This is a classic "Microsoft help desk" answer: while technically correct, it doesn't really help solve the problem. Predicting via entropy distribution for complex systems is hugely more complicated than other methods, and the only places it can really work are things like the ideal gas law and rubber bands. Put together a bunch of systems capable of exporting entropy to each other and interacting, and you'll see the difficulty ramp up absurdly fast.
Since this is meta-level advice, there is no "the problem" in sight. Your criticism would seem to apply to cases not covered by my claim, to wit, cases where the phenomenological macrostate predictions are sharp even though the microstates are uncontrolled. If you're saying the cases that are covered are rare, I do not deny it.
In Zombies! Zombies? Eliezer mentions that one aspect of consciousness is that it can causally affect the real world, e.g. cause you to say "I feel conscious right now", or result in me typing out these words.
Even if a generally accepted mechanism of consciousness has not been found yet are there any tentative explanations for this "can change world" property? Googling around I was unable to find anything (although Zombies are certainly popular).
I had an idea of how this might work, but just wanted to see if it was worth the effort of writing.