Thread by @lizthegrey

Next up is @francesc [ed: <3 <3 <3 my former colleagues!] on why machine learning will make programming better! #VelocityConf

@francesc Is software eating the world? The number of LOC is increasing dramatically -- tens of millions of LOC for Chrome, Windows, etc.

And *cars* have hundreds of millions of LOC. #VelocityConf

He works at @sourcedtech but also runs @justforfunc.

So how can ML on source code help us? It's related to doing natural language processing and data mining. #VelocityConf

Starts with all the data on Github, but you still need more. We need to retrieve all that data from archives, then we need to classify and parse, in order to extract metadata.

Goal: answer questions like "what are the names of the functions I altered last month?" #VelocityConf

We need levels of abstraction to deal with source code than individual characters -- tokens, ASTs, control flow graphs, etc. [ed: and this is taking me back to compilers class...] #VelocityConf

[ed: being mildly sarcastic and ribbing @francesc because I know he'll take it well: and we apply some ML neural network pixie dust to it, which there isn't time to explain...]

and we can do things like predict the next tokens, or handle sequences with RNNs #VelocityConf

So can we generate code with RNNs? [ed: okay, now I'm seeing where this is going]

Sure, we can generate
if err != nil {
return nil
}
#VelocityConf

but we can also detect errors in source code (e.g. use of the wrong variable/missing variable).

Lots of research to be done. [ed: wish there were more concrete results/benefits to show] #VelocityConf

Look for the unexpected/"interesting" and call attention to it for code reviews. including things you wouldn't catch with static analysis.

Goal: assist code review. @francesc remembers getting great code reviews from the Go team at Google, assisted by tooling! #VelocityConf

Bug prediction, automated code review, automated style guide enforcement, education, and eventually code generation and automatic documentation. [fin] #velocityconf