Meta essay links truth to invariance for learning models
AI-generated summary Meta published an essay arguing that learning models should judge a statement true only when it holds across relevant changes in the environment. The essay contrasts this with maximum likelihood training, which trains models to prefer what is common in the data.