LANGUAGE EQUIVARIANCE AS A WAY OF FIGURING OUT WHAT AN AI "MEANS"

Summary from the TLDR AI newsletter

A researcher identified syntax-semantics issues in LLMs and proposed language equivariance to address them. The approach suggests checking if LLMs' moral judgments remain consistent after translation. Language-equivariant LLMs could represent a deeper understanding beyond mere syntax.