Show HN: A working Nintendo in ~1GB of Qwen weights

Original Article ↗ Hacker News Discussion ↗

A new project compiles a full Nintendo Entertainment System—CPU, RAM, and ROM—into about 1 GB of large language model weights, allowing Super Mario Bros. to run purely inside a transformer architecture. Commenters see it as a clever proof-of-concept that highlights both the flexibility and extreme inefficiency of using LLMs to emulate deterministic systems, contrasting it with traditional emulators and normal code. The experiment fuels broader reflections on LLM determinism, future software architectures, and how rapidly improving AI tools may reshape engineering work and team sizes.

Naming and nostalgia

  • Several comments debate calling it a “Nintendo” instead of “NES.”
  • Older commenters associate “Nintendo” with classic Mario-era hardware; others note the term became generic for all game consoles, even causing trademark concerns.
  • Some find the title endearingly “grandma tech” coded (every console is a Nintendo, controllers are “remotes”).

What the project actually does

  • The model emulates the NES CPU and RAM inside a transformer’s weights, using a “Gated DeltaNet” state as the machine memory.
  • CPU, ROM, and RAM live inside the model; the only runtime inputs are controller events.
  • This builds on prior “Qwenputer” work that compiled other CPUs into Qwen models.

Determinism, testing, and “working”

  • One line of discussion: how they know it “works” without testing every possible input.
  • Response: identical to how real CPUs are validated—via extensive conformance test suites, not exhaustive state-space coverage.
  • Emulation community test ROMs with thousands of golden instructions were used; still, commenters note rare timing or edge-case bugs could remain.

Efficiency vs ordinary programs

  • Many highlight the massive inefficiency: nesformer is ~150–290× slower than a real NES, despite replicating a simple fixed-function system.
  • This is used as a concrete illustration of how much more compute LLM-based execution needs vs normal code, and why “LLM as OS each frame” ideas would be wildly expensive.
  • Some argue this is fundamentally no more informative than building Minecraft in Minecraft or a CPU in CSS—technically impressive but practically pointless.
  • Others counter that “anything can run anything” is itself an interesting proof-of-concept and useful inspiration for builders.

Broader implications and ideas

  • Speculation about compiling small VMs, shells, or interpreters directly into LLM weights, enabling internal tool use and stepping through reasoning like a debugger.
  • Questions about whether LLMs could become a “universal computing medium,” though many point out this would be far less efficient than just running code.

AI tooling, productivity, and jobs

  • Tangential but active debate: rapid improvement in AI coding tools, higher throughput with fewer people, and fears about shrinking engineering teams and job loss.
  • Others emphasize using AI to do more with the same headcount, though there is disagreement over how this plays out in shareholder-driven organizations.