intent contract at the heart of your project, it becomes impossible to break things too, logically or experience-wise.
Curious enough, I tried the same prompt with Qwen3.6-27B.
One shot produced a game with no sheeps.
I had to told it to fix two bugs then.
Overall, the graphics and games seems good enough and better than most of the closed models that were shown. However, not surprisingly, falls short of Fable.
I've put the index.html and open code session here:
If you sit down and write that game by hand you will not only finish it in a week but also learn a lot of things along the way and perhaps even discover something about the game and you did not imagine. That is how programming works. It is a search problem.
Also this is a game has very simple mechanics I am sure you can generate as easily with Cursor or some other tools.
More deeply than programming being a search problem, programming is a means to an end.
If the end is a combination of education and product discovery, then yeah maybe, although those are also dimensions of personal productivity that can be amplified by leveraging AI tools.
If the end goals of programming is leveraging computer automation, then nobody actually cares how the automation infrastructure gets established, and the less distractions with low value implementation complexity, the better.
I don't use Cursor so I mentioned it as slightly impartial suggestion but my point is broader. I hear and have seen results from others using Composer 2.5 which is only available in Cursor.
“ can it build a game idea I've had for years, in a single shot?”
Do people do no research or introspection when they’ve had an “idea for years”? There are countless examples of this exact game. I played this on the Gameboy Advance! There’s like 50 of them on the App Store right now.
The standard “this almost certainly exists wholesale in the training data” applies, but I’m also interested in how you carry an idea for years and don’t notice this, or whether the “idea” here was actually “using this thing that’s been remade thousands of times as an AI benchmark”.
There’s nothing wrong with remaking an old classic formula, especially in game dev. It’s the describing it as “an idea I’ve had for years” that rings weird.
It's called Ostrich effect and we all do it. You enjoy toying with your own idea so much, that your brain shields you from the pain of finding out it already exists. Deep down you know it probably exists. It's harmless, unless there's other people's time and money involved.
> You enjoy toying with your own idea so much, that your brain shields you from the pain of finding out it already exists.
Doesn’t look like the author toyed with the idea at all, though, apart from having it in their head. Considering how they describe themselves (Check the About/Home page), if they had toyed with it at all they would have already built it.
I also don’t see why finding out it exists would be “painful”. The game is free and the author didn’t experiment or learn anything from building it, they just prompted it in one go.
I think that's exactly why AI is suited for 99% of stuff we do.
I have pointed out on here before that instances of truly unique human ideas not grounded in nature or previous ideas from others is almost nil, there are not many examples that someone can give me. All human ideas and work is derivative.
Elves? Humans with pointy ears.
Werewolves? Humans mixed with wolves.
Car tyre? Cart wheel...stone wheel/roller.
Etc.
I feel like prior to GenAI, you’d have had to reckon with the true originality of your idea in some form as you did the research. Creatives having to confront their own unoriginality is such a thing it itself is reflected in countless pieces of media.
So it’s interesting to me that the creator here didn’t encounter the tens of physically published versions, or the hundreds of them shipped to digital app stores, or all the codebases on GitHub, in the course of making this. I’m sure they would have done naturally prior to GenAI. Is that good or bad? I don’t know! But it’s interesting to me.
It seems to me that most media genres discover the most interesting parts relatively early, then most subsequent work is deeply derivative. I feel that way about video games, digital music, movies.
I'd wager it's because ideas are simpler to explore orthogonally, giving an overview of what's possible.
I think this is false. New ideas are born every minute, and llms arent going to help people with those for the most part, they'll end up steering you back towards the gradient if you do.
While I agree that it isn't revolutionary that it could implement this from a single prompt, what's surprising to see is how well done this one is compared to the other tries. The controls and movement are smooth, the animations aren't jittery, the ui makes sense, there's a clear progression in difficulty. This model clearly "understands" the implementation of this game far better than the others did.
Yes, this subtly seems worth noting. That smoothness suggests that even if many of the concepts are common/non-original, the bringing together of various pieces in a form that works well on modern mobile browsers is still impressive - browsers are moving targets and even if there are open source versions of this, it's comparatively rare they'd get continual care and attention to stay fully current (unless implemented via an o/s engine)
I also realized this, a quick Google search would’ve told me that this game has been made several times before, also way before I ever had this idea. Apparently it’s a pretty obvious game idea.
