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 ...
Tech Xplore on MSN

Swarming microrobots use spinning flows to turn gears without touching

E pluribus unum—"out of many, one"—is not only a motto for the United States; it's a good credo for microrobots. A research collaboration between Cornell and the Max Planck Institute for Intelligent ...