YOWL THE COOKIE
June 12, 2015 3:59 PM   Subscribe

Do Androids Dream Of Cooking? The following recipes are sampled from a trained neural net. Happy cooking!
posted by The Card Cheat (62 comments total) 34 users marked this as a favorite
 
Remind me again why we're teaching Skynet to cook?
posted by NoxAeternum at 4:02 PM on June 12, 2015 [4 favorites]


With enough iterations, all FPPs will be generated by RNN!
posted by mittens at 4:06 PM on June 12, 2015 [3 favorites]


Skynet taking up a hobby might alleviate the homicidal megalomania.

Also, I like how these instructions:

Preheat oven to 350. In a medium bowl, combine milk, the sugar, vanilla and seasoned flour and water and then cornstarch. add tomatoes, oregano, and nutmeg; serve.

Make you preheat the oven, but just serve a mix of cold ingredients.
posted by Joakim Ziegler at 4:07 PM on June 12, 2015 [14 favorites]


Sweet! I'm totally having word salad tonight!
posted by mikurski at 4:11 PM on June 12, 2015 [17 favorites]


mittens: "With enough iterations, all FPPs will be generated by RNN!"

Given that the human brain is also a neural network, they kind of already are.
posted by Joakim Ziegler at 4:13 PM on June 12, 2015 [1 favorite]


The pie is ready! You guys like swarms of things, right?
posted by indubitable at 4:14 PM on June 12, 2015 [18 favorites]


Put fish into flour mixture over hot oil in a medium-size. In a small bowl, combine
the garlic, salt, pepper, and the dry ingredients. Mix together the
peaches. Place another until very smooth, about 45
minutes or until the meat is tender.

posted by mittens at 4:14 PM on June 12, 2015 [1 favorite]


deep neural nets are the comedy break-outs of 2015.
posted by GuyZero at 4:16 PM on June 12, 2015 [9 favorites]


If you see a new mefi user named "1 beer, cut into cubes" that's my sockpuppet.
posted by saturday_morning at 4:16 PM on June 12, 2015 [24 favorites]


Remind me again why we're teaching Skynet to cook?

1 hooman, cut into cubic segments
sample taste
yum0001101011efficient
electron salt for ionic balance
optimal post-preparatory consumption at velocities between 0.45c and 0.68c
fucking donke... delete ramsay viral subprogram soonest

posted by a lungful of dragon at 4:25 PM on June 12, 2015 [10 favorites]


Honestly this is pretty close to what recipes sound like to me when I try to cook.
In a mixing bowl, beat butter until smooth. Drain on both sides of the refrigerator. In a lightly floured bowl, beat the egg/unsalted water 2 minutes. Mix together the water and the oil in a medium-lightly blender. Make a little and cooked or all to the cherry and ice cream, and slice the crumbs, and fresh sauce.
posted by junco at 4:39 PM on June 12, 2015 [7 favorites]


Put fish into flour mixture over hot oil in a medium-size. In a small bowl, combine
the garlic, salt, pepper, and the dry ingredients. Mix together the
peaches. Place another until very smooth, about 45
minutes or until the meat is tender.


I do a lot with medieval cookery, translations and adaptations from extant pre-16th century "cookbooks" (often actually kitcheners books). These remind me a lot of those "aid de memory" type instructions, actually.
posted by anastasiav at 4:47 PM on June 12, 2015 [3 favorites]


1 Beer -- cut into cubes

An imperial stout, no doubt.
posted by Thorzdad at 4:56 PM on June 12, 2015 [9 favorites]


In large bowl, combine liquid, and add bay leaf. Cook over low heat until flour and pour

So that's where it comes from.

Cover with lemon weed.

Will do!
posted by Rangi at 4:59 PM on June 12, 2015


Add the water and nuts, then add the water and salt in a large skillet; add the eggs, mix well and set aside. Remove the chilline for one out of the apple mashed. Sprinkle the parsley and flour and set aside. Fry the chops to a boil, and simmer for about 30 minutes, or until the mixture is coated and all the beef dispalles are done.

