Understanding World Models and Self-Supervised Learning

This chapter explores the concept of world models and the training process involved in predicting future states. It also discusses the computational resources required for training language models and the importance of new concepts and architectures in achieving human-level AI.

Transcript

chevron_right

Play full episode

chevron_right

Transcript

Episode notes

This episode is sponsored by Oracle. AI is revolutionizing industries, but needs power without breaking the bank. Enter Oracle Cloud Infrastructure (OCI): the one-stop platform for all your AI needs, with 4-8x the bandwidth of other clouds. Train AI models faster and at half the cost. Be ahead like Uber and Cohere.

If you want to do more and spend less like Uber, 8x8, and Databricks Mosaic - take a free test drive of OCI at https://oracle.com/eyeonai

Welcome to episode 150 of the ‘Eye on AI’ podcast. In this episode, host Craig Smith sits down with Yann LeCun, a Turing Award winner who has been instrumental in advancing convolutional neural networks and whose work spans machine learning, computer vision, and more.

Tune is as Craig and Yann explore the intricacies of AI, world models, and the challenges of continuous learning.

In this episode, Yann delves deep into the concept of a "world model" - systems that can predict the world's future states, allowing agents to make informed decisions. The discussion transitions to the challenges of training these models, particularly when dealing with diverse data like text and images. We then discuss the computational demands of modern AI models, with Yann highlighting the nuances between generative models for videos and language.

He also touches upon the idea of the "Embodied Turing Tests" and how augmented language models can bridge the gap between human-like behavior and computational efficiency.The spotlight then shifts to pressing concerns surrounding the open-source nature of AI models, with Yann articulating the legal ramifications and the future of open-source AI. Drawing from global perspectives, including China's stance on open-source, Yann underscores the imperative for a collaborative approach in the AI space, ensuring it's reflective of diverse global needs.

Craig Smith Twitter: https://twitter.com/craigss

Eye on A.I. Twitter: https://twitter.com/EyeOn_AI

(00:00) Preview, Oracle and Introduction

(02:42) Decoding The World Model and Gaia 1

(07:43) Energy and Computational Demands of AI

(08:06) Video vs. Text Processing & True AI Capabilities

(11:17) Embodied Turing Test & Augmented LLMs

(15:38) Is AI a Threat To Society?

(25:04) Where is AI Development Headed?

(31:06) Interplay of Neuroscience and AI**

(33:33) Yann's Vision, JEPA, and Learning Challenges

(39:05) Yann's Career, AI Progress, and Challenges