Ah well, it’s still fun and it does appear to measure how good AI is in creating these kind of games.
If OP is anything like me, they probably played it (or saw and wanted to play it) on the GBA too, and the memory became an idea, forgetting they had actually played it because it did exist.
But also, how original can a game idea ever (now) really be – there's always going to be things you can describe it as 'like' or a mix of, even if not identical. And for such simple things, very little room for being non-identical to whatever they're like.
In case it all just comes from training data, "one shotting" a game would be more comparable to "git pull" and changing some assets than "generating code".
I'm not saying this is how it works, I'm trivializing LLMs with this statement, but when I see someone on linkedin excited about generating checkers and chess my first thought is "you could have done that with git pull for the past 20 years".
I believe that AI is not a signal of all white collar jobs being replaced, it’s a signal that SWE was in its own bubble and this is the pressure to pop it.
Most software is not needed, YCombinator itself works on the philosophy of “maybe 2/300 ideas are good” and even among those their biggest hitters were social media platforms and undercutting existing services using VC money.
It was a big game that didn’t make a lot of sense in retrospect, and now with these AI super coders it just doesn’t make sense faster.
Software ate the world, and AI is the garbage disposal meant to chew up the leftovers.
Same thoughts exactly. I personally started looking into indie game dev and I've just started to realize how naive I was and how hard just game design can be, and that I'll probably never be good at it, and that most of my ideas are pretty garbage (or incomplete at best).
Even with the perfect AI to write, one would need to iterate through many different ideas, play testing constantly, getting people to play test and analyze what they found fun and where they got stuck. And to get the best ideas you'll need to be playing lots of different kinds of games.
I’d guess that large games have an even higher percentage of pre-existing mechanics than small games do. It’s much easier to do something experimental if you don’t have to wrangle investors and entire teams of writers, artists, and developers.
90% of everything is crap according to the late SF write Theodore Sturgeon. That was true before AI and it remains true presently. Does it really matter whether this game was in the training data or not here? I guess if one is trying to assert it can build original ideas (and it can, I've done it), but it seems like this is the equivalent of pulling something from Stack Overflow and customizing it given the description of the problem.
IMO the ability to describe a game and let the AI implement a PoC is pretty wild. It's a signal as to whether such an idea is worth pursuing further to me rather than a finished product. And I am enjoying all the experimentation with existing genres as well as the occasional truly original experience due to the dramatically lower cost of entry. What these efforts lack currently is the playtesting and polish that is hard without a human in the loop. So much like agentic engineering, the productive work is in being a centaur. It surprises me how much pushback this is getting from the demographic that embraced the relatively inscrutable git over simpler alternatives for small teams along with the tower of Babel of equally inscrutable frameworks and APIs.
It's not unlike Martin Scorsese admitting upfront he's using GenAI as a creative tool to visualize scenes for his scripts. The predictable backlash that he dare use AI in any way for any aspect of his craft despite his irrefutable oeuvre is a sign of the times more than a legitimate objection to me. Ask the users of deviantArt to stop working with Photoshop and see how that goes.
Having worked in the game industry in the past and adjacent to Hollywood over my career, they were already top heavy exploitative cultures before AI. And any auteur that thinks they can replace humans with agents is as tuned in to GenAI as the tech CEOs and VCs that happily announced layoffs and instituted tokenmaxxing benchmarks to measure the "incredible" boost in productivity AI enables.
So my question, ahead of the mandatory downvotes for not chanting along with the torch-bearing mob against AI in every way is: beyond the CapEx and the buildout issues (both legit IMO), how is AI impacting you negatively and personally?
You are the second person to respond to my question that’s entirely orthogonal to the actual AI usage here with a very self-conscious screed. Go read my responses to the first one :)
Usually it’s an idea somebody had in a flight of fancy or inspiration but they haven’t really shown much interest in the actual medium prior, so they don’t really have any knowledge of its existence and then they also don’t go out of their way to confirm if it already exists.
