GAIA

The talk explains sample efficiency and presents world models as a way for AI systems to learn and plan with fewer real-world examples. It contrasts differentiable control, where known physics enables optimization, with stochastic settings such as games, driving, and robotics, where other agents and large action spaces make reinforcement learning and test-time planning costly. It surveys model-based RL, synthetic rollouts, video-based world models, and JEPA’s prediction of future states in latent space. The speakers also discuss open problems including inaccurate physics, out-of-distribution adaptation, slow planning, and the scarcity of action and tactile data for robots.

The talk introduces Manus as a general-purpose AI agent that plans complex tasks, delegates them to specialized subagents, and uses tools to complete work. It describes Manus’s use of Claude 3.7 Sonnet, Browser Use, and E2B, along with its reported 86.5% score on the GAIA benchmark. The speakers argue that combining existing models and tools can produce useful AI products, as seen in coding and legal applications. They discuss Manus’s cost, transparency, and user control alongside challenges in agent coordination and the risk that competitors can copy its advantages. The closing lesson is that AI startups need durable differentiation through hard-to-replicate evaluations, embedded workflows, or exclusive integrations and data.