Google built AI agents a world to practice in. The physics still aren’t quite right.
The New Stack Amanda Caswell ● Covered by 5 sources
Google’s new Genie 4 builds live, interactive worlds for AI agents to practice in. The catch: its physics still wobble, so test results may not be trustworthy.
Based on reporting by The New Stack, Amanda Caswell — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Google DeepMind showed off Genie 4 on Tuesday, a world model that generates interactive scenes in real time with simulated physics, human motion, and native audio. The demos look remarkably polished. But the bigger idea is less cinematic: Google wants AI agents to have a place to rehearse multi-step tasks before anyone calls the result real-world ready.
At a press briefing, research director Tejas Kulkarni described the direction as “these hybrid world models,” where Gemini writes the code and text and Genie turns that into a running scene. Genie 4 can spin up test environments on demand, which is the kind of tooling developers always ask for and then immediately argue about. The obvious question is whether those worlds behave enough like the real thing to make the tests meaningful.
Google says Genie 4 leans on several Gemini models to keep the simulation alive. Gemini Omni helps shape the environment. Gemini 3.8 Flash tracks what’s happening, plans the next move, and adjusts when the scene changes. And unlike a finished video, the system is meant to stay live: users can move through the world with a keyboard, shift the camera, or use text prompts to change what’s happening without restarting everything.
The company’s clearest example has Apptronik’s Apollo humanoid playing checkers, with Gemini 4 Argon planning each move while Genie 4 renders the outcome and captions steer the scene. Google is also pointing to Gemini Robotics as a possible fit, since these simulations could let developers test agents in changing environments without rebuilding each one from scratch.
But the physics still aren’t clean. Google says objects can shift position or act oddly during longer sessions, and that makes evaluation messy. A failed run might mean the agent made a mistake, or it might mean the simulation drifted. Google hasn’t shown that Genie 4 can replace traditional physics engines for serious testing, and it also hasn’t published benchmarks for repeated agent runs or consistency across the same prompt.
For now, access is limited to trusted testers. Google plans to open it up to Google AI Ultra subscribers through Project Genie, plus select partners through the Gemini API and Google AI Studio. Pricing, broader release plans, and performance guarantees are still unannounced.
My take — AI-written commentary, not fact-checked reporting
This is classic Google: dazzling demo first, accountability later. A world model that can’t stay consistent is a nice toy for game prototyping, but it’s a shaky foundation for agent testing, which is exactly where the industry wants to pretend certainty exists. The missing benchmarks matter more than the shiny render.
Read more about this at: The New Stack