I think later systems were at a core an attempt to implement something like high-frequency (which is a misnomer, it's more like low-latency) trading over 1960-70s tech at a scope that we still have no means to do now, in 2020s.

"trading" was ideologically prohibited term which didn't help any.

And the whole centralized approach cannot scale.

So there you have it: it can't work, please bring new ideas.

I think people could try social experiments and compare the productivity differences resulting from different management styles on LLMs instead of humans.