[OpenAI Codex](https://openai.com/blog/openai-codex/) ```cardlink url: https://openai.com/blog/openai-codex/ title: "OpenAI Codex" description: "We’ve created an improved version of OpenAI Codex, our AI system that translatesnatural language to code, and we are releasing it through our API in privatebeta starting today. Codex is the model that powers GitHub Copilot[https://copilot.github.com/], which we built and launched in partnership" host: openai.com favicon: https://openai.com/assets/images/favicon.svg?v=2887575c27 image: https://openai.com/content/images/2021/09/codex-og-image-color.png ``` > 私たちは、自然言語をコードに変換するAIシステムである[[OpenAI]] Codexの改良版を作成し、本日よりプライベートベータとしてAPIを通じてリリースします。Codexは、1ヶ月前にGitHubと共同で構築し、ローンチしたGitHub [[GitHub Copilot]]の動力源となるモデルです。Codexは十数種類のプログラミング言語に対応しており、簡単なコマンドを自然言語で解釈し、ユーザーに代わって実行することができるため、既存のアプリケーションに自然言語インターフェースを構築することが可能になりました。現在、私たちは、APIを通じて、OpenAI Codexの上に構築する企業や開発者を募集しています。 論文 [[2107.03374] Evaluating Large Language Models Trained on Code](https://arxiv.org/abs/2107.03374) ```cardlink url: https://arxiv.org/abs/2107.03374 title: "Evaluating Large Language Models Trained on Code" description: "We introduce Codex, a GPT language model fine-tuned on publicly availablecode from GitHub, and study its Python code-writing capabilities. A distinctproduction version of Codex powers GitHub Copilot. On HumanEval, a newevaluation set we release to measure functional correctness for synthesizingprograms from docstrings, our model solves 28.8% of the problems, while GPT-3solves 0% and GPT-J solves 11.4%. Furthermore, we find that repeated samplingfrom the model is a surprisingly effective strategy for producing workingsolutions to difficult prompts. Using this method, we solve 70.2% of ourproblems with 100 samples per problem. Careful investigation of our modelreveals its limitations, including difficulty with docstrings describing longchains of operations and with binding operations to variables. Finally, wediscuss the potential broader impacts of deploying powerful code generationtechnologies, covering safety, security, and economics." host: arxiv.org ```