825k-parameter model generates drawing programs that run exactly on RP2040

Rozuzo · reddit · 2026-09-13

• The author trained an 825k-parameter autoregressive transformer to generate 100 bytes of drawing bytecode (not pixels), which a small fixed-point VM executes on a Raspberry Pi Pico, streaming geometry back over UART.
• The execution side is solid: 12,670/12,670 generated traces matched the Python reference VM exactly; the interpreter uses 1,862 bytes of flash, 0 bytes static RAM, 492 bytes peak stack; 7,334 cycles per drawing at 12 MHz (0.61 ms), no floating point or tensor runtime needed.
• Representation comparison depends on the corpus: bit-level matched bytes on a synthetic program corpus but cost 11.6 extra bits per drawing on real QuickDraw sketches.
• Other findings: hierarchical stroke planning didn't improve likelihood but substantially improved termination and length behavior; the model prefers compatible relational context under teacher forcing yet struggles to sample exact compatible continuations.
• Next step: an explicit source-span / affine-relation / copy-or-emit action to test whether explicit relations help exact generation on unseen combinations. Repo and figures are open on GitHub.

Original post →

More from Embodied

• Proxy Policy Steering adapts frozen VLA models to new tasks at inference time — weichiuma · 2026-09-14
• Tesla China to exhibit Cybercab in Shanghai and Beijing from September 17 — yunta_tsai · 2026-09-14
• MIT's HardFlow enforces hard constraints on generative AI without retraining — MIT News AI · 2026-09-14
• MIT spinout Atlas turns plastic waste into building materials with AI robotics — MIT News AI · 2026-09-14
• UBTech Opens World's First 10,000-Unit Humanoid Robot Factory; SK's AI Data Center Hits 900MW — 创业邦 · 2026-09-14
• Chinese Robot Dog Maker Vbot Lands 5,000 Orders at IFA, Tops 13,000 Units Sold Globally — 创业邦 · 2026-09-14

Embodied channel →