DEV Community

Cover image for I Finally Understood World Models: Memory Is the Hard Part, Not Graphics
Krishna Soni
Krishna Soni

Posted on Originally published at global.krizek.tech

I Finally Understood World Models: Memory Is the Hard Part, Not Graphics

Flowing generated light forms suggesting a synthetic environment
Photo by Google DeepMind on Unsplash

I kept filing "AI generates a playable world" under video generation with extra steps. Then I read DeepMind's Genie 2 write-up properly and the distinction finally landed, and it has almost nothing to do with how the frames look.

What Genie 2 does in one paragraph

Give it a single image — in the published demos, one produced by Imagen 3 — and it generates an action-controllable 3D environment. A human or an agent supplies keyboard and mouse input; the model generates the next observation. First-person, third-person and isometric perspectives all show up. It is an autoregressive latent diffusion model trained on a large video dataset, so object interactions, character animation and physics-like behaviour are emergent rather than hand-authored.

The headline caveat: worlds stay consistent for up to about a minute, and most of the published examples run 10-20 seconds.

The property that actually separates the categories

A video generator produces a clip. A world model has to answer a much harder question: what is behind the player right now?

Genie 2 renders parts of the world that left the view accurately once they become observable again. Turn around and the room is the same room. In the follow-up Genie 3 demos you can paint a wall, walk off, and the paint is still there when you return — DeepMind notes this memory was not explicitly programmed in.

That is the whole ballgame, because inconsistency compounds. Autoregressive generation means each frame is conditioned on the trajectory so far, so small errors accumulate and a naive model quietly rewrites the level behind your back.

How the approaches actually differ

System What it emits Persistence You can export
Genie 1 (Mar 2024) 2D worlds, ~1 fps Very short Nothing
Genie 2 (Dec 2024) 3D worlds, action-controllable Up to ~1 min, most demos 10-20s Nothing
Genie 3 (Aug 2025) Real-time 3D, 720p at 24 fps Several minutes, memory ~1 min Nothing
Oasis (Decart) Real-time Minecraft-like frames Forgets layout quickly Nothing
World Labs A 3D mesh from a 2D image Persistent by construction The mesh

That last row is the honest trade-off. Mesh generation hands you something a game engine can consume, then leaves physics, rendering, character control and every rule to you. Genie hands you a playable thing and keeps all of it inside the model, where you cannot touch it.

Why DeepMind pointed this at agents, not studios

The stated bottleneck is training environments, not graphics. Embodied agents run out of sufficiently varied worlds to learn in.

So the evaluation loop looks like this:

  1. Prompt an image, get a world the agent has never seen.
  2. Give the agent a natural-language instruction — the published example is a room with a blue door and a red door, opened in order.
  3. Probe the environment itself. "Go look behind the house" is used as a consistency test, with the agent as the instrument.

By SIMA 2 (November 2025), agents were improving through trial and error inside Genie-generated worlds, with Gemini scoring attempts — the same loop, closed.

What I would not claim from this

  • It is a research preview. There is no API to build a product on.
  • Nothing is exportable and nothing is deterministic. You cannot tune a jump arc or version a level.
  • Genie 2's consistency evidence is lab-curated demos, not independent benchmarks, and the training data was not disclosed.
  • Session lengths measured in minutes are far short of the hours agent training actually wants; DeepMind says so directly.

The useful reframe for me was this: the reason these systems suddenly feel like environments instead of videos is memory. Fidelity was never the blocker.

Where would you put a world you can play but cannot export — blockout, mechanic sanity checks, or nowhere near production?

Source article: https://global.krizek.tech/feed/deepmind-s-genie-2-unlocking-infinite-virtual-worlds-for-ai-and-gaming-3gntq

Try Altered Brilliance: https://play.google.com/store/apps/details?id=tech.krizek.alteredbrilliance
Join The Power Of Gaming: https://discord.gg/sbYSPcCqJn

Top comments (0)