World models are still in their 'GPT-3.5 era,' says Peter Yang after testing Google Genie 4
petergyang · x · 2026-10-29
After building five interactive worlds live with Google Genie 4 + Gemini (a kaiju attacking San Francisco, ramen at a Tokyo street stall, riding a dragon to storm a castle), Peter Yang argues world models are impressive but still in their "GPT-3.5 era" — great graphics and live audio, not yet mature.
He predicts massive industry impact once they reach GPT-5/6-level capability:
- Movies: a dragon-battle VFX shot that takes a studio months could be generated from a single frame by the director.
- Games: world models excel at graphics and sound, but core loops, progression systems and fun mechanics aren't there yet — still, they'll drastically lower the barrier to building good-looking games.
His overall take: world models are the next big thing in AI after LLMs.
More from Multimodal
- Genie 4 Real-Time Worlds: A Pixel-Art Office Where Agents Pretend to Work — Thionne_WTZ · 2026-10-29
- Veo-generated video of a woman riding a dinosaur is hilariously convincing — EitherCartographer63 · 2026-10-29
- Spice gaussian splat scans released free for commercial and non-commercial use — willeastcott · 2026-10-29
- EGGBOX RUNNERS: A Concept Sneaker With Molded Pulp Outsole and Recycled Foam — Delahuntagram · 2026-10-29
- Voice-to-video workflow demo: building a 3D world in CineMap, rendering it in TapNow — umesh_ai · 2026-10-29
- Zen Brush Stroke Minimalism: a reusable prompt template for ink-wash AI art — LudovicCreator · 2026-10-29