It's Alive! Machine Learning Writes Your Code!
summary
A pipeline extracts functions and comments from GitHub via BigQuery and the TypeScript AST, cleans and tokenizes text, trains an encoder‑decoder model in TensorFlow (training runs take hours on CPU but seconds on Google TPU), converts the model to TensorFlow.js, publishes it to npm and wires it into a VSCode extension that inserts predicted comments. The dataset reaches 300,000 comments (demos on 5 and ~2,000 examples show useful but overfitted, sometimes nonsensical results).
AI-written summary · about
chapters (12)
- 0:00Start
- 1:00WHERE IT ALL STARTED
- 3:13INITIAL EXPERIMENT
- 6:00LEARNINGS #1
- 7:59GOAL
- 8:31OUR PROCESS
- 10:07GATHER DATA
- 13:46CLEANING
- 14:57PREPARING
- 16:00MODEL ARCHITECTURE
- 17:56TRAINING
- 32:55LEARNINGS #2
related talks
36mIt's Alive! Machine Learning Writes Your Code! 🧠🔮
34mIt's Alive! Machine Learning Writes Your Code! 🧠
22mIncreasing developer productivity with Generative AI
22mLetting AI Steal My Job - I'm Not Even Mad, This Is Amazing
07mInterview
37mEnhancing VS Code extensions with GitHub Copilot
40mLevel Up with Copilot
39mLive Coding A Machine Learning Model from Scratch