DeepMind unveils Genie 4 world model with better physics

Genie 3 made headlines last year when its reveal sent gaming stocks tumbling, and world models have since become a major focus for AI labs. DeepMind's pitch is that a world model creates a virtual world without pre-rendering environments or programming interactive elements — you tell the model what you want, and it generates the experience frame by frame in real time.

The company frames this as a key element on the march toward AGI, with potential uses ranging from infinite training environments for physical robots to rapid prototyping for creative fields. Genie 4 also slots into Google's broader model lineup: it uses Gemini 3.8 Flash as a guiding agent to break down goals and evaluate each frame, which DeepMind says helps keep the simulation on track across changing environments.

The limits remain notable. DeepMind notes Genie 4 worlds still only remain consistent for a couple of minutes before they forget how things started — a constraint that may matter for any use that requires sustained accuracy.