Nvidia's Nemotron-Cascade 2 is a 30B MoE model that activates only 3B parameters at inference time, yet achieved gold ...
As engineering velocity accelerates, the bottleneck migrates upstream to the people responsible for understanding customers ...