> Wasn't a key enabler of early transitor work that required capital investment was modest?

They were simple in principle but expensive at scale. Sounds like LLMs.

Is there SotA LLM research not at scale?

My understanding was that practical results were indicating your model has to be pretty large before you start getting "magic."

It really depends on what you're researching. Rad AI started with only 4m investment and used that to make cutting edge LLMs that are now in use by something like half the radiologists in the US. Frankly putting some cost pressure on researchers may end up creating more efficient models and techniques.

NN/ai concepts have been around for a while. It is just computers had not been fast enough to make it practical. It was also harder to get capital back then. Those guys put the silicon in silicon valley.