[ggambetta]

I'm very happy to see this! Not so much because of TDIII (which I played, although not nearly as much as Stunts), but because there seems to be some momentum building around recreating old games using AI agents, and I love that! I had explored some related ideas [0] but throwing Claude at the problem seems super promising. The recent Crimsonland thing [1] was great!

[0] https://www.gabrielgambetta.com/remakes.html

[1] https://banteg.xyz/posts/crimsonland/

[lantastic]

Stunts was the greatest! You could make your own tracks, save replays and (IIRC) even resume gameplay from any point in the replay. My very favorite game of all time.

[alberto-m]

The community is still active! In the meanwhile the car format has been figured out, so that modern-day tournaments like ZakStunts feature vehicles not included in the original game.

[bombcar]

Reaume-from-replay is a feature mode games should have, really.

[noufalibrahim]

This is an interesting area. I've felt that with AI, it would be nice to have a project that I work on "by hand" so that my general skills don't atrophy and I've been writing an implementation of the Kyra engine used by old DOS games like Eye of the Beholder. It's mostly well documented and there are full fledged implementations (like with ScummVM) so this is just exercise for me.

I wrote a decoder for the CPS file format they use for sprites and it worked file for all images except one. It rendered half the image properly and then scrambled the rest. I could see that the sprite information was there but there was some offset problem. I had claude dig into it in detail along and gave it the ScummVM source for reference. I also gave it Ghidra so that it could debug the actual EOB.EXE file but nothing we tried got it to render properly. Even SSI's own code which got from a modding wiki failed to render this image. My final conclusion was that it was a half done asset that somehow found its way into the asset archive and is never used in the game but that's a flaky conclusion given that its name is referenced in the EXE.

I've been having a lot of fun upscaling the sprites used for the cutscenes and remixing the music using AI. It's a game I played a lot as a kid so being able to tinker with it at a low level is a nice distraction.

It's purely a "fun" side project without deadlines or anything so I get to do what I want with it without any hassles about "being productive".

[christkv]

Me and some friends spent so much time playing https://en.wikipedia.org/wiki/Stunt_Island

[spankibalt]

The birth of machinima! One of the giants of 90s PC gaming, nothing even remotely like it on other platforms of the day.

[eru]

I played a lot of Stunts in the 1990s. I only learned in the 2010s that you could actually edit the terrain, too.

[TacticalCoder]

Your blog entry:

https://www.gabrielgambetta.com/remakes.html

is wild!

I love the idea. I used Claude Code / Sonnet 4.6 to get back to compile an old DOS game I wrote back around 1991 and for which I had lost the tooling (compiler / linker / notes / build files). It was on my todo list since years, decades even, but I never got to it. With a LLM it was easy: I didn't let the LLM do everything, I used it to find what needed fixing (like two macros I had used with names that were now clashing with "modern" compilers methods names etc.).

> I hacked together the art using my terrible Gimp skills and some Public Domain and Creative Commons assets from OpenGameArt

IMO AI models are better at generating pictures than at writing code, in that pictures do tolerate sloppy approximations. While code doesn't tolerate slop that much.

Why not use AI for the assets?

[vunderba]

> IMO AI models are better at generating pictures than at writing code, in that pictures do tolerate sloppy approximations.

So does code. You can have poorly written code that runs slow but gets the job done because modern machines are powerful enough to make up the difference. (cough the majority of electron apps cough).

You’re definitely going to get dinged more for AI art. People can’t see the code, but they can see the art - that’s their first experience with the game unless the code is so bad that it literally causes visible bugs.

I work heavily in the generative space and while standalone AI art has come a long way, maintaining visual consistency in sprite assets across an entire game still needs a ton of hand holding particularly if you require animated sprite assets without them looking "uncanny valley".

[ggambetta]

That's a great idea! Just haven't gotten that far yet (that article is pre-LLM, and I'm very early in a new AI-driven reconstruction). But yeah, taking the original extracted sprites and having Nano Banana upscale or redraw them is a great idea.

[vunderba]

>But what if we could run the original binary game code unmodified, just with better art?

There's an entire category of rom hacks devoted to this - replacing just the sprite data.

You should also take a look at 3dSen. It’s an emulator that works with original NES ROMs but transforms the graphics into 3D voxels.

https://store.steampowered.com/app/1147940/3dSen_PC/

[stared]

And a laser puzzle game Chromatron, https://quesma.com/blog/chromatron-recompiled/.

But it gets even wilder, as now creating a Game Boy Advance emulator became a benchmark, https://gbaeval.com/.

[mike_hock]

I used to absolutely love this kind of content. Reading the details of how someone reverse-engineered an old game, maybe figured out how some particular part of it worked. A high-fidelity rewrite is the absolute maximum you could achieve.

Unfortunately, as soon as there's something about LLMs and AI agents in there, it all just becomes ... meh.