You can constrain the solution space all you want, but if you don't have a method to come up with possible solutions that might match the constraints, you ll be just sitting there all day long for the machine to produce some results. So intelligence is not "just knowing which constraints to apply". It is also the ability to come up with solutions within the constraints without going through a lot of trial and error...

But hey, if LLMs can go through a lot of trial and error, it might produce useful results, but that is not intelligence. It is just a highly constrained random solution generator..

I believe that's I and the paper are both saying as well. The LLM is pure routing, the constraints currently are located elsewhere in the system. In this case, both the constraints and the motivation to perform the work are located in Knuth and his assistant.

Routing is important, it's why we keep building systems that do it faster and over more degrees of freedom. LLMs aren't intelligent on their own, but it's not because they don't have enough parameters