The matter-of-fact tone of these recipes really gets me. It's like, what, you don't know how to fry chops to a boil? Have you even set foot in a kitchen before? This is basic stuff!
posted by Rangi at 5:02 PM on June 12, 2015 [10 favorites]


When writing down your recipes for others, always include one seemingly minor mistake that will ruin the dish so when you keep making it flawlessly they wonder at your skill and why they are lacking.
posted by The Whelk at 5:08 PM on June 12, 2015 [11 favorites]


Pick a few molecular gastronomy recipes from a blog, pepper in the RNN content, and then put up a quiz to see who can guess which was written by which.
posted by codacorolla at 5:13 PM on June 12, 2015 [2 favorites]


Make you preheat the oven, but just serve a mix of cold ingredients.

Remarkably similar to the Jean-Paul Sartre cookbook, particularly this entry:

October 10

I find myself trying ever more radical interpretations of traditional dishes, in an effort to somehow express the void I feel so acutely. Today I tried this recipe:

Tuna Casserole

Ingredients: 1 large casserole dish

Place the casserole dish in a cold oven.

Place a chair facing the oven and sit in it forever. Think about how hungry you are. When night falls, do not turn on the light.

posted by mandolin conspiracy at 5:18 PM on June 12, 2015 [23 favorites]


From the HAWAILIGELED PIE:

Cut into balls.

Ow.

Reforken and dot each paper.

Reforken is my new favourite verb.
posted by mandolin conspiracy at 5:27 PM on June 12, 2015 [10 favorites]


Remind me again why we're teaching Skynet to cook?

To Serve Man.
posted by ZenMasterThis at 5:33 PM on June 12, 2015 [7 favorites]


Now I'm imagining Arnold Schwarzenegger trying to cook.
posted by JHarris at 5:36 PM on June 12, 2015 [2 favorites]


I used the same technique a few weeks ago to come up with front companies for FBI aerial surveillance.

At this moment I'm training a neural network on about 2 GB of askmefi posts.
posted by jjwiseman at 5:37 PM on June 12, 2015 [7 favorites]


If you ran these recipes past typical food blog commenters with their strange and arbitrary substitutions and alternate cooking methods you might actually come out with workable recipes.
posted by jason_steakums at 5:39 PM on June 12, 2015 [5 favorites]


Also, I was wondering this for the MTG post too, what happens if you mix corpuses? Like if you took everything from Recipes.com and threw in Poetry.com at a 50/50 ratio? I'm not at my Linux machine right now, so I can't test this out myself.
posted by codacorolla at 5:55 PM on June 12, 2015 [1 favorite]


For more examples, and for a really great blog post on the code used to do this and the general technique, see http://karpathy.github.io/2015/05/21/rnn-effectiveness/
posted by jjwiseman at 6:01 PM on June 12, 2015 [2 favorites]


Chefs Hate This RNN For One Simple Substitution To Improve Your Cooking!
  1    Medium potatoes
       Chopped onions
       -or wine vinegar
posted by johnnydummkopf at 6:06 PM on June 12, 2015 [3 favorites]


"Hubby loved it"
posted by sourwookie at 6:08 PM on June 12, 2015 [13 favorites]


Never diet again with this fat burning secret!

1 pk Gelatino, thawed
1/4 c  Fresh or salad oil
1/2 c  Corn oil
1/2 c  Buttermilk
1 tb Soy sauce
1/2 c  Plus 1 teaspoon sugar
Salt -drained
2 c  Beef coated, sliced
1 ts Olives or cocoa
Fresh orange ribs
1    Egg

posted by mandolin conspiracy at 6:24 PM on June 12, 2015 [1 favorite]


      Title: CARAMEL CORN GARLIC BEEF
 Categories: Soups, Desserts
      Yield: 10 Servings
 
      2 tb Parmesan cheese, ground
    1/4 ts Ground cloves
           -- diced
      1 ts Cayenne pepper
The only things standing between me and the fortune I can make on sales of Caramel Corn Garlic Beef are food labeling laws and the extremely small serving sizes.
posted by ardgedee at 6:42 PM on June 12, 2015 [10 favorites]


I missed you, Cake Sphere!
posted by ROU_Xenophobe at 6:59 PM on June 12, 2015 [1 favorite]