Like I remember in college I had something akin to the idea of “50 people 1 question.” I was starting to become interested in shooting my own documentaries and was particularly interested in man on the street style interviews. I pitched it to a friend who then told me about 50p1q, which baffled him because it was like the hot thing already a year or two prior haha.
Anyway that’s just something I think happens a lot. And now with genAI people don’t throw the idea around even, they quickly do a crappy version of the thing, present it, then find out it exists. Which isn’t terrible I guess but it’s one less filter for my better or for worse.
Looks kinda like "Sheepherds" which came out recently.
However as others have pointed out the idea is a common one, probably because many people are exposed to sheep and sheep dogs and farming. Which further reinforces a previous point I made that all human work is derivative and barely anything actually original.
But that's why it doesn't matter! Make that game/app/website that someone else has made before, make your own interpretation! The beauty and uniqueness is in the skin not the flesh!
I wonder if this is the real problem: it was too good, and a lobby of companies feeling threatened by the competition decided to push the jailbreak narrative as a scapegoat.
There were dozens (if not hundreds) of more complex games made by Fable on Twitter the first day it was released. The only reason this is on HN frontpage is the stupid clickbait title.
Not sure if it would've gone to the front page of Hackernews with that title! I was also trying to make a little fun about the drama around Mythos/Fable: Even though Fable did this really well, to me it does not appear to be fundamentally different from other top models.
It instructs me to rotate my phone. The pasture doesn't get any bigger, but now the top bar blocks half the screen. The tooltip about rotating stays in the middle of the screen. Unplayable. There's a music note indicating sound, but I never heard the dog bark.
It's exactly the kind of unpolished slop I expected it to be.
If this is what you imagined, you need to imagine better.
* Pathfinding is terrible (if I end up inside the fenced area clicking outside doesn’t lead me out).
* Forcing me to go landscape while not even filling the entire screen is terrible (where did you even test this). * Controls are disastrous (I’m either barking all the time or a bark makes my sprite ignore my movements).
You one-shotted this, and I will admit it’s incredible that these agents can create something like this in minutes.
But your statements along with the “most dangerous AI model” in the title are disingenuous. Please do better.
It’s sad that someone can think about a game for years and never really spend the time to just build it out. This is a very simple game even a CS student could build for an assignment. But now we’re supposed to be impressed an AI can one shot it for $20 dollars.
Forces me to rotate to get warning message to disappear (works fine on portrait, but regardless forces me to play with two hands..), when rotate doesnt even fit on phone.
fROnTEnD DeV Is DeAd
DeSiGN Is DeAD
Cool idea tho, could be a fun game if if the UX wasnt so hostile.
But all this was in the training data no?
https://github.com/Nuno1123/chaser https://github.com/tee-lab/collective-responses-of-flocking-... https://shoze.itch.io/sheep-game https://store.steampowered.com/app/3006280/Sheepherds/ https://ameiswhattodo.itch.io/sheepy
Oh wow, this is just a straight ripoff of that last one isn’t it?
I feel like this is a recurring theme. Someone posts about one shotting something and then it’s revealed that you could also have one-shot it buy git cloning the repo it’s based off of
The real point of these posts is karma farming. Posting a good one shot is worth at least a hundred or so upvotes. People need to recognize this for what it is and ignore these kind of things.
3 replies →
[dead]
My Belgian Tervuren and I have a basic herding title and about 4 years of herding experience.
The sheep movement is excellent. You could make it even more realistic by having them favor lusher areas and by having one occasionally bolt spastically (hard mode?)
A handler mode where you play as a human and shout commands at the dog could be cool too!
> My Belgian Tervuren and I have a basic herding title and about 4 years of herding experience.
Do you happen to have any videos of a Tervuren doing actual herding work or competing in herding trials?
I was apparently under the wrong impression that Tervurens are no longer bred for herding. My girlfriend just told me that there are still dedicated herding lines, but when I started looking into it, I could hardly find any footage. Compared to the thousands of videos of Border Collies working stock, there seems to be very little material available.
I'd also be interested to hear more about what your dogs are like. Are they from a specialized herding line?
Do you see them as being closer to the typical working Malinois - very intense, high-drive dogs with a strong prey focus - or more like a blend of the classic German Shepherd and Border Collie traits?
The latter type, at least in my experience, can make excellent everyday companions as long as their owners know what they're doing and provide enough mental and physical stimulation.
