I think his argument is on shaky ground. No matter what you change the proof-of-work function to, it will always be possible to develop specialized hardware that outperforms general purpose hardware.

I accept his argument that it's dangerous for vast amounts of mining power to be centralized to a small number of actors. I don't agree that changing the PoW function would solve this in the medium term. It might work in the short term, as developing specialized hardware would take time.

I suppose one solution would be to change the PoW function at set intervals, randomly selecting one function from a large pool. This would discourage the development of specialized hardware.