Gen AI-Assisted Code Authoring
summary
Meta's Code Compose uses 1.3B and 6.7B models derived from FAIR's encoder, fine-tuned on first‑party Python, Hack and Flow code and supplemented with notebooks; inference runs only in empty blocks or after trigger characters and serves single‑line completions from the small model and multi‑line from the large model via a Rust LSP front end to GPU hosts. Fine‑tuning raises Python exact match from ~23% to 36% python exact match, transfer learning improves Hack/Flow, 61% offline gain comes from language‑causal masking, and suggestion acceptance exceeds 20%+ acceptance.
AI-written summary · about
chapters (5)
- 0:00Introduction to code compose
- 1:09Model training and tuning
- 5:45Architecture and integration
- 7:37Product demonstration
- 8:55Impact and future outlook
related talks
15mKeynote: Pre-training and Fine-tuning of Code Generation Models
33mAI assisted software delivery
26mAl and Software: The Good, The Bad, The Scary
24mHow AI is increasing the surface area for serendipity
25mState of Open Source AI Coding Models
45mProgrammer’s Apprentice Season 2: Future Directions in AI-assisted Coding
01h 00m10 tips to level up your ai-assisted coding
33mAccelerate Code Quality in the AI Age