The reason I'm asking is that I've been considering getting another Malinois. I used to own a fourth-hand Belgian Shepherd who eventually turned into a reasonably good everyday companion, but it was a long and sometimes difficult road. There were also certain situations where he could never really be trusted.
He also had a habit of finishing arguments that my Pointer started in a very Malinois-like fashion, while my Pointer was generally too stupid(and proud) to learn anything from that experience.
Feel free to write a wall of text if you feel like it.
Come by!
[dead]
I think it’s impressive that an LLM can take you to a local maxima in one-shot.
But once you start maintaining it, improving it and fixing bugs, you’ll eventually need to rip it apart and put it back together again while understanding how it all works.
This is why I think the better approach isn’t to one-shot but to have the architecture in your head and build it up piece by piece, with the AI accelerating the code writing.
I’ve found it very easy to maintain, add features to and fix bugs in software I’ve written entirely with LLMs, and in languages and frameworks with which I’m unfamiliar. You just ask the LLM to explain the code and then work with it to come up with the fix.
How big are those projects.. I dont think this is good for your mental health or physicaly your brains health. Problem solving keeps your brain strong. The laziness in us is inclined to take shortcuts, don't do it. Its like driving your car 3 blocks instead of walking, your physical health will suffer.
8 replies →
Exactly this! People think the one shot gets them to 95% complete on an implementation of their vision. Issue is, it actually gets you the AIs vision adjacent to what you want, and coercing that into the actual implementation you want is now 95% of the work.
This is really no different from working with humans. A visionary founder has to spend tremendous effort to get their engineers to will a vision into existence. This will be the key skill with AI.
LLMs are good now at looking at existing project and suggesting big refactors for technical debt removal and new better architectures after the project grew organically for a while
I think this is true for projects beyond a certain complexity. I have 100% vibe coded projects with tens of thousands LOC, and haven't seen any real issues with fully automated maintenance. Will that approach work in every scenario, absolutely not, but the size and complexity of projects where it does is growing with each new model release.
> you’ll eventually need to rip it apart and put it back together again while understanding how it all works.
This is what spec .md files are for, skill issue
Strongly agree, well said. The one-shot is sexy purely because that first demo is so impressive. Going from zero to working app in minutes.
Like you said, working and maintainable are very different things. One-shot hits a wall the moment you need to do anything non-trivial after the initial generation. Bug fixing is extremely hard, even with AI assistance. Same with feature additions. It's pretty much black box at this point on. AI that wrote it now goes in loops wasting tokens without being able to can't reliably fix it either, because it has no memory of the architectural decisions it made (or didn't make, for that matter) the first time round.
What I realized is that the failure here is the absence of a shared mental model between you and the code.
I'm a product designer with average front-end know-how, and a solid understanding on HTML/CSS and how the web works, coming from the era of hand-coding html/css files. After vibe-coding a few products early this year, purely to learn how AI works, how to design AI interaction patterns etc., I built something called Intent Model. (largely inspired by SDD / BDD.
Intent model is a structured, typed artifact (basically a JSON contract) that captures actors, entities, journeys, rules, and constraints before I write (or make the AI write) any code. It sits upstream of everything. Think of it like a condensed, strict distillation of your PRD / BRD / requiremnt doc.
When you hand the AI a well-defined intent file instead of a vague brief, this one-shot becomes structured and bound by rules. Now you're giving it an architecture and to conform to. You define (or make the AI define) the precise variable names, their types, lifecycle, user roles, responsibilities, business rules and constraints in the file. Every generated artifact can trace back to a decision you made deliberately, reviewd and signed-off.
In the design world, we already do this by using design tokens. We can tell the AI that it needs to strictly use design tokens and not use stray properties like a hex color value or raw values not defined in the token contract. This is easily auditable by AI as well.
The result is you can still move absurdly fast and still maintain the understanding, which the one-shot approach throws away. This way, you know why every piece exists because you defined the intent before the AI implementated it.
AI is the accelerant, and you're the architect. The intent is the blueprint you generate to guide/harness the AI.
The best part is, once you have an intent contract at the heart of your project, it becomes impossible to break things too, logically or experience-wise.
