41–50 of 160 posts
Well that's the thing, "describing" an idea to the point where you are explicit enough to get the actual b behavior you want, you are basically writing code. Granted, you might have to add superfluous constructs and syntax to make it fit the programming languages we currently have, but that is a different kind of problem.
"DeepCoder uses a technique called program synthesis: creating new programs by piecing together lines of code taken from existing software" - So essentially just copying others code as opposed to generating the logic on its own?.
DeepCoder works with a very simple functional-ish toy DSL to solve very simple toy problems. It doesn't copy code, because the DSL is unique - and it's so simple there's no need to copy code, and nowhere to copy code from.
What it actually does is use an RNN to speed up a dumb search through the space of all possible programs written in the DSL by "learning" which code constructs are most common.
This works surprisingly well, but it's not obvious the process is generalisable to code written in production languages to solve complex problems with explicit logic and possible thread timing issues.
(It may well be. Naively I would expect a level of complexity beyond which the improved search stops working and/or is too slow to be useful. But I may well be wrong about that.)
Anyway - it's very, very interesting research. The article doesn't come close to explaining it or doing it justice, so it's worth reading the source paper.
A nuance that is underappreciated by a lot of people / the media is the degree to which any of the AI generates X models (faces, drawings, fills in photos, increases sharpness, etc.) simply copy and paste and interpolate from the training set. Unlike general supervised learning problems, for many of the deep learning "generative" models that get posted to HN regularly, there is no objective "test set" to measure gene…
Eh, I'm sure many people copy and paste from Stack Overflow. (Not me ofc, I would never:) ) So I guess we are all out of jobs in a couple years...
This is not a new thought. The problem with it, however, is that such a description would have to be very precise or else leave room for different interpretations which in turn could lead to very different programs being generated. In particular, as programmers know, it's often the edge cases that pose the main difficulty not the typical case.
Describing what your program should do in sufficient detail will probably end up being not very far from the actual program itself.
>AI learns to write its own code by stealing from other programs I initially took this to mean that the AI learnt how to generate the source code which makes up its own program, a bit like a quine I guess.
Oh dear New Scientist. _Steal_? What was wrong with the original article title?
(From this article: http://www.inc.com/betsy-mikel/mark-cuban-says-this-will-soo...)
Earlier quoted context omitted.
Ah, but it will help us debug errors faster if we train it on a troubleshooting dataset! :) Programmer: why is my program not working? AI: Have you tried turning it off and on again?
I'm not sure if you're aware, but this has already been done, pretty much literally: https://arxiv.org/abs/1506.05869