Years of reading cryptocurrency stuff has made me skeptical of articles wholly based on speculation about things that COULD happen, with no details on how to get there.

"X might do Y someday" is just the gentleman's way of saying "X doesn't do Y right now".

How would a LLN take over research? My understanding is that further improvements (once scale taps out) will rely on algorithmic improvements - ie, nontrivial math in a highly abstract and technical field that's evolving year-by-year. That's exactly the kind of thing LLNs seem really bad at.

It’s not new or will lead to some singularity any soon. We use computer to design and produce the next generation computer since the 70s and we achieved continuous improvement. And yes, the improvement seems more linear than exponential, partly because we must pursuit many goals at once. I don’t see any difference in AI development. We will continue to make improvement and at the same time we also have maintain many other objectives, thus the models will get more and more complex but the performance only get gradually better over time.