Hacker News Reader

Comment by goldenarm | Hacker News Reader

Funny idea but wow the gameplay is terrible and full of bugs. Bad hitboxes, shells that bounce in front of holes, bad physics.

AI still cannot oneshot 2D platformers, the most documented videogame genre with so much source code available.

View on HN

One-shotting is a bit of a red herring, imo. Let's say I wanted to build a "Super Dario" meme platformer. I would do some research, find a platformer on Codepen or Github that "feels" good, tweak it to my liking, change the sprites, and voila.

I constantly have to tell agents to just "look it up online" and "don't hallucinate your own components" because people have already done this a million times. Ironically, being more lazy could make these models more useful.

So you're saying that it's not fair to rate a coding model on its ability to code, and instead the best way to use it is to tell it to find existing human-written code online rather than generate actual code on its own?

Isn't part of the human coding that Super Dario tells us LLM are going to replace also made of what you are saying GP is saying?

If you're testing how good LLMs are at compressing information, then I think that's a fair test. Personally, I don't really think that's where their strength comes from (especially considering how much more useful local models that are orders of magnitude smaller than Claude/OpenAI-tier models have gotten). In other words, we already have a "super-intelligence"—it's called the internet, so just use the darn thing.

Yes for any slightly non-standard web UI components I either tell the LLM to search the web and find examples or I provide them myself.

Usually the ones I provide myself are better than the ones they find themselves.

I'm not super clear how they search the web or what they do and don't have access to. It seems to be quite fast.

TBH, I wonder how much one-shotting posts are just lying about one-shotting it.

Many for sure.

Strange, I use M3 and Sonnet a lot, they never hallucinated. Gemini used to make up some function names, but it was in long context windows. My AGENTS.md does specific they should use context7 when unsure about a library, but still…

On the "bright side", it's impossible to die!

That's the power of venture capital!

This was my favorite bit of meta-commentary: it doesn't matter how poorly you play, you never lose!

Even survived the Hegseth Bullet Bill attack

More like keeps "extending by 1 more week"

you mean amodei

It seems that you do lose a life when hit by an enemy or falling in a pit, but the latter freezes until you click the toggle sound button, and running out of lives resets the counter to 3 and halves your "valuation".

I think that supports the premise more than it defies it

Strong agree. Very much part of the joke.

Not to mention it feels like 30hz on my 120hz monitor. It's actually impressive how it managed to mess that up.

But yeah guys, AI sure will be writing all the code just give it another 6 months...

It feels like each new SOTA model brings the average quality a bit closer to the asymptote of well-made non-vibecoded software but it can't quite cross it.

I don’t know if it’s showcasing that AI is still that bad, or if it’s purposely built to be this bad in a funny type of way.

Just amazing how right a few people at Nintendo got this in 1985, and its still so easy to get it absolutely wrong.

Nintendo wasn’t even trying, it was all their old sprite characters from Japan market that didn’t do well smashed together in a rushed title for an NA launch. It surprised them.

Huh? This is very much not my understanding of SMB1’s development. Are you maybe thinking of Donkey Kong (made to replace a failed arcade game) and/or the US version of SMB2 (a re-skinned non-Mario Japanese game)?

Nope, SMB1 was a smash up. SMB2 was made by folks that had no idea what they were doing so SMB3 went back to basics. I very much remember the sequence of events. Mario Bro's (and Donkey Kong) were games on the Famicom in Japan that were decent. The issue was with Arcade's in Japan making more money than home systems so they introduced the Nintendo VS cabinets. This helped but it wasn't enough and so in 1985 they greenlit the Super Mario Bros proposal while dealing with the NA launch and the NYC event. Shigeru Miyamoto carried the development and he deserves the credit but corporate was throwing shit at the fan and seeing what sticked. He just so happened to have a plan that required less work from creative.

Any links to read more about this?

Most of it has been lost to internet history but there are still some remnants left.

https://grantland.com/features/the-rise-of-nintendo-video-ga...

https://iwataasks.nintendo.com/interviews/wii/mario25th/4/0/

https://www.reddit.com/r/AskHistorians/comments/2w1zl4/when_...

Thanks!

Shigeru Miyamoto has spoken many times about the process of perfecting the controls for SMB, which is quite the opposite of "wasn't even trying".

Oh he absolutely cared about the programming and process of game development, don't get me wrong. It's that Nintendo (corporate) needed something for the NA market and he was eager to fill that gap with already used characters (such as Mario). Anything else is corporate propaganda they have been telling as a creation myth to make it sound more fairytale-like than it was. They were going broke. They needed money. North America was ripe for the picking. The risk paid off.

Yep. Additionally having experimented pretty heavily in this space, I'll state that even with state-of-the-art models (like NB Pro or GPT-Image-2), good luck getting them to generate sprite sheets with consistent frames for animation particularly if you're shooting for a pixel art aesthetic.

I've made a few sprite-based animations with AI assistance but it still required a fair amount of manual photoshop (or in my case Aseprite) post-processing.

If you're not going for pixel art, use the video to video models (the character swap ones that take in a video reference of the animation and a still of the character (the tiktok deepfake models)).

You can e.g. download a run cycle from Mixamo, use Blender to render a video from whatever angle you wish, and feed that into the model along with your sprite on a green background, it's pretty much perfect.

Thanks for the info! I'll give this a shot. Most of my knowledge is from over a year ago, when I would set up OpenPose-style positioning with a controlnet at a high strength to guide an image-to-image workflow in comfyui. But it was pretty tedious and prone to errors.

It’s a tough problem but Retro Diffusion is making some incredible progress

https://retrodiffusion.ai/

Nice. I need to go check them out again. I really liked their original Retro Diffusion models but this a while back in the SDXL days. It felt like it was trained on more natural language, so prompting it was quite nice for that era.

But if they’ve been working on it specifically for animation, that’s pretty neat.

They have a few different specialized models now and have put a lot of work into generating sprite sheet animations. The latest thing is a pipeline for ensuring pixel perfect results that also look good stylistically.

You’re absolutely right! /s

Reply on news.ycombinator.com