You either manage that in a somewhat predictable way which ends up still being subject to being dominated by specialized hardware or you do it in a completely freeform way which is a lot of work and is a huge security risk if you make a mistake.
One does not need to switch to completely different algos with different risks to make something ASIC proof. Small changes to a variant of ASIC resistant SHA3 without compromising or changing the fundamentals of security could be implemented in a random matter. You understand that all ASIC's would be worthless if we switched to 3 rounds of SHA256(1 example of many) but a gpu would merely need to upgrade its software, right?
More details --
https://bitcointalk.org/index.php?topic=359323.0Good reading... This guy's not bad when operating within his domain.
[...] No finite collection of fixed algorithms (Even a large set) can be ASIC proof (in fact, large sets probably just lead to ASIC monopolies due to higher NRE). But if you change the POW periodically in ways which aren't predicable months in advance, and in ways that can't just be generalized with anything more specialized than general purpose consumer hardware... then I do think you would actually have achieved a fairly high degree of asic-proof-ness. There is just the question of the costs of periodic changes being worth the benefits, and what cadence is required to make investment unwise.