AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

Google DeepMind on Genie 3: Real-Time Interactive World Models

· The a16z Show

Google DeepMind researchers Shlomi Fruchter and Jack Parker-Holder join a16z partners Anjney Midha, Marco Mascorro, and Justine Moore to discuss Genie 3, a foundation model that generates fully interactive, persistent worlds in real time from text prompts. Unlike Genie 2, which relied on image prompting, Genie 3 takes text directly as input and frames each frame individually without an explicit 3D representation such as NeRF or Gaussian splatting, an approach the team says is critical for generalization. The standout capability discussed is Genie 3's "special memory": minute-plus world persistence where objects a user has looked away from reappear correctly. Fruchter and Parker-Holder describe this as both a deliberate design goal and a surprising result when it worked, noting that conflicting objectives—real-time generation, high resolution, and long memory—were set as a deliberate technical challenge. They highlight emergent physical behaviors that emerged with scale, including realistic water, lighting, and terrain-aware interactions (characters slowing on uphill ski slopes, swimming in water, reacting appropriately to puddles). The team positions Genie 3 as an environment model rather than an agent, designed to be composable with other agents (Demis Hassabis publicly referenced combining it with DeepMind's SIMA agent). They see major applications in robotics training—potentially closing the sim-to-real gap—embodied AI for AGI, entertainment, and education. Genie 3 is currently a research preview with no public timeline for developer access, and the team emphasizes they don't yet know which downstream applications will matter most. They also distinguish Genie 3 from Veo 3, noting different priorities (Veo 3 has audio and higher cinema-quality threshold; Genie 3 is interactive and lacks audio).

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.