Robots downloading Skills (Like in the Matrix!)
AI Summary
This video is from Machine Learning Street Talk featuring Dr. Maxwell Ramstead and Jason Fox discussing the creation of a physical AI company named Numinal and the shortcomings of large language models (LLMs) for embodied AI. They explain the importance of embodiment—that cognition is deeply connected to the body’s interaction with the physical world, and AI systems need situated experience to truly understand and act in the real world. The speakers critique current AI’s confinement to “data space” without real-world grounding, likening it to Plato’s cave allegory where only shadows of reality are observed. They elaborate on how language is a compressed, abstract representation far removed from direct worldly experience, making language models inherently limited without embodiment. The conversation covers philosophical notions of cognition, embodiment, and emergence, emphasizing that intelligence is not just a computational process but an embodied, situated activity tied to interaction dynamics with objects and environments. They discuss technical and business challenges in building physical AI systems that can actively learn and generate new data through exploration, highlighting the need for compositional, modular AI models that can adapt dynamically. The talk also describes the importance of feedback loops, real physical testing, and the limitations of current large models when deployed in robotics. Ultimately, they advocate for AI architectures inspired by biological brains, leveraging active inference to create embodied intelligence that goes beyond large language model capabilities, aiming to deploy AI safely and effectively in real-world physical applications.