LeCun separates joint embedding architectures from predictive ones
Yann LeCun, Turing Award-winning AI researcher
LeCun explained the mathematical losses behind joint embedding methods and why a predictor helps on asymmetric tasks, and answered critics of weak results in papers using the name.
In the Daily Brief of 9 November 2026
Yann LeCun set out the difference between Joint Embedding Architectures and Joint Embedding Predictive Architectures in posts on X. He described the mathematical losses involved and traced the family back through Siamese networks to Barlow Twins, VICReg and DINO.
His central technical point concerned the predictor. A predictor, he argued, is useful when a task is asymmetric, as temporal prediction in a world model is.
He also answered a recurring criticism: that many papers titled with Joint Embedding Predictive Architecture report weaker empirical results than well-engineered systems such as DINO or V-JEPA. LeCun drew a line between foundational research on a method and the downstream engineering that scales it,.