Curious enough, I tried the same prompt with Qwen3.6-27B.
One shot produced a game with no sheeps. I had to told it to fix two bugs then.
Overall, the graphics and games seems good enough and better than most of the closed models that were shown. However, not surprisingly, falls short of Fable.
I've put the index.html and open code session here:
https://github.com/da-x/when-ai-fails/tree/qwen3.6-27b/shepa...
Would love a PR with this!
If you sit down and write that game by hand you will not only finish it in a week but also learn a lot of things along the way and perhaps even discover something about the game and you did not imagine. That is how programming works. It is a search problem.
Also this is a game has very simple mechanics I am sure you can generate as easily with Cursor or some other tools.
More deeply than programming being a search problem, programming is a means to an end.
If the end is a combination of education and product discovery, then yeah maybe, although those are also dimensions of personal productivity that can be amplified by leveraging AI tools.
If the end goals of programming is leveraging computer automation, then nobody actually cares how the automation infrastructure gets established, and the less distractions with low value implementation complexity, the better.
Cursor has access to the latest models so it should be equivalent, right?
Or is there some other AI usage described in this article that is not supported by cursor?
I don't use Cursor so I mentioned it as slightly impartial suggestion but my point is broader. I hear and have seen results from others using Composer 2.5 which is only available in Cursor.
“ can it build a game idea I've had for years, in a single shot?”
Do people do no research or introspection when they’ve had an “idea for years”? There are countless examples of this exact game. I played this on the Gameboy Advance! There’s like 50 of them on the App Store right now.
The standard “this almost certainly exists wholesale in the training data” applies, but I’m also interested in how you carry an idea for years and don’t notice this, or whether the “idea” here was actually “using this thing that’s been remade thousands of times as an AI benchmark”.
There’s nothing wrong with remaking an old classic formula, especially in game dev. It’s the describing it as “an idea I’ve had for years” that rings weird.
It's called Ostrich effect and we all do it. You enjoy toying with your own idea so much, that your brain shields you from the pain of finding out it already exists. Deep down you know it probably exists. It's harmless, unless there's other people's time and money involved.
> You enjoy toying with your own idea so much, that your brain shields you from the pain of finding out it already exists.
Doesn’t look like the author toyed with the idea at all, though, apart from having it in their head. Considering how they describe themselves (Check the About/Home page), if they had toyed with it at all they would have already built it.
I also don’t see why finding out it exists would be “painful”. The game is free and the author didn’t experiment or learn anything from building it, they just prompted it in one go.
4 replies →
I think that's exactly why AI is suited for 99% of stuff we do.
I have pointed out on here before that instances of truly unique human ideas not grounded in nature or previous ideas from others is almost nil, there are not many examples that someone can give me. All human ideas and work is derivative.
Elves? Humans with pointy ears. Werewolves? Humans mixed with wolves. Car tyre? Cart wheel...stone wheel/roller. Etc.
I feel like prior to GenAI, you’d have had to reckon with the true originality of your idea in some form as you did the research. Creatives having to confront their own unoriginality is such a thing it itself is reflected in countless pieces of media.
So it’s interesting to me that the creator here didn’t encounter the tens of physically published versions, or the hundreds of them shipped to digital app stores, or all the codebases on GitHub, in the course of making this. I’m sure they would have done naturally prior to GenAI. Is that good or bad? I don’t know! But it’s interesting to me.
11 replies →
Just because AI can give you a recipe for an sandwich doesnt mean everyone who sells or buys or experiments making sandwiches are going to stop.
The question is not whether the ingredients are original. They almost never are. The question is whether the synthesis is any good
It seems to me that most media genres discover the most interesting parts relatively early, then most subsequent work is deeply derivative. I feel that way about video games, digital music, movies.
I'd wager it's because ideas are simpler to explore orthogonally, giving an overview of what's possible.
I think this is false. New ideas are born every minute, and llms arent going to help people with those for the most part, they'll end up steering you back towards the gradient if you do.
6 replies →
While I agree that it isn't revolutionary that it could implement this from a single prompt, what's surprising to see is how well done this one is compared to the other tries. The controls and movement are smooth, the animations aren't jittery, the ui makes sense, there's a clear progression in difficulty. This model clearly "understands" the implementation of this game far better than the others did.
