The gigapixel example could be done with fourier features which takes about a few minutes to train (on colab-like resources). Definitely still a huge improvement though (and based on more clever hashing techniques than optimization).