Yann LeCun talks about what's missing in large language models and about his new joint embedding predictive architecture may be a step toward filling that gap.