Genie 4 Has No Game Engine. So How Does It Know What W Means?
Whatzat?
0:00 / 0:00
Genie 4 Has No Game Engine. So How Does It Know What W Means?
11 просмотров · 6 дн. назад
Whatzat?
20 подписчиков
11 просмотров · 6 дн. назад
Google DeepMind's new world model, Genie 4, took a single Van Gogh painting and let people walk around inside it with a keyboard.
But that street has no 3D models, no game engine and no physics code. So how does it follow the keys you press?
We put a game engine, which computes the world from meshes and coordinates, side by side with Genie, which predicts the next frame one picture at a time.
Then we check the original sources on how memory appears without a 3D model and how long it lasts, why small errors pile up, and how a runner can pass straight through a cactus.
Where did Genie first learn that W means "forward"? We trace it back to the eight latent actions in the 2024 Genie paper.
Plus where it is already used (Waymo's driving simulator, robot practice), the "end of game engines" take and the other side of it, and our own view.
Sources: Google DeepMind blog, "Genie 4: A new generation of world simulation" and "Genie 3: A new frontier for world models" (2025); Bruce et al., "Genie: Generative Interactive Environments" (arXiv 2402.15391, 2024); Valevski et al., "Diffusion Models Are Real-Time Game Engines" (arXiv 2408.14837, 2024)
0:00 The street: walking into a painting
1:16 How a game engine draws a street
2:48 Genie draws the next frame
5:20 How does it remember without 3D?
7:53 Who taught it the W key?
10:02 So what changes?
10:43 Another view: the end of game engines?
12:16 My take, and what to watch
12:44 Recap
Subscribe and let's figure out the next 'Whatzat?' together!
Music: "Funky Chunk" Kevin MacLeod (incompetech.com) · Licensed under Creative Commons: By Attribution 4.0 — https://creativecommons.org/licenses/...
Video:
Genie 4: A new generation of world simulation — Google DeepMind, 2026-10-27 · • Genie 4: A new generation of world simulation
Genie 3: An infinite world model | Shlomi Fruchter and Jack Parker-Holder — Google DeepMind, 2025-08-21 · • Genie 3: An infinite world model | Shlomi ...
Genie 3: Creating dynamic worlds that you can navigate in real-time — Google DeepMind, 2025-08-05 · • Genie 3: Creating dynamic worlds that you ...
Introducing Genie 4 — Google for Developers, 2026-10-27 · • Introducing Genie 4
Introducing the Jaguar A I-PACE, equipped with Waymo's autonomous driving technology — Waymo, 2018-03-27 · • Introducing the Jaguar A I-PACE, equipped ...
CVPR 2026 DriveX Workshop Keynote 03 Mingxing Tan (Waymo) - The Waymo World Model — Mobility UCLA, 2026-06-27 · • CVPR 2026 DriveX Workshop Keynote 03 Mingx...
Project Genie | Experimenting with infinite interactive worlds — Google DeepMind, 2026-01-29 · • Project Genie | Experimenting with infinit...
Inside a Google Cloud TPU Data Center — Google Cloud, 2023-08-29 · • Inside a Google Cloud TPU Data Center
Photos:
https://deepmind.google/blog/genie-3-...
https://deepmind.google/blog/genie-4-...
p.1·4 · https://arxiv.org/abs/2408.14837
p.1·3·4·8 · https://arxiv.org/abs/2402.15391
https://deepmind.google/blog/genie-2-...
https://docs.unity.com/en-us/engine/c...
https://waymo.com/blog/2026/02/the-wa...
https://www.pocketgamer.biz/googles-p...
https://techcrunch.com/2026/10/27/goo...
https://techcrunch.com/2026/05/19/goo...
"Google DeepMind 2026.svg" — Google, Public domain · https://commons.wikimedia.org/wiki/Fi...
"Van Gogh - Starry Night - Google Art Project.jpg" — Vincent van Gogh, Public domain · https://commons.wikimedia.org/wiki/Fi...
p.1 · https://arxiv.org/abs/1803.10122
#Whatzat #Tech #Technology #WhatIsThat #Genie4 #WorldModel #GoogleDeepMind