The full sentence in the whitepaper is: "These constraints were supposed to protect hash from GPU and ASIC implementation, but a GPU miner appeared on the scene in 2 weeks after this technology got public attention." Thus, contextually we know that his meaning in the word "protect" is "ensure they do not exist". He considers the very existence of a GPU miner a failure of the algorithm, when, in fact, a GPU miner can and should exist as long as it the performance gap is closed. Currently GPU miners are 2-3x as performant / efficient as CPU miners, and by dga's calculations they shan't exceed ~5x the performance / efficiency. Thus the algorithm has completely succeeded at what it purports to do, and has met its primary goal.
If we're nitpicking, then closing the gap would mean that GPUs are no faster than (the fastest) CPU.
A 2x-3x gap is a narrowed, not a closed gap.
That is true, but you and I are resigned in the knowledge that a purpose-built device will always be able to outperform a general purpose device, even if the cost of that purpose-built device is fiscally prohibitive. Thus, the gap cannot ever truly be closed, in the truest sense of the word. Thankfully, this is clarified somewhat: "It is appropriate that some users can have a certain advantage over others, but their investments should grow at least linearly with the power."
Holistically the take away is and should be that the performance gap needed to be reduced between CPUs, GPUs, FPGAs, and ASICs. CryptoNight delivers on that goal quite sufficiently.
Note: I much prefer Cuckoo Cycle over CryptoNight, and am watching its ongoing development with expectation and excitement. Although we do not expect to switch PoWs anytime soon, Cuckoo Cycle is on a very short list of candidates for future consideration.