Human-AI generated pizza | Hacker News Reader

Human-AI generated pizza

howtogeneratealmostanything.com

34 pointsmatt_the_bass12 comments

A char-RNN is a bad way to go about it. The results are effectively random - there's a ton of structure in recipe-space, but forcing it all through the bottleneck of text (and then sampling!) is going to hide a lot of it. You want something like "Sukiyaki in French style: A novel system for transformation of dietary patterns" https://arxiv.org/abs/1705.03487, Kazama et al 2017's `word2vec` (using structured recipes like https://open.blogs.nytimes.com/2015/04/09/extracting-structu... ); or IBM Watson's recipe generation ( https://www.eater.com/2014/2/26/6272773/food-truck-is-actual... https://www.newyorker.com/magazine/2016/11/28/cooking-with-c... ).

I think the intent (and this blog in general) is more for fun rather than an academic exercise.

If you used better approaches, the recipes sampled would be a lot more fun to evaluate and discuss. With generative stuff, there's a sort of inverse uncanny valley: it's only when it gets very close to perfect that it's really interesting to look at, because that's where you get the most creativity and reactions like 'I can't believe that worked' or 'maybe...'. For example, I find that sukiyaki paper a lot more fun to read than OP, because the recipes are meaningful rather than random gibberish. ('cranked red onion' is not actually amusing, it's just gibberish).

But overall it _did_ capture most of the structure. Other than a few adjectives, everything it required was real, and most of the combinations were exotic but good.

However, thanks for sharing the Kazama paper, that looks really interesting.

Word2vec approach would also create pizzas that follow the traditional ingredients pairings

Recipes aren't dense. If you sample ingredients and then random neighbors, you'll get new recipes. You can also do much more like 'style transfer' for recipes (which is the point of the paper I linked), which are definitely not traditional.

You have roughly 3 main toppings per pizza before it becomes a cluster-fuck, maybe 50 ingredients if you're stocked. There are 125k combinations, half of which obviously won't sell, 49% of which might sell as a novelty, with the remaining 1% being the combinations people mostly know about.

If the menu has 20 new combinations per day, I'm having a hard time seeing why an AI would be better able to derive a new list per day than a motivated high-school cooking class is.

If a mechanical robot had some precision, I'd be more interested if an AI was trained to place the ingredients in optimal locations (r, theta) on the pizza, so as to maximize enjoyment and minimize cost.

Most of those sound like a recipe for going out of business.

Some of those recipes looked great. I wish I wasn’t gluten intollerant.

It’s definitely possible to get GF-pizza, the dough is a little strange (dryer and chewier, usually). My girlfriend is gluten intolerant too and we live somewhere without too many exotic food ordering options but there’s a number of pizza places that offer a gluten-free base. Ask around, you might be surprised!

This kills the Italian

TL;DR - it's canadian bacon and pineapple, thin crust.

Reply on news.ycombinator.com