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 ...
Interesting Engineering on MSN

Quantum data teleported 19 miles across German capital with 95% peak accuracy

Researchers in Berlin have teleported quantum data across a 19-mile loop of commercial fiber ...