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 ...
Nature

Augmented genetic decoding: global, local and temporal alterations of decoding processes and codon meaning

Alternative genetic decoding can be either a global phenomenon that affects the decoding of most or all mRNAs, or a local recoding that affects particular sites in a few mRNAs. Global alterations in ...