Program Synthesis Demo(research.microsoft.com)
research.microsoft.com
Program Synthesis Demo
http://research.microsoft.com/en-us/um/people/sumitg/pubs/mobisys13.mp4
10 comments
And a link to his homepage where you might find more work on program synthesis http://research.microsoft.com/en-us/um/people/sumitg/
I expect this will lead to a whole new category of bugs: "I wasn't precise enough in saying what I wanted, so now the code doesn't do what it should."
Interesting technology, but I doubt it will ever replace or even be used in a proper piece of code.
Interesting technology, but I doubt it will ever replace or even be used in a proper piece of code.
"I wasn't precise enough in saying what I wanted, so now the code doesn't do what it should." is not, in any sense, a new category of bugs.
You put in a query and get out some code. Concrete code not connected to the query anymore. The code doesn't shift out from under you later, if that's what you're implying.
Why wouldn't it be used in a "proper piece of code"? You ask Visual Studio/XCode/Eclipse for something and it gives you back a snippet that you can use as a starting point.
Why wouldn't it be used in a "proper piece of code"? You ask Visual Studio/XCode/Eclipse for something and it gives you back a snippet that you can use as a starting point.
But it would be a good alternative to bash one-liners.
oh absolutely, if they can get it general enough to do that. If they manage to include this in powershell (since it's MS), they could have the worlds greatest command line.
"Get all current running processes, select the ones that use over 5% CPU, and display them to me sorted descending by CPU usage"
More verbose than the normal, but damn it would make things easier for the non-initiated.
"Get all current running processes, select the ones that use over 5% CPU, and display them to me sorted descending by CPU usage"
More verbose than the normal, but damn it would make things easier for the non-initiated.
The danger is that you wind up with another AppleScript - something that might read more easily for the non-initiated, but is no easier to write, because it only accepts one particular finicky subset of natural language while dishonestly implying that it'll magically understand all of it.
I would hope that Microsoft is smarter and has learned from those mistakes. It does seem to be their intention to do it with full natural language processing. We'll have to wait and see.
Pretty neat. I noticed something a bit off about the "tint a random photo blue" example around 2:22. Two of the three search results are fine, but the second search result is
pics := media -> pictures
pic := pics -> random
pic -> tint (colors -> sepia)
c := colors -> blue
Isn't that the point though? That the search results show different possible interpretations of the text entry, and this is one of the possible ones?
I wouldn't say "Tint a random photo sepia, and then set an arbitrary unrelated variable to blue" is a possible interpretation of the text "Tint a random photo blue".
The animated (translation) transitions seems everything but ergonomic to me.
For instance, static touch controls grid are pretty efficient, see:
http://www.malacria.fr/pdf/gutwin14-fasttap.pdf http://www.malacria.fr/video/fasttap-video.mp4
For instance, static touch controls grid are pretty efficient, see:
http://www.malacria.fr/pdf/gutwin14-fasttap.pdf http://www.malacria.fr/video/fasttap-video.mp4
Interesting research. Of course, you could always hack the Home button to act as a Command key and hold the device in landscape orientation without having to give up a piece of the screen.
This system could be used together with speech recognition (Siri, Cortana, etc. personal agent) to automate tasks and write little apps on a mobile device.
Very interesting demo. It reminds me of Microsoft Australia's work with on(X)... but this is closer to code completion on steroids.
I have attended one of Sumit's guest lectures, and his work on Flash Fill is pretty cool.
So this is what we were all working on at ICFP last year :)
Am I the only one hearing Daft Punk's https://www.youtube.com/watch?v=XfH3erWacsQ?
Le and Gulwani's paper does cite Little, but the refutation is a bit weak.
> Little and Miller propose a code completion tool that synthesizes the most likely Java expression in a code context from a set of keywords [23]. Smart-Synth is different, in that it synthesizes a complete script and does not require extra contextual information.
From the video, a "complete script" appears to be 1 or 2 simple expressions. And not requiring contextual information sounds like it can't take advantage of contextual information.