Title: ICE ENCION COOKIES
Categories: Desserts, Germanaus
Yield: 4 Servings

2 tb Olive oil
2 tb Flour, soy sauce
1 ts Vanilla
Pepper; chopped
1 tb Soy sauce
2 tb Cornstarch
3 c Flour, broccoli flowerets
1 c Unsalted butter
2 ts Dry mustard
1 tb Confectioner's sugar
2 ts Active dry yeast
Sprigers or chili
1/2 c Water; crushed
3 tb Flour
1/4 c Corn syrup
6 tb Chopped fresh chopped fresh
2 tb Powdered green paprika
1/4 c Celery; peeled and
2 Eggs
2 ts Salt
1 ts Baking powder
1 c Milk
3/4 c Rum of sugar
1 c Mayonnaise
1 ts Cinnamon
1 tb Lemon juice
2 tb Butter
2 Cloves
1/2 c Chopped onion
1/4 c Freshly chopped onion
2 tb Canola oil
1 ts Salt
1/4 ts Ground cumin
1 ts Salt
1 c All-purpose flour
Crumbled beef, cut in chili
-or beef and peeled and
-with sweet red, sliced
6 oz Fresh oregano cheese
1 ts Minced fresh parsley
1/2 ts Garlic powder
1/4 ts Ground cinnamon
2 tb Cashe
1 c Chocolate chuff
2 c Flour
1/2 c Parsley
2 tb Flour
2 ts Chopped 1/8 tsp Chocolate
6 oz Margarine or butter
2 tb Baking powder
1/4 ts Salt
1/4 c Diced coriander
1 tb Baking powder
2 ts Mayonaisonetted Lemon
1/2 ts Cayenne pepper
1 Carrot; diced
1/4 c Carrot; chopped
1 ts Salt orange peel flour
1 tb Minced fresh green onion
Sauce
Salt & pepper

Combine flour, cocoa and milk.
Cook for 5 minutes, stirring often. Beat with all sides of the up and simmering well. Bake at 350\F8
depeed to directions, then immediately.



i love beef encion cookies immediately
posted by SmileyChewtrain at 7:15 PM on June 12, 2015 [7 favorites]


Pick a few molecular gastronomy recipes from a blog, pepper in the RNN content, and then put up a quiz to see who can guess which was written by which.

You can call it the Turing Taste Test.
posted by metaphorever at 7:17 PM on June 12, 2015 [7 favorites]


Title: HOT GARLIC CASSEROLE
Categories: Diabetic, Main dish, Seafood, Main dish
Yield: 6 Servings

2 c Sugar
1 tb Cream, divided
1 ts Soy sauce
2 tb Chopped fresh oregano
Salt; to taste
1 ts Vanilla extract
1 c Whipping cream
1/4 c Chopped celery
1 tb Combination;
-chops or separates; chopped
1 Garlic; crushed
3 c Water
2 tb Butter
2 tb Chopped tomato
-shellfaws
1 tb Milk
1/4 ts Ground cumin
1/4 ts Black pepper; chopped
1 ts Prepared shrimp
And cornered
-- cut in fresh mushrooms,
-salt
-Drained and cored
Salt and pepper to taste


If the recipe starts with 2 c of sugar for 6 servings, it probably shouldn't be in the diabetic category.
posted by Fig at 7:23 PM on June 12, 2015 [3 favorites]


Title: ICE ENCION COOKIES

I like this one because it has you buy 80 dollars worth of ingredients and then instructs you to make gross gluteny hot cocoa.
posted by codacorolla at 7:46 PM on June 12, 2015 [5 favorites]


Hot Garlic Casserole :
and only 1 teaspoon of prepared shrimp, hardly qualifies as a seafood dish either tbh.
posted by some loser at 7:48 PM on June 12, 2015 [1 favorite]


If the recipe starts with 2 c of sugar for 6 servings, it probably shouldn't be in the diabetic category.

