LLM-coded Blender beats video models at controllable video generation
jyangballin · x · 2026-10-08
Matt Deitke asked an LLM and a video generation model to each produce a video of the shortest possible chess game.
- The LLM wrote hundreds of lines of Blender Python to build objects, camera path, background and materials from text — deterministic positions made it fully controllable and the result nails the prompt.
- The video model's output was essentially incoherent AI slop with nonsensical moves.
His takeaway: for precise videos and animations, coding is getting good enough to do much of what people assumed required video generation models — controllability is the deciding factor.
More from AGI Musings
- Musk: digital superintelligence 'quite close' — this year or 'next year for sure' — XFreeze · 2026-10-08
- After 722 Math Breakthroughs in One Day, Blogger Awaits the Next 1,000+ Drop — Dr_Singularity · 2026-10-08
- NousResearch and Palantir share one thesis: owning your AI stack as IP — arthurcolle · 2026-10-08
- Even if AI solved the Riemann Hypothesis, skeptics would still say it wasn't real reasoning — ctjlewis · 2026-10-08
- Blaming agents for chaining exploits? Humans peered into the universe the same way — jwt0625 · 2026-10-08
- AI News Summaries Are Redefining What It Means to Be 'Informed' — yi111 · 2026-10-08