I agree: one thread pool for all tasks (I'm personally a fan of Intel's Threading Building Blocks, so...)

However, if you only target, say, 4 cores and you have 4 subsystems (physics, AI, rendering and gameplay lets say) and all four constantly use the CPU (ie you don't have too much downtime) then the extra complexity and overhead (moving data between cores, shared data, locking) may not be worth it over just pinning each subsystem to its own core.