

Google DeepMind Lead Researchers on Genie 3 & the Future of World-Building
448 snips Aug 16, 2025
Join Jack Parker-Holder, a research scientist specializing in generative models, and Shlomi Fruchter, a research director focusing on AI in gaming, as they dive into the groundbreaking Genie 3. They discuss how this model creates fully interactive worlds from text in real-time and its revolutionary 'special memory' feature. The duo explores the evolution of generative models, the potential in gaming and robotics, and the future advancements, including Genie 4 and 5. Their insights make for an exhilarating glimpse into the future of AI-powered experiences.
AI Snips
Chapters
Transcript
Episode notes
Real-Time Worlds From Text
- Genie 3 generates fully interactive, persistent worlds in real time from simple text prompts.
- The team combined lessons from prior projects to achieve a leap in capability and interactivity.
Memory Was A Targeted Breakthrough
- The team planned for long-lived memory and pushed conflicting goals (resolution, realtime, memory) together.
- Hitting a minute-plus memory in real time was surprising even to the researchers when it worked well.
The Paint-On-Wall Moment
- Jack recounts staring at a paint example and doubting the model until he froze frames to confirm persistence.
- He says seeing the same painted wall return after looking away was a genuine "whoa" moment.