News for hackers. Stuff that matters.
Re:The Unreasonable Effectiveness of Recurrent Neural Networks by myth_buster on Thursday May 21, 2015 @04:10PM (#9585116)
This is quite incredible. The stylistic similarities of generated Shakespearean saga, Linux code etc was quite startling. Perhaps we can train a Haiku/Fortune cookie generator which could occasionally be quite profound.
Re:The Unreasonable Effectiveness of Recurrent Neural Networks by seiji on Thursday May 21, 2015 @04:56PM (#9585338)
> Linux code etc

People are always worried about "computers taking factory jobs" resulting in mass unemployment, but the truth is, a rudimentary AI with acceptance tests on output will obsolete every programmer alive.

Hell, half the programming people do these days is just gluing APIs together then seeing if it actually works. It doesn't take 16 years of rich inner human life experience to accomplish that, just exhaustive combinational parameter searching on the subset of API interactions you're interested in evaluating.

Re:The Unreasonable Effectiveness of Recurrent Neural Networks by myth_buster on Thursday May 21, 2015 @05:59PM (#9585637)
Douglas Crockford touches on this aspect in this entertaining and insightful talk [0]. I'm guilty of what you state and I think a large part of "programming" is rudimentary boiler plate coding/configuration and staring into the Abyss. I think our role will be to design algorithms and come up with creative solutions/hacks (which would be difficult for a program) and designing a workflow/flow chart and feeding it into a program which spits out binaries and flag for edge cases. A whole swat industries and economies (read outsourcing) will become redundant and only outsourcing done would be to the generator.

[0]: https://www.youtube.com/watch?v=taaEzHI9xyY

Re:The Unreasonable Effectiveness of Recurrent Neural Networks by raverbashing on Friday May 22, 2015 @05:21AM (#9587264)
Who do you think will write the acceptance tests? (to be honest they're sometimes more complex than the code itself. E.g. write the acceptance tests for x=a/b for a and b as inputs )

I'm all for it, it's going to be a productivity gain. It's like going from a manual screwdriver to a motorized one.

Re:The Unreasonable Effectiveness of Recurrent Neural Networks by speechduh on Thursday May 21, 2015 @04:22PM (#9585168)
That particular stuff is actually pretty typical. I have a textbook that shows similar results on Shakespeare using N-grams from years ago.
Re:The Unreasonable Effectiveness of Recurrent Neural Networks by gipp on Friday May 22, 2015 @01:32PM (#9589774)
Capturing writing style with ngram-based input and individual-character input are very, very different tasks. That's several ballparks higher in difficulty.

With ngrams, Markov models are perfectly sufficient. With individual characters, complex concepts need to be remembered across many, many characters of input.