This Technology Scares OpenAI (Here's Why)
Jeff Hawke, CTO at Odyssey, provides a deep dive into the emerging field of "world models"—AI systems that generate continuous, interactive simulations. He draws parallels to the "GPT-2 era" of LLMs, outlining the current state, core research challenges like coherence and control, and the immense potential for applications in gaming, robotics, and content creation. Hawke also clarifies the confusing terminology, distinguishing canonical world models from spatial intelligence and generative video models like Sora.
World models represent a new category of AI, distinct from both Large Language Models (LLMs) and generative video models. They are designed to predict potential futures, generating a continuous and interactive stream of “intelligent pixels.” Unlike generative video models that produce discrete, bounded clips, a world model’s output can be manipulated and adjusted as the simulation evolves. A developer can seed a model with an image or text and then stream and interact with the potential futures it generates.
Training, Architecture, and the “Next Future” Paradigm
The primary training data for these models is large-scale public video. The volume of video produced daily on the internet is so vast that the core challenge shifts from data scarcity (as was the case with early language models) to data selection.
Architecturally, these models are underpinned by Transformers, but the problem framing is different from LLMs. While an LLM is often described as a “next-token” predictor, a world model is better understood as a “next-future” predictor. Given the current state of the world and some context, the model predicts what could happen next when a change is introduced into the environment. This allows for querying potential futures and even testing counterfactuals based on different action conditioning.
The “GPT-2 Era” of World Models
The current state of world models is analogous to the GPT-2 era for LLMs, a period of mass exploration rather than mass commercialization. The limitations are similar to those of early language models:
- Stability and Coherence: The primary challenge is generating a continuous stream of video that remains stable and coherent over time. Early models would quickly accumulate errors and degrade into visual noise, described as “psychedelic rainbows.” Recent progress has extended this stable generation time from around 15-30 seconds to 1-2 minutes.
- Near-Future Prediction: The models are currently limited to predicting the near future.
- Prompt Sensitivity: Similar to GPT-2, the outputs can be highly sensitive to the initial prompts.
The computational profile also differs from LLMs. While LLMs are often memory-bound (requiring a large fetch from memory for a small amount of compute), world models are flops-limited. Generating a full frame involves computing a large batch of tokens at once, requiring a small memory fetch but a massive amount of computation, pushing GPUs to their maximum output.
A Taxonomy of “World Models”
The term “world model” is used inconsistently across the industry. A clearer breakdown includes:
- Canonical World Models (Odyssey’s approach): These models learn how the world evolves. Originating from robotics and reinforcement learning, they create a neural simulator that takes a learned state and an action to predict a future state. This enables dynamic, interactive experiences.
- Spatial Intelligence (e.g., World Labs): These models learn how the world appears. They focus on 3D structure and visual appearance, making them useful for integration into existing 3D toolchains like Autodesk, Unreal, and Unity.
- Proxy World Models (e.g., Sora, LLMs): Generative video models like Sora are considered a wholly separate category. Similarly, while LLMs may develop internal, abstract representations of the world, these are not true world models as they aren’t trained on detailed visual observations of the world.
Applications: From Robotics to Interactive Content
While originating in narrower domains like autonomous driving and games, general-purpose world models have a broad range of potential applications:
- Gaming: Enabling new experiences like “choose your own adventure” games with complex, hard-to-script character interactions.
- Robotics: World models can serve as “intelligence infrastructure.” They can help overcome the data-starved nature of robotics by providing a rich understanding of how the world works, which significantly improves the sample efficiency needed to train a robot for a new environment. This is distinct from behavior models (policy learning), which focus on learning what action to take in a given state.
- Interactive Media: Use cases range from interactive billboards in retail that react to shoppers to displays at live events that respond to crowd behavior.
Research Challenges and Future Directions
The two core research problems that need to be solved are:
- Coherence: Achieving a consistently plausible, continuous stream of pixels.
- Interactivity: Developing robust methods for controlling and manipulating the generated stream.
Long-term memory is another significant problem, but it is secondary to solving the foundational issues of coherence and control.
Many techniques from the LLM world are being adapted:
- Post-training: The pipeline of pre-training, mid-training, and post-training to add capabilities is directly applicable.
- Reinforcement Learning (RL): RL can be used to improve outcomes. An advantage for world models is the potential for verifiable RL, where a 3D-consistent environment can provide a computational reward signal, reducing the need for human-in-the-loop feedback.
- Retrieval-Augmented Generation (RAG): While not yet implemented, a RAG-like framework could allow models to pull from external databases of video assets, which would be useful for both long-term memory and for studios wanting to use their own content libraries.