Aug 16, 2025 · 42m · a16z
Google DeepMind Lead Researchers on Genie 3 & the Future of World-Building
Jack Parker-Holder · 17m spoken Shlomi Fruchter · 12m spoken Justine Moore · 2m spoken Erik Torenberg · 17s spoken
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of The a16z Podcast, Google DeepMind researchers Shlomi Fruchter and Jack Parker-Holder discuss Genie 3, a groundbreaking real-time foundation world model capable of generating interactive 3D environments from simple text prompts. They explore its architectural breakthroughs, emergent physical simulation capabilities, persistent spatial memory, and broad potential applications ranging from creative software to robotics and embodied AI.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The host as informed peer 2.7 Guest teaching 3.6 Guest disagreement 0.4 The host pushing back 0.1
speaking balance: gold is the host, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 13:45 Rejecting LLM Reasoning Premise
Shlomi politely rejects Marco's premise that scaling Genie creates emergent LLM-style reasoning, clarifying that world understanding is fundamentally distinct.
Hardest push from the host ▶ 20:48 Challenging Genie 3 vs. Veo 3 BrandingHost Mike directly presses the guests on why the project wasn't branded as Veo 3 Real-Time given the overlap in capabilities.
Biggest teaching moment ▶ 11:33 Explaining Generative vs. NeRF RepresentationsShlomi educates the host on why avoiding explicit 3D geometric representations like NeRFs was crucial for maintaining spatial generalization across generated frames.
The host holds their own ▶ 32:50 Citing Hassabis on SIMA Agent ComposabilityHost Mike demonstrates strong background knowledge by referencing Demis Hassabis's recent remarks on combining SIMA agents with Genie world models for robotics training.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The host as informed peer | Guest teaching | Guest disagreement | The host pushing back | Why |
|---|---|---|---|---|---|---|
| Genie 3 World Generation Capabilities Highlight | 2 | 3 | 0 | 0 | Marco asks about interactive memory and data collection strategies. Jack explains how Genie 3 synthesized progress from Genie II, GameNGen, and Veo II. | |
| The Magic of Real-Time Control and Low Latency | 2 | 2 | 0 | 0 | Justine highlights the perfect release timing relative to viral X/Reddit trends and asks about primary use cases. Shlomi emphasizes that core generative world capabilities underpin all applications equally. | |
| Reinforcement Learning Origins and Unlimited Environments | 3 | 4 | 0 | 0 | Mike points out a specific visual detail in the blog post regarding paint persistence. Shlomi educates the hosts on why they avoided explicit 3D representations like NeRFs in favor of pure generative frame-by-frame modeling. | |
| Memory Duration Trade-offs and Real-Time Latency | 2 | 5 | 1 | 0 | Marco asks if scaling Genie brings emergent LLM-like reasoning. Shlomi gently corrects the framing, noting that while world understanding improves, 'reasoning' is not the accurate term for this architecture. | |
| Simulating Environmental Physics and Unlikely Prompts | 3 | 4 | 0 | 0 | Justine notes specific physics dynamics like character swimming. Shlomi explains the technical tension between world consistency and following unlikely user text prompts. | |
| Direct Text Controllability and World Generation | 3 | 4 | 1 | 1 | Mike asks why this was branded Genie 3 instead of Veo 3 Real-Time. Shlomi highlights functional differences, such as interactive navigation and lack of audio, justifying separate research tracks. | |
| The Convergence vs. Divergence of Generative Modalities | 4 | 4 | 1 | 0 | Mike sets up a high-level technical question regarding whether video generation and world models will converge or diverge. Jack and Shlomi break down why separate technical trade-offs make distinct models optimal for now. | |
| Research-Driven Innovation Versus Downstream Use Cases | 2 | 3 | 0 | 0 | Justine asks whether research direction is guided by downstream applications or raw technical exploration. Shlomi clarifies that pushing capability boundaries comes first, with downstream applications following organically. | |
| Future Horizons: Multiplayer Worlds and Experiential AI | 2 | 3 | 0 | 0 | Marco proposes potential future directions like multiplayer world models. Shlomi discusses broader future horizons including therapeutic simulations for phobias and public speaking. | |
| Transforming Robotics and Embodied AI Training | 4 | 5 | 1 | 0 | Mike cites Demis Hassabis's recent comments regarding SIMA and Genie agent composability in robotics. Jack provides a deep dive on why generative world models bridge the sim-to-real gap far better than standard lab simulations. | |
| Developer Access Timelines and the World Model S-Curve | 3 | 3 | 0 | 0 | Justine introduces the framework of S-curves to evaluate where world models sit relative to LLMs. Jack and Shlomi contextualize current progress and answer a humorous closing question about living in a simulation. |
Statements from this episode (18)
Parker-Holder: Genie 3 outputs look real to non-expert viewers
“And it, it's at the point where, like, a human who is not an expert, like, will watch it and think it looks real, right?”
