Experimental: Syntax AI with React.js | Hacker News Reader

Experimental: Syntax AI with React.js

blog.codebox.sh

72 pointsjonsharratt14 comments

> We scrape the data from the public GitHub source code to generate a model that stores prop type names and the type in which they have been set. Once we have a huge set of sample data we then feed this into AWS Machine Learning as a multi-class classification model.

I would literally pay for a tutorial on how that is done (not the scraping github part... that part seems straight forward)... the ML part.

Anyone know of a tutorial on how to build classification models like that?

Many classification models are available as open source.

I'd start by downloading Weka [0] and play around with the different algos they have and feed in some training data, csv or similar.

When you find a model with a good result (for example, look at the confusion matrix) you can try to find an open source implementation of that model in your favorite language and take it from there.

[0] http://www.cs.waikato.ac.nz/ml/weka/

Are there any efforts to feed an AST into a neural net and see what comes out of it? Like a coding assistant that "understands" what your code does or something?

I work in this space. We have some pretty sophisticated pilots in process right now for such an assistant. We can talk privately if you want to learn more. Expect to see a lot of this coming out over the next few years.

That's super interesting, would you mind emailing me at [redacted]?

Emailed! Thanks

used to work in related area, would love to talk: cammarata.nick@gmail.com

Reached out. Thanks for your interest

Can you talk a bit on how you convert the AST to your NN input?

Do you traverse the AST and send it to a recurrent net? Do you compress the AST into a different representation first?

Well there are always trade offs. NNs aren't very good with trees, too symbolic so usually compressing it into a different format is better. We have a few we use depending on the goal, but always looking for trying new ones since this really is the crux of the problem.

That would be superb to try, for now I can't wait for demos of what is possible with neural but it is way out of my depth.

@addcn sounds mega interesting what you are up to :D

Well I hope it can be made easier to digest. We've all but confirmed at this point that the knowledge needed to build an effective programming AI is not available in any corpus of source code. There's a lot of information brought out in runtime and understood by humans that's critical to the learning. Without strong collaboration between both humans and bots nothing practical will come of any of these efforts.

The ML bit seems to be very interesting, the more impressive part out of this.

Voice to code generation not so much these days. Esp with code such as React (easy boilerplate, standard component structure, uniflow of data from parents to children)

Reminds me of https://kite.com/

Reply on news.ycombinator.com