Yes, this subtly seems worth noting. That smoothness suggests that even if many of the concepts are common/non-original, the bringing together of various pieces in a form that works well on modern mobile browsers is still impressive - browsers are moving targets and even if there are open source versions of this, it's comparatively rare they'd get continual care and attention to stay fully current (unless implemented via an o/s engine)
1 reply →
I also realized this, a quick Google search would’ve told me that this game has been made several times before, also way before I ever had this idea. Apparently it’s a pretty obvious game idea.
Ah well, it’s still fun and it does appear to measure how good AI is in creating these kind of games.
Well … it’s a measure of how good it is at reproducing a game that probably already exists in multiple forms in its training data.
2 replies →
I agree - it's worth doing just for fun.
I did the same recently just for fun - I really enjoyed "Gravity Force" on the Amiga - itself a lunar lander variant.
Could a model build a Gravity Force like game I could run in-browser? Yep! (I never made it as good as Gravity Force - just got the basics down)
If OP is anything like me, they probably played it (or saw and wanted to play it) on the GBA too, and the memory became an idea, forgetting they had actually played it because it did exist.
But also, how original can a game idea ever (now) really be – there's always going to be things you can describe it as 'like' or a mix of, even if not identical. And for such simple things, very little room for being non-identical to whatever they're like.
This is a thought I've had about genAI.
In case it all just comes from training data, "one shotting" a game would be more comparable to "git pull" and changing some assets than "generating code".
I'm not saying this is how it works, I'm trivializing LLMs with this statement, but when I see someone on linkedin excited about generating checkers and chess my first thought is "you could have done that with git pull for the past 20 years".
I believe that AI is not a signal of all white collar jobs being replaced, it’s a signal that SWE was in its own bubble and this is the pressure to pop it.
Most software is not needed, YCombinator itself works on the philosophy of “maybe 2/300 ideas are good” and even among those their biggest hitters were social media platforms and undercutting existing services using VC money.
It was a big game that didn’t make a lot of sense in retrospect, and now with these AI super coders it just doesn’t make sense faster.
Software ate the world, and AI is the garbage disposal meant to chew up the leftovers.
Well, “an idea I’ve had for years” and “something that has never been done before” are not the same thing.
This is fair! I am possibly attaching some notion of originality to the word “idea” in the context of a project that others don’t.
1 reply →
Same thoughts exactly. I personally started looking into indie game dev and I've just started to realize how naive I was and how hard just game design can be, and that I'll probably never be good at it, and that most of my ideas are pretty garbage (or incomplete at best).
Even with the perfect AI to write, one would need to iterate through many different ideas, play testing constantly, getting people to play test and analyze what they found fun and where they got stuck. And to get the best ideas you'll need to be playing lots of different kinds of games.
I played it in Twilight Princess and I did not have fun and Link probably didn't either.
Most small games are recombinations of existing mechanics anyway
I’d guess that large games have an even higher percentage of pre-existing mechanics than small games do. It’s much easier to do something experimental if you don’t have to wrangle investors and entire teams of writers, artists, and developers.
90% of everything is crap according to the late SF write Theodore Sturgeon. That was true before AI and it remains true presently. Does it really matter whether this game was in the training data or not here? I guess if one is trying to assert it can build original ideas (and it can, I've done it), but it seems like this is the equivalent of pulling something from Stack Overflow and customizing it given the description of the problem.
IMO the ability to describe a game and let the AI implement a PoC is pretty wild. It's a signal as to whether such an idea is worth pursuing further to me rather than a finished product. And I am enjoying all the experimentation with existing genres as well as the occasional truly original experience due to the dramatically lower cost of entry. What these efforts lack currently is the playtesting and polish that is hard without a human in the loop. So much like agentic engineering, the productive work is in being a centaur. It surprises me how much pushback this is getting from the demographic that embraced the relatively inscrutable git over simpler alternatives for small teams along with the tower of Babel of equally inscrutable frameworks and APIs.
It's not unlike Martin Scorsese admitting upfront he's using GenAI as a creative tool to visualize scenes for his scripts. The predictable backlash that he dare use AI in any way for any aspect of his craft despite his irrefutable oeuvre is a sign of the times more than a legitimate objection to me. Ask the users of deviantArt to stop working with Photoshop and see how that goes.
