The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Common Sense as an Algorithmic Framework with Dileep George - #430
Nov 23, 2020
Dileep George, Founder and CTO of Vicarious, is a trailblazer in AI and neuroscience. In this engaging discussion, he delves into the importance of mimicking the brain to achieve artificial general intelligence. He highlights the interconnection of language understanding and explores innovative frameworks like Recursive Cortical Networks and Schema Networks. The conversation also touches on the evolution of robotics, improvements in generative models, and the balance between general and narrow intelligence in AI development.
AI Snips
Chapters
Transcript
Episode notes
Human Intelligence's "Goldilocks" Assumptions
- Human intelligence relies on a "Goldilocks" set of assumptions about the world.
- These assumptions allow for broad capabilities and efficient learning, unlike current AI.
Language Understanding and Simulation
- Dileep George uses the example of "John pounded a nail on the wall" to illustrate how language understanding connects to internal simulation.
- Humans answer follow-up questions by simulating the scene, unlike systems like GPT-3.
Connecting Language, Perception, and Action
- Building true common sense AI requires connecting language understanding with perception and action.
- Perception systems must be generative and controllable by top-down language queries, unlike current discriminative systems.