OpenAI Unveils Codex, a Massive AI Model for Generating Software Code
OpenAI introduces Codex, a 12 billion parameter AI system that translates natural language into functional program code. The model solves over 70% of programming challenges in a new benchmark test, significantly outperforming its predecessor.
OpenAI announces Codex, a 12 billion parameter AI model designed to generate program code directly from natural language descriptions. Built upon the foundation of the GPT-3 language model, this new system represents a major leap in automated coding capabilities. The underlying technology already powers GitHub's Copilot tool, but Codex itself serves as the specialized engine fine-tuned specifically for translating human instructions into executable software.
To train Codex, the OpenAI team collects Python code from 54 million public GitHub repositories, ultimately filtering the data down to a 159 GB dataset. The researchers adapt the GPT-3 text tokenizer to better handle the unique structure of programming languages, even adding special tokens to manage Python's critical whitespace formatting. This focused training allows Codex to understand the distinct syntax and logic patterns found in software development.
The system undergoes testing on HumanEval, a newly created open-source benchmark consisting of 164 programming challenges validated by unit tests. While the standard GPT-3 model fails to solve any of these problems, Codex passes the unit tests for 28.7% of challenges on its first attempt. When the system generates 100 different solutions per problem, it successfully produces at least one correct answer for 77.5% of the tasks.