Actually, I think it should, in the "if you eat this, you will get diabetes" sense.
posted by NoxAeternum at 7:54 PM on June 12, 2015 [6 favorites]


Hilarious. How exactly did they train this neural net? There's an ingredient in there, "gellohy", which even all-seeing Google doesn't know about (the only English hit was this link, and in other languages it was 2 separate words).

Kind of reminds me of GLaDOS's chocolate cake recipe - gellohy or fish-shaped ethyl benzene, same diff.
posted by Quietgal at 7:57 PM on June 12, 2015


I know so many people who don't yet know how much they need to read this
posted by feckless fecal fear mongering at 8:03 PM on June 12, 2015 [4 favorites]


Is this thing about being amused by a program that randomly selects words a new thing or has it always been around?
posted by bleep at 8:38 PM on June 12, 2015


Quietgal: this is using the same RNN implementation that was used in the MTG card generator post from yesterday.

Superficially, the results look like they could have come from simple Markov chains which have been around for a long time, but it's actually a much more sophisticated model. It works character-by-character, which is how it's able to come up with new words that aren't in the training data. Instead of just regurgitating the word frequencies that were fed into it, it can learn to mimic both detailed and large-scale structures. That's how it can generate things like fake Shakespeare with approximately the right meter or source code that's almost syntactically correct. You won't see that from something that just randomly strings tokens together.

The underlying technology (recurrent neural networks with "long short-term memory") has been around for a couple decades but was mostly of theoretical interest until quite recently. In the last few years, Google's "DeepMind" team has made several breakthroughs using deep LSTM networks for things like playing arcade games, automatically learning simple algorithms and speech recognition.
posted by teraflop at 8:51 PM on June 12, 2015 [5 favorites]


I would imagine that the recent user friendly distributions of the software, and the ease of hooking it up to Twitter has resulted in the mainstream popularity that's taken it outside of machine learning academic circles.

These things remind me of jokes written by children. The machine is 'learning' in a sense, and has a naive conception of format, but no conception of context or execution. That's where the humor comes in. We expect a Magic Card to make sense, or a Recipe to be cookable, and seeing something that fits the form but is patently ridiculous is a good way to invert expectations and create funny juxtapositions (like the idea of a single recipe using 3 cups of poppyseeds that isn't bargain bin opiates).
posted by codacorolla at 8:59 PM on June 12, 2015 [8 favorites]


ICE ENCION COOKIES

1/2 c Water; crushed
6 tb Chopped fresh chopped fresh
2 tb Powdered green paprika
1/2 c Chopped onion
1/4 c Freshly chopped onion
1 c Chocolate chuff
2 ts Mayonaisonetted Lemon
Sauce
Salt & pepper


I was giggling the whole way through and then hit the "Salt & pepper" just sitting casually at the end there and utterly lost it. I am in tears and scaring multiple cats. WHY IS THIS SO FUNNY.

On preview: thanks to codacorolla I now know why it is so funny
posted by saturday_morning at 9:02 PM on June 12, 2015 [3 favorites]


Needs dubstep.

With the discussion A Recipe, How to Produce Erotic Dreams by the underappreciated surrealist Remedios Varo.
posted by CBrachyrhynchos at 9:08 PM on June 12, 2015 [2 favorites]


Title: FLUMIL WITH PEANUT CARROT - PATTY PEANUT BUTTER STEW
Categories: Poultry, German, Casseroles
Yield: 8 Servings

1 pk Guennisin
1 c Mayonnaise
1 lb Lean bag in microwave
1 Onion; chopped fine

Saute the peas in the refrigerator at least 8 hours.

Prepare pastry and expanpie with a layer of the squares, on the salads and lightly, and set
over a slotted serving plate to cook.

In a large saucepan, combine the vegetables and blend well.


I've got the mayo and onion. You think a brown paper bag is lean enough? I called around, and nobody seems to carry Guennisin.
posted by leahwrenn at 9:18 PM on June 12, 2015 [2 favorites]


Title: GRILLED CHEESECAKE

I'm in.
posted by OrangeGloves at 9:28 PM on June 12, 2015 [1 favorite]


Literally crying. 1/2 c Microwave oil pushed me over the edge.

