Teaching a Machine to Code
summary
A production code-completion system is built from 700 cleaned C# repos (200k documents) using lightweight sequence/Markov models and a heavier AST-embedding deep network; a simple frequency recommender fits in a 1 MB model and yields about 58% accuracy, while the deep model reaches about 71% accuracy but is ~5x larger and ~5x slower until quantized; data shows a 14% cold-start rate for first invocations and the top 50 classes cover ~28% of call sites, and the system is deployed across C#, Python, Java and C++.
AI-written summary · about
chapters (8)
- 0:00The evolution of compilers
- 2:11Data in software development
- 5:20Analyzing code characteristics
- 7:11Representing code for AI
- 10:20Opportunities for AI in tools
- 14:49Building an intelligent IDE
- 20:09Simple vs. deep learning
- 24:20Challenges and future outlook
related talks
43mMachine Learning for Developer Productivity
37mMachine Learning on Source Code
39mTraining AI To Code Using The Largest Code Dataset (Project CodeNet)
34mHow to train a fast LLM for coding tasks- 39mTraining AI To Code Using The Largest Code Dataset (Project CodeNet)
01h 05mSelf-Writing Code
46mCode + AI: Will Robots Take Our Coding Jobs? ML Applied to Programming
11mGen AI-Assisted Code Authoring