Interview
Decart’s Dean Leitersdorf on AI-Generated Video Games and Worlds
- Product Launch & Capabilities: Descartes launched "Oasis," a fully playable AI game engine that operates without traditional game engines, allowing users to move, break, and place blocks in real-time.
- Technical Breakthrough: Oasis is the first video model capable of real-time inference and interactive response on existing H100 hardware, bypassing the industry belief that Blackwell chips are required.
- Architectural Approach: The system utilizes a "prompt-to-pixels" model where the prompt consists of user keyboard actions and previous frames rather than text, generating the next frame in real-time.
- Training Efficiency: The Oasis model converges from start to finish in 20 hours, whereas comparable diffusion models at other top labs typically require two weeks.
- Vertical Integration Strategy: The company is fully vertically integrated, optimizing from CUDA kernels and custom garbage collectors up to model architecture, enabling a 10x cost advantage and faster iteration.
- Infrastructure Reliability: Descartes built custom solutions to handle unreliable GPU clusters, including a 200-page internal document of failure modes and a custom distributed file system to overcome missing cloud storage.
- Data Processing Innovation: The team developed a small model to label gameplay recordings with user actions, enabling the training of the video model with synthetic data generated in parallel on the same cluster.
- Long-Term Vision: Dean Leiterstorff predicts the convergence of transformer-based state models and pixel-based rendering models within the next 2-3 years to combine consistency with dynamic flexibility.
- Market Definition: The company aims to transition from "UX" (User Experience) to "GX" (Generated Experience), where AI dynamically creates content matching human interaction rather than users navigating static interfaces.
- Competitive Moats: Sean McGuire argues that hardware optimization and low-level systems mastery (analogous to Google's early PageRank implementation) constitute a deeper, longer-lasting moat than application-layer features.
- Distribution Dynamics: The interview notes a new "mini network effect" driven by TikTok, where virality creates an immediate brand barrier that can persist for years before long-term tech moats solidify.
- Strategic Alternatives: The founding team considered building an NVIDIA competitor or a new cloud provider but selected the "generate experiences" avenue as the primary path to a trillion-dollar valuation.
- Rapid-Fire Insights:
- Dean Leiterstorff holds a PhD from Technion at age 23, later surpassed by his younger brother at age 21.
- He first programmed a computer at age 13 using bots for RuneScape, which resulted in a ban after 24 hours.
- The company name "Descartes" was chosen to reflect the "4 AM" state of AI, resembling Nikola Tesla, with Leonardo da Vinci as his favorite scientist.
- Forward-Looking Statements: Leiterstorff estimates it will take at least five years to realize the full potential of the "generated experiences" model, though the timeline could be shorter given the current pace of AI development.
- Content Creation Future: AI-generated games will likely precede paid AI-generated novels, as platforms enabling human creativity to generate content will emerge before AI can independently produce high-fidelity creative works.