A Forgotten 1949 Memo, A 1957 Map Of Meaning, And The New Proof That Separately Trained AIs Can Learn The Same World.

Researchers Randall Balestriero and Yann LeCun published a paper proving that Joint-Embedding Predictive Architectures trained separately on different modalities can learn the same world representation up to a rotation, provided they use good-view pairs of observations. The author argues that this mathematical proof validates Warren Weaver's 1949 vision of a shared basement of human language and offers practical guidance for training AI models efficiently with fewer samples. Open article →