Delivering AI on Code: Live Demo of source{d}
summary
source{d} builds an open pipeline that crawls public git repos, deduplicates forks into root repositories, stores them in a custom Siva archive format, and trains models on a 3 TB public Git dataset. Demos use char RNN + embeddings on Colab GPUs to generate Go/Python-like code (often starting with Apache v2) and a trained identifier-splitting model fixes camelCase/snake_case reliably. Vocabulary sizes hit ~500k tokens; structural AST embeddings are next to enable repo similarity and automated code-review signals.
AI-written summary · about
chapters (12)
- 0:00Start
- 0:13Machine Learning on Source Code #MLonCode
- 2:33Deep Learning Revolutions
- 6:59What is Source Code
- 12:04Learning from Source Code
- 12:15Learning sequences
- 13:33Recurrent Neural Networks
- 14:52Training the neural network
- 22:54Word embeddings
- 25:41Universal Abstract Syntax Trees
- 25:53Structural Embeddings
- 27:11Awesome Machine Learning on Source Code
related talks
37mMachine Learning on Source Code
16mMachine Learning on Source Code
01h 03mIt's All About AI
25mMachine Learning for source code analysis
01h 13mBuilding Production-Ready AI Systems in 90 Minutes
07mInterview
39mTraining AI To Code Using The Largest Code Dataset (Project CodeNet)- 39mTraining AI To Code Using The Largest Code Dataset (Project CodeNet)