The funniest thing I saw this week was this (excerpt from this gist:

Title: CINNAMON COOKIES ------------------------
 8 oz Piece of trout of milk (1 c.)
      Oil
      Sour cream
      -cut into 1/3 of the salt
 6    Cloves garlic, minced
 4    Small beef broth
      -deveined
      Salt & pepper
 3 c  Strawberry cooked rice
 1 ts Salt
 2 tb Cornstarch
 1 ts Salt
 2 c  Powdered sugar
 2 ts Ground coriander

Preheat oven to 350. Place on lightly floured milk in skillet. Drain off each piece and
remove from the warm water to coat. Serve on a
spray lined, you adaptive and around the toppings to the mixture as possible.
The coctunday recipe.  In a large skillet, over medium high
heat, stirring occasionally, until meat is moistened.  Let rise
approximately 4 minutes.

Remove from heat and cook over moderately high heat for about 7 minutes.  In a small saucepan,
covered, for 6 to 7 minutes, add corn in the oil.  Add the lemon juice
and water in a double boiler or skillet. Spoon in the flour mixture
to a boil. Drain the fruit.  PLACE STEW ONCOAT. WITH COOKIES
WITH THE LOW THE PAN. PER SERVING: 77g; PRO: 4g; MC
: From oven the chilies.

I saw this in class when we were going over Recurrent Neural Networks (RNN). Basically, you shove a large corpus of example text in one end and out the other end you get a model of that text. This model encodes properties of character probabilities given the previous characters. You can feed it a starting word and it can repeatedly spit out another likely character. I like to imagine a monkey hammering away at a keyboard, kinda like this:

via GIPHY

And this monkey is remarkably good at creating things which resemble recipes. For a more in depth dive into RNNs, I highly reccommend this which is also where I got the implementation of what I will use below.

Ask HN: Who is Hiring? (<Insert Month Here>)

I found the recipes so hilarious that I had to make more of them. Since I’m in the process of finding a job, I am very familiar with HN job posts. After reading so many of them, you get a sense for certain trends which get a bit tiresome. To bring some cheer, I’ve unleashed the power of RNNs on 3ish years of posts (back to 2013 though I’m missing quite a few). This, yields roughly 13 million characters.

As I mentioned above, I used the ready-to-bake implementation from karpathy which was super nice. It uses the torch framework which actually has an Amazon EC2 Machine Image (AMI) ready-to-launch here with all the CUDA packages etc. This meant I was training my RNN in a few minutes on on a g2.2xlarge which is 8 cores and a GPU with 4GB of memory. I trained a network with 256 hidden nodes per layer with 3 hidden layers. This fit in only 500MB of GPU memory. While I could have gone bigger and gotten better results, I didn’t want to spend a ton of money. I stopped it after about 8 hours of training (I’m cheap) and then sampled some text. It’s not as clean as I would have hoped, but some of them are kinda funny. You can find the sampled output and the input here.

Visually, the generated postings look like they should, with all the right formatting and even usernames and X days ago before some of them. Overall, the English has a bunch of mispellings and the sentences are not very coherent. Here’s my favorite one:

borts 277 days ago  

GitHub - http://careers.stackoverflow.com/jobs/77213. New York City,
Crypark and International Software Developer. Ramper healthcare in Santa
Barbara, and MahoutOperaction (http://aclima.io) -
http://smarkets.com/labs/risk/why-perks-clime/WingifESignaltw
[similar-wifith-get-stakemonacide.comPangea Recruiting / email Carouff
experience is a plus, competitive pay).Sorry "where we can see and performance
and resident price, performance and evaluation in a small startup organization
in designs, and competitive positions with experience in large scale
development, experience near conversation and general products that have a
specializing enterprise data and manufacturing mobile development where we
spend our various politics mobile development and enhancements to understand
hardware of paid products are innovative hardware, as well.  And we’re looking
for someone who can progress and e-commerce and test-driven data structures.

* Fulfilling systems and lead

* * Senior software engineer, you have experience managing and experience
learning and more customers* Strong contractor work for the folks of
online edge of computer vision system in a manufacturing in just the best
infrastructure.  More about you up and espons the team 

 *

I’d like the revisit this when I have a personal machine with a GPU (or 2… or 3).