Dispatch News of 27 October 2026 3 min
Genie 4, explained three ways
Genie 4 paints a world you can walk through and hear. Text steers it while you are inside it.
In one sentence A planner now sits inside the loop of a world model, watching every frame.
Imagine a video game where nobody built the world. You type "a beach at sunset". A computer paints it on the spot. You walk forward and it paints the sand ahead of you.
Genie 4 is a new program from Google DeepMind that does this. It came out on October 27, 2026.
This one makes the sound too. You hear the waves and the wind. People in the world can even talk back to you.
You can change the world while you stand in it. Type "make it rain" and the sky turns grey.
A second program watches the whole time. It is like a helper who keeps the story straight.
It still makes mistakes. Walk away from a tree and come back. Sometimes the tree has moved. The makers say so themselves.
Try this at home. Close your eyes and picture your kitchen. Now picture walking out and back in. Is everything where you left it? Your brain is a world model too. It is very good at that part.
Genie 4 is Google DeepMind's interactive world model. It was announced on October 27, 2026. It generates a scene in real time from keyboard input and text prompts.
Three changes stand out against Genie 3:
- Audio is native. Ambient sound, music and dialogue come from the model with the frames.
- Text steers the scene mid-session. A user can change the weather or add a person or object.
- The loop is hybrid. Gemini models wrap the generator.
The hybrid loop is the part to study. Gemini Omni anchors appearance and behavior to the real world. Gemini 3.8 Flash is a guiding agent in the loop. It observes the scene and breaks goals into steps. It reviews incoming frames to check progress. Gemini 4 Argon turns user tasks into steps shown as captions.
That is a planner wrapped around a simulator. It is the digital agent pattern from T-20.
DeepMind names the open problems plainly. Object permanence drifts in long sessions. It drifts after many prompt edits. It drifts as the user moves through the scene. This is the closed-loop drift from T-12. Physics, audio accuracy, legible text and spatial accuracy are also open.
The post gives no resolution, frame rate, session length or cost. Treat any benchmark claim as unverified until those numbers appear.
What to measure when access opens:
- Revisit consistency after one, five and twenty minutes.
- Drift after ten prompt edits.
- Delay from key press to new frame.
Every video and audio output carries a SynthID watermark.
Genie 4 turns a sentence into a world people can walk through and hear. Google DeepMind released it on October 27, 2026 as a research preview.
Access is narrow today. Trusted testers have it now. Google AI Ultra subscribers and early Gemini API partners come next.
For an enterprise the near use is rehearsal. A team could walk through a scene that is costly to stage for real. Think of a storm response or a first site visit.
The limits matter more than the demo. DeepMind says details can shift over long sessions. A training scene that moves a switch teaches the wrong habit.
Three questions decide a pilot:
- Does the scene stay the same when a trainee returns to it?
- What does an hour of generation cost?
- How will you label generated material? Every output carries a SynthID watermark.
The decision today is to watch and prepare. I would recommend to name one rehearsal use case now. Then test it the week access opens.