The Speakers

Genie 3 is a general-purpose world model that can generate an unprecedented diversity of interactive environments from a single text prompt. This marks a significant advance from static video generation to fully interactive simulations of worlds.

Our model is the first foundation world model that allows real-time interaction at 720p resolution and a consistent 24 frames per second. Genie 3 maintains consistency for minutes of continuous interaction, showing marked improvements in realism and coherence over previous-generation models. Furthermore, Genie 3 introduces “promptable world events”, allowing users to model counterfactuals and alter the state of the world with text prompts on the fly. Genie 3 also demonstrates a rich understanding of the world, capable of modeling complex physical properties such as water and lighting, simulating natural ecosystems, and generating imaginative fictional and animated worlds.

We believe that world models are a key stepping stone along the path to AGI. By making it possible to train AI agents in an unlimited curriculum of rich simulation environments, Genie 3 opens a new frontier for research in embodied AI and general-purpose agent development.

Prof. Finn’s efforts to endow robots with language-conditioned goals and rapid adaptation speak directly to LAW 2025’s mission of marrying language models with agent and world models: her methods show how high-level instructions can be grounded in physical control and updated on the fly.

Real-world AI systems must be robust across a wide range of conditions. One path to such robustness is genuine understanding — a model having a coherent internal model of the world. But it is unclear how to measure, or even define, understanding. This talk will propose theoretically-grounded definitions and metrics that test for a model's implicit understanding, or its world model. We will focus on two kinds of settings: one that tests implicit world models behaviorally, and another that tests them via their internal representations. In applications ranging from testing whether LLMs learn the rules of games to whether foundation models acquire Newtonian mechanics, we find that models can make highly accurate predictions with incoherent world models. Such incoherence creates fragility when a model attempts related but subtly different tasks. Building generative models that meaningfully capture the underlying logic of their domains would enable robust deployment; these results suggest new ways to assess and improve how close a given model is to that goal.

AI world models typically focus on prediction, treating planning as expensive downstream inference. Hippocampal cognitive maps suggest an alternative: representations whose primary purpose is making planning computationally trivial. I present a framework where place cell populations encode multi-scale transition probabilities through geometric structure. Inner products between neural embeddings directly represent how easily one can reach any location from another, transforming navigation into simple gradient ascent—no search trees, no rollouts needed. A time-scale parameter naturally creates hierarchical representations from fine-grained local precision to coarse-grained global connectivity. Non-negativity constraints induce emergent sparsity without regularization, while efficient recursive composition enables "preplay"—discovering shortcuts before physical exploration. I discuss implications for language models, vision systems, and agent architectures, arguing that planning-ready geometric representations—not just predictive models—are essential for flexible goal-directed behavior in AI systems.

World Model, the supposed algorithmic surrogate of the real-world environment which biological agents experience with and act upon, has been an emerging topic in recent years because of the rising needs to develop virtual or robotic agents with artificial (general) intelligence. In this talk, starting from the imagination in the Sci-Fi classic Dune, and drawing inspiration from the concept of "hypothetical thinking" in the psychology literature, I discuss several schools of thoughts of world modeling, and lay the ground of a Physical, Agentic, and Nested (PAN) world model whose primary goal is to simulate all actionable possibilities of the real world for purposeful reasoning and planning via thought-experiment, in order to perform long-term, causal, and coherent actions toward a goal (or goals), rather than optimizing short-horizon visual metrics such as frame level fidelity or motion realism. We propose a Generative Latent Prediction (GLP) architecture that builds on stateful latent space, long-term and close-loop action-conditioned latent reasoning, inference grounding over realizable world states, and training through both SSL and RL. And we present PAN, built on the GLP architecture that brings together perception, state, action, and causality within one model to supports open-domain interactable world simulation. Extensive experiments show that PAN achieves strong performance in action-conditioned world simulation, long-horizon forecasting, and simulative reasoning compared to other video generators and world models.