I wonder if we're moving to one huge general model looking at a problem and JIT synthesizing a TRM to solve it. Maybe keeping a cache of recently used TRMs for efficiency.

Related: https://www.youtube.com/watch?v=AOem-S2xSqE