Having worked in the game industry in the past and adjacent to Hollywood over my career, they were already top heavy exploitative cultures before AI. And any auteur that thinks they can replace humans with agents is as tuned in to GenAI as the tech CEOs and VCs that happily announced layoffs and instituted tokenmaxxing benchmarks to measure the "incredible" boost in productivity AI enables.
So my question, ahead of the mandatory downvotes for not chanting along with the torch-bearing mob against AI in every way is: beyond the CapEx and the buildout issues (both legit IMO), how is AI impacting you negatively and personally?
You are the second person to respond to my question that’s entirely orthogonal to the actual AI usage here with a very self-conscious screed. Go read my responses to the first one :)
1 reply →
Usually it’s an idea somebody had in a flight of fancy or inspiration but they haven’t really shown much interest in the actual medium prior, so they don’t really have any knowledge of its existence and then they also don’t go out of their way to confirm if it already exists.
Like I remember in college I had something akin to the idea of “50 people 1 question.” I was starting to become interested in shooting my own documentaries and was particularly interested in man on the street style interviews. I pitched it to a friend who then told me about 50p1q, which baffled him because it was like the hot thing already a year or two prior haha.
Anyway that’s just something I think happens a lot. And now with genAI people don’t throw the idea around even, they quickly do a crappy version of the thing, present it, then find out it exists. Which isn’t terrible I guess but it’s one less filter for my better or for worse.
(Exaggerated:) Guy who would never pay 20€ to another dev for such a game, pays same amount for AI.
Applause to Anthropic: mission accomplished!
There's always been a premium paid for custom work, though. OP just paid off-the-rack prices for a custom-made suit.
That's no surprise, telling the AI what to do is basically the IKEA effect.
https://en.wikipedia.org/wiki/IKEA_effect
45 minutes and $20 ? You can make it in 5 minutes with DeepSeek!
I even did it for free in their web chat.
https://jsbin.com/sigesemeyi/edit?html,output
DeepSeek Flash was also able to do it, even with reasoning disabled, but Pro gave much nicer graphics.
Your Fable version is prettier though!
Edit: And has better gameplay. And sounds. OK, nevermind! Fable wins this one :)
Update: I've been having a lot of fun asking various models to make this game.
> make me a simple html5 canvas game, top down, where you play as a dog who herds sheep into a pen
I like Gemma 31B's version (free model on OpenRouter): https://jsbin.com/mufirivixo/edit?html,output
In this version, the sheep all get stuck in the corners and edges. This doesn’t happen in the Fable version.
The Fable version has some kind of "gravity" that pulls the sheep together.
Some of the other games mentioned here use fancier flocking mechanics (like boids!) https://news.ycombinator.com/item?id=48518518
Love the:
> // Sky-like background
comment for the ground! :) No doubt most of the training data is for sky...
So it created a trivial game that a teenager could’ve built as a part-time project while acquiring deep knowledge.
Humans can do lots of things, I don't see how that's relevant. This post is about AI progress.
Looks kinda like "Sheepherds" which came out recently.
However as others have pointed out the idea is a common one, probably because many people are exposed to sheep and sheep dogs and farming. Which further reinforces a previous point I made that all human work is derivative and barely anything actually original.
But that's why it doesn't matter! Make that game/app/website that someone else has made before, make your own interpretation! The beauty and uniqueness is in the skin not the flesh!
But isn't getting an LLM to n-shot something just going to produce non-unique, non-original interpretations of an idea?
I’m sure I saw a blog post about this same mechanic being made by llms back a year or so ago too
Well it was the same person at least
https://news.ycombinator.com/item?id=43360648 https://news.ycombinator.com/item?id=43298945
I sure do miss Fable. It just knew how to do things and do them well. Sad it’s now blocked.
I wonder if this is the real problem: it was too good, and a lobby of companies feeling threatened by the competition decided to push the jailbreak narrative as a scapegoat.
There were dozens (if not hundreds) of more complex games made by Fable on Twitter the first day it was released. The only reason this is on HN frontpage is the stupid clickbait title.
