Neural CPU runs Doom E1M1: 28.7M tokens, 2h21m for a verified real frame

SalamanderFickle4315 · reddit · 2026-10-30

• Built on Qwenputer's idea of compiling CPU logic into neural weights and recurrent state, the author used a standalone Qwen-derived checkpoint to render a real 320×200 Doom E1M1 frame.
• The run autonomously generated 28,699,242 protocol tokens, taking about 2h21m end-to-end; cumulative native Metal inference hit 4,193.5 tokens/sec (earlier MLX backend: 501.5 warm tokens/sec—different scopes, not a head-to-head).
• The framebuffer was extracted from neural memory and independently verified against the expected raw pixel hash.
• Not real-time gameplay; the 10,000 TPS goal remains unmet. Doom's program initializes the machine, and the neural circuit executes it. Open-sourced as qwenputer-doom on GitHub.

An absurdly hardcore proof that "weights as a computer" can actually run a game.

Original post →

More from Fun

• Dad and Son Build a 6MB Single-File 3D Open-World Game with AI, 60 FPS on Mobile — rschu · 2026-10-30
• Spies targeting Jensen, Sam, Dario and Elon would be 'weird' not to happen, says dev — rickasaurus · 2026-10-30
• Hugging Face employee's mom gets 🤗 emoji nails to support open source — pcuenq · 2026-10-30
• 'Girls only want one thing': an $496,639 8x NVIDIA B300 deep learning server — Kyrannio · 2026-10-30
• Aella clarifies: an Anthropic employee attended her events, sparking AI-circle drama — astralmatrix · 2026-10-30
• Starlink helps rescue Swedish sailor stranded off South Africa's coast — elonmusk · 2026-10-30

Fun channel →