Source: raw/AI_Can_Now_Create_Worlds_You_Can_Change_in_Real_Time.md (main explainer) · raw/AI_News_-_The_Only_6_Stories_Worth_Knowing_This_Week.md (hands-on test) · raw/newsletter-theneurondaily-com-1fa0d4fd45.md (The Neuron Daily, October 28, 2026, lead story)

Creator: Nathaniel Whittemore, The AI Daily Brief; Matt Wolfe; The Neuron Daily · URL: https://www.youtube.com/watch?v=HOgj6cJ9oxQ (AI Daily Brief) and https://www.youtube.com/watch?v=eDiQNFa-zYg (Matt Wolfe) · Duration: not recorded · Platform: YouTube, email newsletter (fetched 2026-11-01)

Genie 4 is Google DeepMind’s fourth world model, announced the week of October 26, 2026. A world model generates an environment you move through with the keyboard, frame by frame, instead of a clip you watch. Genie 4 adds three things to Genie 3: you can change the world with text prompts while you are inside it, it generates sound, and a Gemini model runs alongside it to plan multi-step events. It is still a research preview with two-minute sessions, so its practical use today is quick previsualization for film and game ideas, not finished footage.

Key Takeaways

  • Live prompting is the new capability. Earlier Genie versions took one prompt up front and let you explore the result. In Genie 4 you can type mid-session to add an object or character, change the weather or guide the camera (Neuron: “add a dragon”; Wolfe: “Godzilla stomps through”).
  • It is several Google models working together. Per the AI Daily Brief reading of Google’s post, Gemini Omni (Google’s video model) “anchors the simulations in real-world appearance and behavior”, Gemini 3.8 Flash acts as “a guiding agent in the loop” that watches frames, breaks goals into steps and checks each step succeeded, and unspecified audio models add sound.
  • First Genie with audio. Ambient sound, music, sound effects and speech. Whittemore says the speech still has the “hollow and echoey” quality older video models had.
  • Better physics and people. Google claims simulated aerodynamics and collisions, and characters that “move naturally” and can be steered in real time.
  • Sessions last two minutes, double Genie 3’s 60 seconds (AI Daily Brief; Wolfe saw the session freeze at the two-minute timer).
  • Consistency still breaks. Google itself says objects, physics and spatial details “can get inconsistent, especially in longer sessions” (Neuron). In Wolfe’s test, after a 360-degree turn some objects had moved and a new cupcake had appeared.
  • Robotics is Google’s stated long-term use. Google has tested Gemini 4 Argon as the planner: it breaks a task into steps that are fed to Genie as text captions while the simulation renders, so a simulated robot can practise a task (a chess demo per AI Daily Brief; Neuron describes a humanoid playing checkers).

Hands-on (Matt Wolfe’s early access)

  • Setup screen: describe an environment and a character, optionally from a sketch, or “roll the dice” for a random world. Third-person or first-person view.
  • Controls: WASD to move, arrow keys for the camera; on-screen indicators show which key is pressed.
  • Suggested prompts appear at the side (for example “police drone”); clicking one adds it to the scene a few seconds later. Typing your own prompt takes over the camera until it finishes, which he found annoying.
  • Failure he hit: asking Godzilla to swat the drone froze the session with “Something went wrong.”
  • Sound: a train passing through a candy-land world came with music and effects.
  • His verdict: not a game you would play for hours, and not good enough for final footage, but useful for roughing out a scene before shooting it or testing a game idea before building it in Unreal or Unity.

Who can use it

Sources disagree on access.

  • The Neuron (Oct 28): trusted testers now; Google AI Ultra subscribers (via Project Genie) and selected Gemini API and AI Studio partners next.
  • AI Daily Brief: “Google is making Genie 4 fully available for Gemini Ultra subscribers to try out”, as it did with Genie 3.
  • Matt Wolfe: early access via a waitlist with people let in slowly.

Treat Ultra access as likely but check your account before planning around it.

Context from the AI Daily Brief

  • Lineage (as Whittemore tells it): Genie 2 was shown in December 2024; Genie 3 was previewed later and opened to the public in August 2025, adding keyboard and mouse navigation and persistence (paint a symbol on a wall, look away, and it is still there), which Google called an emergent property of autoregressive frame generation.
  • Reactions: Ethan Mollick built a “neo-Gothic, brutalist city inhabited by robots and horses” in about a minute and then added dragons and an otter NPC live: “glitchy, but also impossibly advanced compared to a couple of years ago.” Justine Moore (a16z) made a Lovecraftian game and a Grand Theft Auto pastiche.
  • Market: Genie 3’s release had hit gaming stocks; Genie 4’s did less damage, which Whittemore reads as investors judging the technology still some way from replacing games.
  • The larger point he draws: Google lags on language models but leads on multimodal work, and Genie 4 shows the value of combining several model types into one system.

Try It

  1. If you have Google AI Ultra, open Project Genie and check whether Genie 4 is enabled; otherwise join the waitlist.
  2. Use it as a previz sketchpad. Before a client shoot or an animated spot, describe the set and the main character, walk the camera through the planned shots, and screen-record the two-minute session as a reference for the crew.
  3. Script your mid-session prompts in advance. You have two minutes, and typing takes over the camera, so keep a list of the 3-4 changes you want to test (a prop, weather, a second character).
  4. Check persistence before trusting a layout. Do one full camera turn and compare; if props move, treat the session as mood reference only.
  5. Do not plan deliverables on it. Neither source considers the output good enough for final footage; keep using your normal video models for that (see the August video-model wave).

Open Questions

  • Exact access tiers and pricing are not confirmed by a Google primary source in the vault.
  • Whether the robotics demo was chess or checkers; the two secondary sources differ.
  • Whether “Gemini Omni” and “Gemini 3.8 Flash” are the exact product names in Google’s post; they come from auto-captions and the newsletter, not the announcement itself.
  • How long a session will run when Genie 4 leaves research preview.