This confirms something I've noticed about GPT-3, or at least GPT-3 as it is trained using the public internet as a corpus...
This response reads exactly like a so-called "recipes" website, in which the writer gives their whole life story, side-notes, and wanders around over several paragraphs before finally getting to the damn recipe.
This makes me think the public internet is not the most sanitary input for training. That type of "recipe" evolved, IMO, to snatch the highest SEO rankings, adding a bunch of keywords, snippets, affiliate links, etc, instead of just giving me the text of the recipe. And now GPT-3 has learned the same SEO tricks (at least when you give it my input, which is a very click-baity opening, to be fair...)
This confirms something I've noticed about GPT-3, or at least GPT-3 as it is trained using the public internet as a corpus...
This response reads exactly like a so-called "recipes" website, in which the writer gives their whole life story, side-notes, and wanders around over several paragraphs before finally getting to the damn recipe.
This makes me think the public internet is not the most sanitary input for training. That type of "recipe" evolved, IMO, to snatch the highest SEO rankings, adding a bunch of keywords, snippets, affiliate links, etc, instead of just giving me the text of the recipe. And now GPT-3 has learned the same SEO tricks (at least when you give it my input, which is a very click-baity opening, to be fair...)