Genie 4 Builds Worlds You Can Walk Through. Is That a Path to AGI?

On October 27, Google DeepMind announced Genie 4, a “world model.” Think of it as a flight simulator that an AI builds by itself: you move through a place, change it by typing, and it reacts. Google says understanding and simulating the world is a crucial part of building AGI. Does a good simulation move us closer?

What Genie 4 can do (Google’s claims)

• You can type to change the weather, move the camera, or add a person or object while the world runs. Gemini adapts your prompt to fit the scene.
• It is DeepMind’s first real-time world model to generate audio natively, including dialogue that responds to you.
• A model called Gemini 3.8 Flash acts as a guide in the loop. It breaks goals into steps and checks each frame.
• In a robotics demo, a separate model, Gemini 4 Argon, planned the steps while Genie 4 rendered a humanoid robot playing checkers.

For now it is a research preview for trusted testers. Google has said it is coming to Google AI Ultra subscribers and early API partners, but has given no date. It has published no benchmark results or technical report.

What it cannot do yet (Google’s own list)

Visual details and object permanence can drift over time. The model supports only a few minutes of continuous interaction. It cannot simulate real places with perfect accuracy, and readable text usually appears only if you put it in the prompt. Complex interactions between several agents are still open research. Press reports put the session limit at about two minutes, so the exact figure is unclear. Physical interactions are still a weak point.

The case that world models lead toward AGI

• Demis Hassabis (Google DeepMind) has said AGI is five to ten years away, and that one or two more breakthroughs are needed, including reliability, reasoning, and memory. He sees world models as a key part of an eventual AGI system.
• Yann LeCun, who left Meta to start AMI Labs, says language models are not a path to human-level intelligence because they cannot build a world model that predicts what happens next and connects cause and effect.
• Fei-Fei Li’s World Labs takes a different route and builds persistent 3D scenes instead of video generated in real time.

The case for caution

• Looks real is not the same as understands. A TechRadar critique argues that faking plausibility is easier than capturing the rules behind behavior. It also notes that AGI has no official definition, so the claim is hard to test.
• Even the world-model camp is split on method. LeCun’s favored approach, JEPA, predicts abstract changes in the world rather than pixels. We did not find a comment from him on Genie 4 itself. A researcher at the Beijing Academy of AI described Genie 3 as closer to a “world-model-style video generator” than a physics-and-causality model (36Kr).
• Genie 4 is not one model learning physics. In the checkers demo, a separate Gemini model did the planning, and Genie drew the result. Google’s own name for its direction is “hybrid world models”, which combine language and code reasoning with pixel simulation. The Batch’s editors also asked whether a plan can truly fail in Genie 4, because the Gemini guide may steer the world toward the intended result.

Our read

Genie 4 is solid evidence that AI can make convincing, controllable simulations. It is weak evidence about AGI. The nearest use is cheaper practice for robots and agents and quick previews for creators, a theme we also covered in our posts on NVIDIA’s GR00T 2 and the AGI debate.

What to watch: an independent benchmark, longer consistent sessions, and proof that robots trained in Genie worlds work in the real one.