Article

dbreunig/nesformer

Simon Willison·2026-10-30

— The SEO Cover · Daily Brief · Oct 31, 2026

On "dbreunig/nesformer"

This is a flex disguised as a proof of concept. Yes, you can technically run NES games inside an LLM's weights—the model learned to emulate a 6502 CPU well enough to execute Mario frame by frame. But "very slowly" is doing heavy lifting; what you're actually watching is a parlor trick that burns compute for something a $20 used cartridge already solved. The real takeaway isn't that LLMs can do computation (they can, inefficiently), it's that people keep testing the boundaries of what these models will memorize and execute, which tells you something about their actual learned representations but nothing about their practical value. For practitioners: this confirms LLMs aren't magical computation engines, just very expensive lookup tables that can approximate arbitrary behavior if trained on enough examples.

From the source: dbreunig/nesformer Inspired by qwenputer by Divyansh Agrawal, Drew Breunig baked a working emulator for a Nintendo Entertainment System 6502 CPU into the weights of Qwen3.5-4B. You can now play NES ROMs such as Super Mario Bros (very slowly) running inside Qwen! In a comment on H

Open original

Topics

AI Agents

AI Tools

Quoting Terk Terkowitz

llm-gemini 0.34

Quoting John Gruber

Holo4: powering generalist computer-use agents

Welcome Beam: a 501B open model you can own, customize, and scale

llm-tools-datasette 0.1.1