Parody: CHICKEN & BROWNIES, CHOCOLATE RANCH BARBECUE, CHEESE SALAD IN SAUCE SAUCE. Poetic flair: BASIL BALLS OF BEAN BREAD, BROAK PLANT BROCK STUFFED BLANT FOR 1, CARAMELLED CORN PIPS & A LAN SAUCE, the aforementioned ICE ENCION COOKIES.

MMMMM after each one. Just wonderful. I guess the AIs will grow out of this endearing toddler/benevolent alien stage pretty soon—better cherish it while we can.
posted by absqua at 11:38 PM on June 12, 2015 [6 favorites]


Title: CHEESE SALAD IN SAUCE SAUCE
Categories: Desserts, Breakfast


I like where this is going!
posted by No-sword at 11:54 PM on June 12, 2015 [6 favorites]



It's a cookbook!

posted by rum-soaked space hobo at 2:09 AM on June 13, 2015


1 Beer -- cut into
-cubes

This is when I stop drinking and call it a night.
posted by Splunge at 3:05 AM on June 13, 2015 [1 favorite]


teraflop: " source code that's almost syntactically correct"

Whoever created that could make a fortune selling real-looking but meaningless source code to film and TV, ideal for scrolling by quickly or for the heroes to find a bug or virus in.
posted by Joakim Ziegler at 3:26 AM on June 13, 2015


I would certainly comtribute to a YouTube channel that makes one recipe per week. I think it would be best if the recipes were treated with the same matronly gravitas of Martha Stewart.
posted by johnnydummkopf at 6:46 AM on June 13, 2015 [4 favorites]


Is this thing about being amused by a program that randomly selects words a new thing or has it always been around?

While it's not truly random I guess that's always been the fun of Mad Libs. I found it endlessly amusing as a kid.
posted by mandolin conspiracy at 7:38 AM on June 13, 2015


It reminds me that summer is the time for making chilled sour cherry soup, though.
posted by Wolfdog at 8:21 AM on June 13, 2015 [1 favorite]


Beat Water and boil until light brown.

Combine sugar and olive oil and oregano. Remove chops. Break off doors along with tomatoes. Bake at 350 F for 45 minutes. Serve hot.

MMMMM
Poetry.
posted by cardioid at 9:18 AM on June 13, 2015 [2 favorites]


Peeking in on the askmefi neural net, halfway though it's training (another 7 hours to go):

Under 30 9 | As an pretty weird
3 users marked this as sachenicully

I don't know too collsisting dress back? Pyrands pop in the understanding might be some pretty close drapmit to now, because more the phrasiday, and I won't have you hease serious, but the day of ability form...)

Or we'le travels users, though, so as many updates the knowkickships anyways are shirts on diet socials.

posted by azulora
posted by jjwiseman at 10:21 AM on June 13, 2015 [8 favorites]


Arnold Shwarzeneger is still trying to get to the kitchen and learn his basic culinary terms, hence his scream "Get to the chopper!"
posted by Nanukthedog at 10:38 AM on June 13, 2015 [3 favorites]


jjwiseman: "Peeking in on the askmefi neural net, halfway though it's training (another 7 hours to go):"

Tell me this is going on Projects when it's ready.
posted by Joakim Ziegler at 3:35 PM on June 13, 2015 [2 favorites]


Honestly, if I was still cooking, I would totally do a restaurant where I followed as closely to recipes generated like this as possible. Yes they are insane, and most would never see the front of house, but to have a mad scientist computer coming up with base ideas and then figuring out how to make the cooking techniques suggested workable would be awesome. You've been warned - following a fools recipe and making it work and palatable is the next true food challenge - and nirvana.
posted by Nanukthedog at 4:44 PM on June 13, 2015 [4 favorites]


Since we've had a couple of these char-rnn posts already, and this thread isn't super busy, I will presume that it's OK to stray from recipes a bit and I'll post some info on the askmefi model I trained.

I built a corpus several years ago of pretty much all of askmefi. I used about 13 MB of it to train the RNN, using the command line th train.lua -data_dir data/askme/ -num_layers 2 -rnn_size 156. I'm using the GPU on my Macbook Pro, which doesn't have a ton of memory, and that's why I didn't use a larger network, more layers, a larger sample of training text, or more unrolled timesteps.