Some random examples:
https://x.com/fe_yukichi/status/2064635098411180374 https://x.com/akiraxtwo/status/2064780732082651402 https://x.com/kieradev/status/2064482704763085202 https://x.com/VincentLogic/status/2064699740936356065 https://x.com/XiaohuiAI666/status/2064994538591223911
No, the reason is that it is a follow up on a (multiple) threads march last year and shows the progress of ai.
The article’s title seems needlessly dramatic, the article itself doesn’t reference the LLM’s danger.
The title could have been just “Shepherd’s Dog: A game by Fable 5”.
Not sure if it would've gone to the front page of Hackernews with that title! I was also trying to make a little fun about the drama around Mythos/Fable: Even though Fable did this really well, to me it does not appear to be fundamentally different from other top models.
Yeah, fundamentally the same: Worthless.
5 replies →
Brilliant marketing here in the title
Enjoyed playing it, here's the direct link to play as otherwise you have to click from the article to the GitHub and then find the correct demo link
https://vnglst.github.io/when-ai-fails/shepards-dog/claude-f...
Thanks for that, I messed up copying the links into the article!
When you say €20 worth of tokens is it fair direct API call price or subsidized claude code?
Direct API access I’m afraid, it was not my intention to spent it all in one go on this. But after 12€ I didn’t want to stop anymore
[dead]
This game need a "bah-ram-ewe" cheat code where the sheep dog turns into a well-mannered pig who politely asks the sheep to return to their pen.
Playing on iphone13 mini.
It instructs me to rotate my phone. The pasture doesn't get any bigger, but now the top bar blocks half the screen. The tooltip about rotating stays in the middle of the screen. Unplayable. There's a music note indicating sound, but I never heard the dog bark.
It's exactly the kind of unpolished slop I expected it to be.
ok, after playing this game, I started respecting shepherding dogs' skills even more.
Oddly, I wonder if this is not a great benchmarking prompt.
Yeah, but can it one-shot Sven Bomwollen?
2002 - a strange game!
Pretty great game, am having some fun
I didn't even have to play. Immediately after opening, some notification about rotating my phone is obscuring the instructions and I cannot read them.
Damn I couldn't load it on my Nokia n95 from 2007 either. Damn bruh, these silly devs should make this stuff work on everything.
I am on a flagship samsung that runs for example the Red Alert 2 browser port well.
OP is just pushing slop, the 80% part anyone gets for free. (well 20 bucks)
> It's really fun and exactly how I imagined it.
If this is what you imagined, you need to imagine better.
* Pathfinding is terrible (if I end up inside the fenced area clicking outside doesn’t lead me out). * Forcing me to go landscape while not even filling the entire screen is terrible (where did you even test this). * Controls are disastrous (I’m either barking all the time or a bark makes my sprite ignore my movements).
You one-shotted this, and I will admit it’s incredible that these agents can create something like this in minutes.
But your statements along with the “most dangerous AI model” in the title are disingenuous. Please do better.
In which harness?
It’s sad that someone can think about a game for years and never really spend the time to just build it out. This is a very simple game even a CS student could build for an assignment. But now we’re supposed to be impressed an AI can one shot it for $20 dollars.
BAA VRAM EWE
He should ask AI to tell him that #aaa text on #eee background is not acceptable.
The "world's most dangerous AI" framing is funny here because the result is basically the least dystopian use case imaginable
Now next game - The Boy who cried wolf! Wolf!
Forces me to rotate to get warning message to disappear (works fine on portrait, but regardless forces me to play with two hands..), when rotate doesnt even fit on phone.
fROnTEnD DeV Is DeAd
DeSiGN Is DeAD
Cool idea tho, could be a fun game if if the UX wasnt so hostile.
[dead]
That’s one tired sheepdog.
This was my second attempt, I'm still learning! Besides, the wolf was freaking me out.
Always fun having a go, mind you Michael Nyman had some thoughts on all this: https://www.youtube.com/watch?v=xn1_vUe_Vws
For interest, some shepherds run two dogs, each on a different whistle or voice command pitch.
"a game idea I've had for years"
Bruv, there are already countless games with this exact mechanic...
As far as I can tell it is possible to get this sort of quality game with a properly tuned harness out of one of the cheaper models.