Hacker News Reader

how did they train it ?

Not sure what is their data source but it is not something new, wolfram alpha do this since many years and more in amazing way. It is based on linguistic processing system to perform these tasks.

I remember playing around with wolfram alpha. It was really interesting to see it search the web, render formulas and solve worded math problems. I thought it was really neat when it could calculate things like atomic weight of a compound or what something would weight on the moon.

Okay.

Make the AI trace control flow.

Then learnt language syntax or teach language syntax,

They build a model for all control flows and label it.

When user prompts for a code, it does NLP, then find the matching label and returns the control flow

I think it is a transformer much like GPT-3. Having been trained on public GitHub repos, the bulk of READMEs, Q&As plus the textbooks and occasional novels might just be enough to teach it conversations and the definitions. The Python evaluation stuff might have been picked up from textbook interaction samples and (assuming it will figure out the correspondence) jupyter notebook files.

Either way, how did they label the control flow?

A wild guess - trace the control flow

Reply on news.ycombinator.com