Teaching GPT to write code
Research PaperIntroduced Codex, a GPT model fine-tuned on code from GitHub that could generate functional programs from natural language descriptions, powering GitHub Copilot and launching the AI-assisted coding revolution.
A GPT-3-family model fine-tuned on 159GB of Python code from GitHub. It could solve 28.8% of HumanEval problems (a benchmark of programming challenges) on first attempt, rising to 72.31% with 100 samples.
The paper introduced HumanEval — 164 hand-crafted programming problems that test code generation from docstrings — which became the standard benchmark for code generation.
Codex became the backbone of GitHub Copilot, launched in partnership with Microsoft, bringing AI-assisted coding to millions of developers.