Yann LeCun: We Need Predictive Architectures Beyond Generative Models

Yann LeCun, Chief of AI at Meta, explained why simply training neural networks to predict the next frame in a video falls short of true understanding. Instead, he advocates for a new approach: teaching AI to predict in an abstract representation space, not just raw pixels. This method, known as Joint Embedding Predictive Architecture (JEPA), enables systems to grasp the underlying principles of the physical world—like gravity and object permanence—by focusing on what’s meaningful and predictable, rather than every possible detail.

LeCun believes this shift will move AI beyond today’s generative models, paving the way for machines that reason, plan, and adapt more like humans.

🎥 Watch the full interview on this channel: https://youtu.be/BytuEqzQH1U

Support the show on Patreon! http://patreon.com/aiinsideshow

=====·SUPPORT·=====
PATREON: http://www.patreon.com/aiinsideshow

=====·ALL OUR SHOWS·=====
AI INSIDE PODCAST: http://www.aiinside.show
TECHSPLODER PODCAST: http://www.techsploder.com
ANDROID FAITHFUL PODCAST: http://www.androidfaithful.com

=====·CONTACT·=====
BUSINESS AND SPONSORSHIP INQUIRIES: jason(at)yellowgoldstudios(dot)com

#AI #MachineLearning #yannlecun

Видео Yann LeCun: We Need Predictive Architectures Beyond Generative Models канала AI Inside podcast