Researchers baked 3x inference speedups directly into LLM weights — without speculative decoding

Researchers from the University of Maryland, Lawrence Livermore, Columbia and TogetherAI have developed a training technique that triples LLM inference speed without auxiliary models or infrastructure ...

Content scoring tools work, but only for the first gate in Google’s pipeline

Google’s first-stage retrieval still runs on word matching, not AI magic. Here’s how to use content scoring tools accordingly ...