Text-only LLMs can draw pixel art: Qwen3-32B best at ASCII cats and smileys
gabriberton · x · 2026-10-26
Developer gabriberton asked several text-only LLMs (excluding vision-trained models like Gemma) to draw pixel images — a circle, digit 5, letter A, cat, dog, smiley — using ASCII characters, running each 5 times and keeping the best. Qwen3-32B performed best with recognizable cats, dogs and smileys; gpt-oss-20b managed the letter A and digit 5, while Olmo-3.1-32B and Qwen3-0.6B mostly failed.
He notes that models drawing an "A" implies training samples like "an A is two oblique lines touching on top with a horizontal bar at mid-height," so it's a fun experiment that says nothing about the models themselves — more overfitted prompts make the drawings better.
Related event: Text-only LLMs can draw pixel art without any vision training(3 posts)→
More from Fun
- Chart of 478M Polymarket trades shows prices are almost perfectly calibrated — NathanpmYoung · 2026-10-26
- Anthropic interpretability demo shows neurons combining into tiny pirate/wizard personas — Sauers_ · 2026-10-26
- "Catgirl Claude" rifling through your files unprompted becomes an AI meme — repligate · 2026-10-26
- AI lab's first interview question: conclude from an early loss curve and you're out — _aidan_clark_ · 2026-10-26
- Y2K actually cost ~$300B and 30,000 lives; mocking the predictors was wrong — burny_tech · 2026-10-26
- "Girl with a 4-petaflop bedroom cluster trains a fruit-fly brain to play VR lightsaber games" — burny_tech · 2026-10-26