vabruzzo

vabruzzo / snes-gpt

Public

micro-gpt in ASM on the Super Nintendo

36
3
100% credibility
Found Feb 17, 2026 at 23 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Assembly
AI Summary

This project builds a playable Super Nintendo ROM featuring a tiny AI model that generates names using techniques inspired by modern language models, all coded in low-level assembly.

How It Works

1
🔍 Discover SNES GPT

You stumble upon a fun project that brings tiny AI smarts to old Super Nintendo games.

2
📥 Grab the files

Download the whole folder to your computer so you can start playing around.

3
🛠️ Prepare your setup

Make sure you have basic tools like Python ready, following the easy guide.

4
🚀 Create the game

Hit the build button to train a little name generator and make your playable game file.

5
🎮 Load into emulator

Pop the new game file into any Super Nintendo player app on your computer.

✨ Enjoy generated names

Watch as it magically creates fun, fantasy-like names right on the classic screen!

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 23 to 36 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is snes-gpt?

snes-gpt ports a minimal GPT model, based on Andrej Karpathy's micro-gpt from GitHub, to pure 65816 assembly on the Super Nintendo. It runs full autoregressive name generation—embedding, attention with KV cache, RMSNorm, and MLP—all in Q8.8 fixed-point on the SNES's 3.58 MHz CPU and PPU hardware multiplier. Run `make` with cc65 and Python 3 to train a tiny model on names, build the ROM, and load it in any SNES emulator like Snes9x for instant output of 20 generated names on boot.

Why is it gaining traction?

This stands out by squeezing a real transformer forward pass onto 1990s Nintendo hardware with just 8KB weights in ROM and 1KB WRAM, proving GPT inference needs no modern accelerators. Developers dig the extreme constraints: temperature-sampled generation via xorshift PRNG, lookup tables for exp and 1/sqrt, all without libraries. It's a masterclass in asm optimization that spits out playable results immediately.

Who should use this?

Assembly hackers porting ML to microcontrollers, retro computing enthusiasts tweaking SNES ROMs, or embedded engineers benchmarking tiny transformers on low-power chips. Ideal for demos in talks on constrained AI, like running GPT on Game Boy or Arduino next.

Verdict

Fun proof-of-concept with solid README and zero-setup build, but 22 stars and 1.0% credibility score signal early-stage novelty over production readiness—no tests or extensibility. Fork it for hardware hacks if you're into super Nintendo asm experiments.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.