aduffy

Why not just use one of the Common Lisp flavors then? Racket has been rising in popularity last few years: https://racket-lang.org

Weld is a project of the Stanford Future Data Systems Lab. It's a project by Matei Zaharia, who also created Apache Spark and Apache Mesos.

Encoder-Decoder models are a very common technique in sequence-to-sequence models for deep learning. These models have had very big wins lately in NLP tasks such as translation, POS tagging, dialogue generation, etc.

The Tensorflow documentation has an okay writeup about seq2seq models: https://www.tensorflow.org/tutorials/seq2seq

The author of the library also has a small blurb about it on his blog: http://www.wildml.com/deep-learning-glossary/#seq2seq

There is already a seq2seq in the tree under contrib, is this one different from/replacing it?

https://github.com/tensorflow/tensorflow/tree/master/tensorf...

Great, just another thing I need to delete

My favorite: the new `object` type. Non-nullable and cannot refer to primitives.

The slide is taken very hard out of context, which is easy to see if you scroll one slide to the left. It's a commentary on how the software industry as changed as a whole since the suit and tie era.

The rest of the deck, however, creeps me the fuck out.

Do you have any details on your experience training the models? Was this using vanilla RNNs, LSTM cells, did you use an attention mechanism, etc.? How long did you train for?

Interestingly enough, this code is replicated in Scala Native as an example for use of the language superset[1].

[1] https://github.com/scala-native/scala-native-example/blob/ma...

-Xverbose:class JVM arg could also have sufficed in case you don't want to use strace.

Are they simply taking this out of beta? I've been using paper for over a year now, it's a lot like Quip but integrated with Dropbox

You can now have a website secured by a certificate issued by a Google CA, hosted on Google web infrastructure, with a domain registered using Google Domains, resolved using Google Public DNS, going over Google Fiber, in Google Chrome on a Google Chromebook. Google has officially vertically integrated the Internet.

The Algorithm Design Manual by Skiena.

Each section contains a story of some situation he was in where he faced a problem which he solved by applying one of various algo techniques (DP, divide and conquer, etc.). After reading CLRS for a class, it was nice to see how some of the most common textbook algorithms have been applied by a notable computer scientist.

If you maintain a large open source project during work hours then that arguably is "sponsorship". Looking at the Keras commit log it seems he pushes frequently.

Cool, it smelled like MC :)

Curious if the poorly written prose was generated by AI or a person. Funny stuff though!

The difference is that it's not dealt with entirely in the compile phase. i.e. language features that are checked at compilation and known to be true that are not needed at runtime.

The Haskell `newtype` example I gave was meant to illustrate this, as newtype's are respected by the type system and then are treated as the underlying type at runtime.

This is actually a very helpful comment. I used to think "zero-cost" meant "at compile-time", as in `newtype` in Haskell, etc. I'm guessing that's what the parent commenter thought as well, and I'd guess is what most people think when they hear the phrase.

Not exactly a book, but Yale CS hosts a set of notes from a course of theirs, almost 500 pages.

http://cs-www.cs.yale.edu/homes/aspnes/classes/465/notes.pdf

I wrote the comment before the parent made their edit so the logic did match what it was at the time.

Having said that it would've been a very simple extension to make it a recursive algorithm which includes the else case, and arguably would be clearer than the for loop solution.

With Haskell this isn't an issue because lists are evaluated lazily, but what I was trying to get at and failed to make clear was that other languages which allow functional-like behavior can result in unnecessary allocation. For example in Python doing multiple maps/filters will allocate multiple lists unless you use the itertools versions in Python 2.7.

I could imagine the above being expressed very concisely and elegantly in a language like Haskell:

    f = sum $ map (\x -> if x == 'a' then 2 else 0)

I tend to think that as the accumulation condition gets more complicated, the for loop begins to be the clearer choice, but of course there's always a way to simplify into a number of map steps. This can lead to a bunch of allocation overhead for the intermediate lists unless iterators are used in languages where those are supported.

EDIT: Whoops, read and replied on mobile, missed the parseLine() call :/

EDIT EDIT: Looks like my code matched the parent as of when I read it, before the edit. wyager is correct that I messed up my point-free syntax

HN: come for the links, stay for the 20+ replies about the benefits of mirror writing.

I would argue that "a couple people I follow on Twitter said it was cool" is not good criteria for investment. Then again, the post didn't mention any information about how serious the reach-outs were. I'm guessing they may have been introductory emails rather than the "HERE TAKE MY MONEY" messages the author seems to imply.

"Hmm"

Smirks, puts on tinfoil hat

"Doesn't look like anything to me?"

I'm guessing it has to do with the fact that you can build go binaries not linked with libc, they also expose syscall.RawSyscall anyway that all the other syscall wrappers call out to last time I looked

That's a good point, I should've included that in the post

And still no code for "close encounter", tsk tsk

This is happening on Chrome 55 dev channel build FWIW. Anyone else experiencing this?

What's wrong with exporting to csv? Seems like it's supported: https://github.com/apache/spark/blob/master/sql/core/src/mai...