Show HN: A working Nintendo in ~1GB of Qwen weights(github.com)
255 points by dbreunig 18 hours ago | 122 comments
tl;dr: A hobbyist compiled a working NES 6502 CPU, 2KB RAM, and 32KB cartridge ROM directly into the weights of a Qwen3.5 model, with the Gated DeltaNet state acting as the machine's memory — no training involved. The model generates the CPU's execution trace token-by-token while a host program handles PPU rendering, controller input, and NMI timing. It runs NROM-mapper games like Super Mario Bros. at roughly 150–290× slower than real hardware, with traces that exactly match a reference emulator.
HN Discussion:
  • ↑Excitement about future applications like embedding VMs and interpreters inside LLM weights
  • ~Appreciation for the hack while noting it highlights LLM inefficiency versus deterministic programs
  • ↑Surprise and fascination that RAM is also simulated inside the model, not just the CPU
  • ↓Skepticism about the point of the project since digital logic trivially reduces to weights
  • ↓Dismissal due to AI-written documentation quality