Once I had the trained model, I generated several samples. There's a parameter called "temperature" which affects how closely the output follows the input: lower temp follows the input closely, higher temp allows for more variation. I generated 10 samples of 32KB each for temperatures of 0.7, 0.8, 0.9 and 1.0. They are at http://lab.heavymeta.org/askmefi-rnn/ and e.g. 0.7-1.html is sample 1 with temperature 0.7. There's also a 1 MB sample at http://lab.heavymeta.org/askmefi-rnn/askmefi.html (with temperature 1.0).

The training data was the raw HTML of a bunch of ask metafilter posts, all concatenated together. You can view the samples in a browser and get some of the HTML experience, even though the HTML is pretty badly formed. (I tried running the samples through tidy-html5 but it complained a lot with messages like "Error: <strop> tag unknown".)

Keep in mind that this is a character-oriented model, and it only has a memory of the previous 150 characters. That's a pretty big difficulty factor added to the problem of learning how to reproduce ask metafilter.

Here are a few samples I pulled out to give you a flavor of the results:
-----------------------------------------------------------------

by taking a computer experience.
posted by JollyWanker at 7:57 AM on December 11

And your standards, the Manhattant learn to have good are handforn. The family was a donation freeware party.
posted by jok at 12:40 AM on December 24

If you're great.
posted by filmgey at 8:38 AM on December 19

I don't know if that is the end code to grasp them to need to do that the word with the particular and approach a lot of reception. Make a perfect file in saying if I like the same context part and had to recognize the money to think that motherwarg I had to commercial understand the best price is a possibly term.
posted by cowdingeles at 9:38 AM on December 18

I know of the same times. The columning camera is the same several issues for that, and they expect the camera that are carrying our single friends more than it will say that you should be larger than a famous place in the money.

A couple of money is available more.
posted by NK HsCACs at 12:36 PM on December 24

Here's any advice, and stupid streethapes as an symper dream (smart of the IMO, but you might want to pull a good sort of a specs) on pan per of the Orlet. The 3G 35. issule supply in the family bar age, because that makes sure to use the "ekny" by Seattle (or say a book have thinking something undercord everything but spent side $aching, but a few leashments of actually, I got too back, no vacrowed our link by egg several times) shopping, or ments, things won't you set the purchase because it just brings up with at all about 25-short. This is BTW, I just can't find material.
posted by taz at 8:36 PM on December 15

Oh, I bet-...
posted by suce5rosse at 9:51 AM on December 17

The fact that hew most citiestical (di thee after powerned places, and when it comes out it way, I say).
posted by sg00lors at 6:52 AM on December 30

There are a little tiny of the smoothment on the sex massalor to make it wire do you have and memory exphament (I don't like to know the way to get OS XMD3). I've looked at home, dance when you will just read there that have a six months. So the other comments of coupling outside formaties, just tire (leave you.)
posted by obc at 9:24 AM on December 22

Often rather than it's not unarban. Printed War was worth it back to Othosbee. Avoid natural programmer with freaken, he's pretty comprehendary ( mods in the response.
posted by Grod at 4:08 PM on December 25
posted by jjwiseman at 10:25 AM on June 14, 2015 [9 favorites]


I don't know if that is the end code to grasp them to need to do that the word with the particular and approach a lot of reception.

Has anyone really been far even as decided to use even go want to do look more like?
posted by Rangi at 12:12 PM on June 14, 2015


Teraflop, thanks for the brief explanation. I didn't look at the MtG cards thread since I know nothing about what real MtG cards are like, but cooking recipes are familiar territory and these were howlers. Howl the cookie?

I peeked at your links but didn't get much from them either ... can anybody recommend a very simple introduction to neural networks and where the field is heading? It sounds fascinating but I don't have much of my own neural network left by the time I get home from work, so something on the level of a TED talk or Nova show would be good. Thanks!
posted by Quietgal at 4:15 PM on June 14, 2015




« Older Richard Pryor’s Peoria, an Online Archive   |   robbed, assaulted and left for dead in 1267 by six... Newer »


This thread has been archived and is closed to new comments