← Back
technique
JEPA
Last discussed Aug 8, 2026
Joint Embedding Predictive Architecture - a training technique developed by Yann LeCun for world models.
Recently discussed in
Why Robotics Still Isn't Solved - But Could Be Soon | YC Paper Club
Guanming Wang · Aug 8, 2026
Are LLMs a Dead End? (Investors Just Bet $1 Billion on “Yes”) | AI Reality Check | Cal Newport
Cal Newport · Mar 26, 2026
What creators say
GW
Guanming Wang
mentioned
Used for: Learning to predict future observations using MSE loss between current and future latent encodings
vs: Contrasted with distribution-based approaches that treat the future as a dist...
"Some people, like LeCun, they were doing JEPA. They think about maybe we can have two different encoders — one encoding the current observation and the other encoding a future observation. By doing loss on the current observation latent and the fu..."
More from Guanming Wang →
Cal Newport
mentioned
Used for: Training world models with high-level representations
"If you've heard of JEPA, Joint Embedding Predictive Architecture, this is a new training technique that LeCun came up with for training a world model"
More from Cal Newport →Keep exploring
Topics
About this data
- ✓45,000+ mentions extracted from real podcast transcripts
- ✓Every quote links to the source episode
- ✓Sponsored mentions are always labeled
- ✓Critical takes are shown, never filtered out
Get an email when a new creator recommends it