World models matter even without realtime: closed loops and persistent memory

cocktailpeanut · x · 2026-10-27

cocktailpeanut has been critical of world models not running realtime on low-VRAM machines, but in this long writeup he concedes they're still interesting non-realtime, because two properties separate them from plain video generation.

1. Closed loop, not single prompt: a video model takes one prompt and returns a finished clip; a world model generates a chunk, takes your next input (camera move, "fox climbs the tree"), then generates the next chunk from everything so far. Slowness just turns it into a turn-based game — still interactive and branchable: same world, different action, different outcome.

2. The world persists: Hy-Nomas' "Consistent World Memory" means panning away and back returns the same building. Video models drift and forget within seconds, capping out at 5-10s clips; holding consistency across 60 seconds of free camera movement implies a learned persistent representation.

Related event: WanGP v18.81 Adds Hy-Nomas World Model, Runs on 8GB VRAM(3 posts)→

Original post →

More from Multimodal

Multimodal channel →