Google just published something that deserves more attention than it's getting.

It's called TurboQuant: a compression method that shrinks a model's memory footprint up to 6x — with zero accuracy loss and no retraining required. No new architecture. No bigger dataset. Just smarter math applied to existing weights.
It's easy to miss because it's not flashy. No new chatbot, no benchmark chart. But this is the kind of research that actually expands where AI can run: phones, embedded systems, edge devices — places a full-size model simply can't fit today.
This is the same direction I work in with BitFace, a 1.58-bit Vision Transformer for face recognition on edge hardware. The field keeps proving the same point: the bottleneck was never intelligence. It's memory and compute. And the biggest wins right now are coming from compression research, not bigger models.