Nvidia puts 30 years of high value knowhow in a 13B LLM (twitter.com)

They do mention that their expectation is that the 70B model will provide even better performance. I expect that you are correct and that they determined the 13B was capable enough to serve as a base model. Why incur additional training time before getting preliminary results.

It's funny how the sttng model of computer-assisted research has gone from “laughably naive about what a computer even is” to “available to anyone for 20 dollars a month” in less than two years (from the consumer perspective).