Interview
Decart’s Dean Leitersdorf on AI-Generated Video Games and Worlds
- The company plans to launch OASIS 3, an advanced iteration of the OASIS 1 platform designed to enable real-time "magical mirror" interactions where users manipulate objects and alter environments via digital reflections.
- Gen AI is projected to bridge the gap between imagination and visual output, potentially transforming current applications and replacing traditional game engines with AI-driven systems that allow natural language-based "reactive modding."
- The industry is expected to converge on a dual-model architecture within the next two to three years, utilizing one transformer model for game state management and a second for pixel rendering, though full convergence requires additional time.
- Future platforms will include next-generation social media comparable to TikTok or Snapchat, alongside specialized simulators such as those for fighter pilots, building on the interactive video capabilities.
- The "Generate Experiences" (GX) paradigm is identified as a primary focus intended to reach every global consumer, with the company targeting a trillion-dollar scale within at least five years, though adoption timelines may extend to 10 or 15 years.
- While AI-generated creative content for consumer sale remains distant, immediate efforts will focus on platforms enabling users to generate such content, with monetization expected to mature as adoption accelerates.
- The company maintains a strategy of full vertical integration, optimizing from electrons to pixels, claiming a "10x plus advantage" and the ability to execute 20-hour training runs compared to the two-week typical industry standard.
- Vertical integration is anticipated to provide a competitive "tech mode" advantage of reaching the market one to two years ahead of peers, allowing for a move to end-to-end solutions to avoid the downtime associated with unoptimized public cloud infrastructure.
- Real-time video delivery is not expected to be feasible within the next year without deep vertical integration down to bare metal components like CUDA kernels and hardware optimization.
- Current instability in GPU clusters for AI training is projected to stabilize over the next half year as cloud providers address reliability, eventually offering stable storage and optimized nodes to reduce the need for proprietary distributed file systems.
- Hardware is positioned as the ultimate long-term moat due to the extended manufacturing timelines required for new facilities, contrasting with faster software development cycles and temporary brand effects driven by social media mentions.
- The company remains open to future development of an NVIDIA competitor and a new cloud provider, with hardware competition serving as a lingering strategic option alongside the primary focus on experiential platforms.
- Material risks include the temporary nature of "tech modes," the historical difficulty of replicating deep system optimizations, and the uncertainty regarding Google's full leverage of its hardware advantages on the application layer.
- The organization operates under the philosophy that the opportunity to overcome fundamental limitations rather than solve existing problems occurs once every 10 to 15 years, necessitating rapid execution to convert early technological advantages into lasting market dominance.