Something is... Different About Google's Genie 3
Speakers
Summary
- Google has released "Genie 3," a World Model AI that simulates interactive reality at 720p/24fps, described as a more significant step toward AGI than ChatGPT.
- The technology allows for the generation of infinite, physics-compliant 3D worlds from prompts, threatening traditional game engines like Unity and Roblox.
- Waymo (Alphabet) has utilized this tech to build advanced driving simulators, solving "long-tail" edge cases (e.g., elephants on highways) to accelerate autonomous driving.
- The hardware requirements are immense; a single instance of Genie 3 requires four NVIDIA H100 GPUs, signaling massive sustained demand for compute during the inference phase.
Trade Ideas
Following the Genie 3 release, "Unity... dropped 30%... Roblox... dropped 13%." The speaker notes that AI engines can now emulate games in seconds, threatening the traditional 5-10 year development cycle. If a World Model can generate a playable, interactive environment from a prompt, the value proposition of complex, expensive game engines (Unity) and user-generated platforms (Roblox) diminishes significantly. The barrier to entry for game creation is collapsing to zero. Short/Avoid traditional gaming engines as their "moat" (complex physics and rendering engines) is being eroded by generative AI. Roblox explicitly announced their own world model, suggesting they may pivot successfully; the market sell-off may be an overreaction (oversold bounce).
"Genie by Google DeepMind runs on four H100 GPUs... per instance." Ejaaz notes that compute is a "never-ending resource that we will need." Unlike LLMs which are training-heavy, World Models are inference-heavy. Every 60 seconds of generated video requires massive, real-time GPU power. As World Models scale to consumers and robotics, the demand for H100-class chips shifts from "training clusters" to "always-on inference," drastically increasing TAM. Long NVIDIA as the pick-and-shovel play for the high-compute requirements of simulating reality. Supply chain bottlenecks; potential development of more efficient, non-GPU hardware (TPUs) by Google itself.
"Tesla's been using world models to train autonomous driving models... it is statistically better than the average human driver." The validation of World Models by Waymo (Google) confirms the specific architectural path Tesla has taken. As World Models mature, the ability to simulate infinite edge cases (training data) solves the final hurdle for Full Self-Driving (FSD) regulatory approval and safety. Long Tesla as a beneficiary of World Model technology applied to real-world robotics and autonomy. Competition from Waymo (which uses higher fidelity sensors/Lidar mapped into Genie 3); regulatory delays.
Ejaaz Ahamadeen
Co-Host, Limitless Podcast (Bankless)
Google DeepMind released Genie 3, which "simulates reality" and understands physics, allowing users to generate interactive worlds. Waymo used it to build the "most advanced driving simulator ever created." Google is transitioning from text prediction (LLMs) to reality simulation (World Models). This secures their dominance in two massive verticals: AGI development and Autonomous Driving (via Waymo), solving the data scarcity problem for training robots and cars. Long Google as the leader in World Models, which serves as the foundational software for the physical AI revolution. High inference costs (requires massive compute) could impact margins; regulatory scrutiny on AI-generated content.
