They also correctly identify the function of LLMs as a glue between siloed AIs. We have barely seen the beginning of that, but as the AI race continues, it seems likely that some LLM models will be created that will have less human language, and more "interop language". Where nowadays LLMs can be somewhat probed for words and relationships, we'll have zero chance to probe an LLM using tokens that are part of some made up (by the AIs) interop language. Black boxes inside black boxes.
A naive approach will be to "democratize AI", and that will surely be better than centralized AIs responding to every query... but won't solve the deepening of inscrutability.
One point made me chuckle: when they showed the graphs for cualitative jumps above certain network size. Recently someone commented about the "diminishing returns" and "asymptotic growth" of making a LLM larger and larger... but also so it was in these models: diminishing returns all the way up to a point... followed by a sudden exponential jump until the next asymptote. The truth is we don't know where the asymptotes and exponential jumps lie, we don't even have a remote hypothesis about it.