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 ...

Testing the Unpredictable: Strategies for AI-Infused Applications

A recent SD Times Live! Supercast shed light on practical solutions to stabilize the testing environment for dynamic AI applications.
The Debrief

All-domain Anomaly Resolution Office Hosts Private Workshop with Civilian Researchers, Universities, and Government Agencies

A recent AARO report outlines new standards for data collection, AI use, privacy, and civilian collaboration in advancing UAP research.