Fruchter: DeepMind achieved real-time interactive environment generation
“We weren't sure, like, how, how big it's gonna be but we definitely felt, I felt definitely that we have something that was kind of like for a long time coming, basically being able to generate environments in real time. I think a lot of work that was done in,…”
Parker-Holder: Veo 2 launched one week after Genie 2
“So we had this Genie II project that was much more sort of like, three environments that it could generate, and it wasn't super high quality. It felt like it coming from Genie I, but it wasn't the same quality as things like VO II, which you know, the Saints o…”
Parker-Holder: DeepMind targeted minute-plus spatial memory for Genie 3
“And then for Genie three, we basically went much more ambitious on the same sort of approach, right? And we made it like a headline goal for ourselves. It's like, can we make the memory be what it is, right? We said we want minute plus memory and real time and…”
Fruchter: Genie 3 generates environments without explicit 3D representations
“One thing that we didn't want to do, and we didn't want to build an explicit representation, right? So there are definitely methods that are able to achieve consistency, and they did that through an explicit, some three, you know, their nerves, and other metho…”
Fruchter: Genie 3 currently caps spatial memory retention at one minute
“There was not like, there is no like fundamental limitation, but Currently the current design we limited to one minute of this type of memories.”
Parker-Holder: Genie 3 terrain physics emerge from scale, not explicit logic
“And I think that that really is the property of scale and breadth of training. So this is very much like an emergent thing. I don't think there's anything Like really specific we do for this, right?”
Parker-Holder: Genie 2 suffered transfer issues from image prompting
“So I think that that's actually a really important capability that we didn't have with Genie too as well, right? Because we relied on image prompting. And so, there was some transfer issue, like, where you rely on imagine to generate the image, and that often …”
Fruchter: Veo cannot currently navigate environments or take actions like Genie
“Genie allows you to navigate environment and then maybe take actions. And that's not something that Veo at this point can do.”
Fruchter: Genie 3 is a research preview, not a commercial product
“Gini Free is pretty much a research preview, right? It's not something we are releasing at this point.”
Parker-Holder: DeepMind kept Veo 3 and Genie 3 as separate projects
“We obviously made a choice that we want VO three and Gini three to be separate projects this year, right?”
Parker-Holder: Veo 3 achieves higher visual quality than Genie 3
“VO three is clearly a higher quality threshold than Gene three, right?”
Parker-Holder: World models are the fastest path to embodied AGI
“I still think, honestly, for my, what I'm excited about for AGI, and which is more embodied agents I really believe this is the fastest path to getting these agents, like, in the real world”
Fruchter: Current world models remain far from accurately simulating reality
“I think they are very far from actually simulating the world accurately and being able to do, to kind of put a person in there and then do whatever they want.”
Parker-Holder: Genie 3 is an environment simulator, not an autonomous agent
“We designed it to be an environment rather than an agent, right? So Genie three is very much like an environment model. Like we don't see it as like an agent itself that can like think and act in the world. It's more just a general purpose sort of simulator in…”
Parker-Holder: Top robotics simulators still face significant sim-to-real gaps
“The robotic simulations are even the best ones, and we have some of the best ones at DeepMind and Majoko, right, which we work with. They're still quite far away from the real world, right? And so you have the sim to real gap.”
Parker-Holder: Genie 3 combines real-world data with simulation for robotics
“Really what we think with Genie three is it's the best of both, right? Because you're taking a real world data driven approach, right? But then you've got the ability to learn in simulation. So It kind of combines the good parts of each of those. And so that's…”