OpenAI Codex is a program to train AI agents in the world’s most popular games, starting with Dota 2 and StarCraft II. We’ve seen significant progress in research on general intelligence by training agents in these environments, but training on them has been prohibitively expensive for most AI researchers. Our goal is to help democratize access to training on these environments.
Codex will be open to anyone: you can apply for access at our website. It’s built on top of OpenAI Universe, which lets you use any game or application running on your computer as an environment for training AI agents. We’re releasing a large batch of new Universe environments today, including some using commercial games such as Dota 2 and Civilization V that we’ve bought licenses for on behalf of all users.
Rationale:
We’ve been working on a project to train an AI system to solve games, simulations, and other environments using self-play. We’ve now built a system that can solve many environments from the Arcade Learning Environment (ALE), including games like Montezuma’s Revenge and Pitfall, achieving or surpassing human gameplay performance. We believe that self-play is likely to be an important component in the future of AI.
Today we’re releasing OpenAI Codex, which allows anyone to train an AI agent on a wide variety of environments, including the games we’ve trained on. The OpenAI Codex platform will also allow us to scale up our training efforts in ways that are difficult to do on our own. We hope both researchers and enthusiasts will help us identify which problems are most fruitful to work on next.
The OpenAI Codex engine runs as a client-side web application in your web browser. It includes a wide variety of environments based on ALE (see here for more details). It also includes a mode where you can play against other players’ agents by connecting to our servers through WebSockets.
We created this project because we want to encourage more people to explore self-play, and because we want help accelerating our research through distributed training
Taking the first step towards building a training platform for AGI.
Creating a curriculum for AGI safety research, and then opening it up to the world.
Building an open-source toolkit for playing games with deep neural networks.
OpenAI’s mission is to build safe artificial general intelligence (AGI), and ensure AGI’s benefits are as widely and evenly distributed as possible. We believe that a crucial part of this mission is to create better tools for researchers in the field, and we’ve decided to take our first big step toward that goal by open-sourcing OpenAI Codex.
We believe that large-scale codebases will be an important component of artificial general intelligence (AGI), and we want to develop better tools to enable researchers to create them. Since the beginning, we’ve been working on an AGI training platform called OpenAI Codex. Today, we’re open-sourcing Codex’s core components:
We’re releasing OpenAI Codex, our program for training AI on games. It includes a neural network that can learn to play Atari games from raw pixels, and models of imaginary games that we’ve developed to train and test general learning algorithms.
This year we released Universe, a platform for measuring and training an AI’s general intelligence across the world’s supply of games, websites and other applications. We developed Universe using our own hardware infrastructure; now we’re making it available for anyone to use.
The OpenAI Charter describes the principles that guide us as we execute on our mission.
Trained on about 50 million human-contrived actions in these games, the neural network acquired a proficiency greater than most humans at all 57 games we tested from Atari 2600, attaining an average score of 75% of the best possible human score.
The OpenAI Codex is a free program for people who want to work on creative AI projects. It’s open to anyone: a writer, an artist, or a machine learning researcher. The Codex will help you get started with creative AI, and it has all the tools and resources you need.
This announcement includes a lot of material from the Codex, but the best way to go through it is to read it on the site. So go ahead and get started at openai.com/codex!
Codex gives you access to an interactive training environment where you can create your own machine learning models using cutting-edge technology like reinforcement learning and TensorFlow. You can explore these tools by reading tutorials or by trying out examples as you go through them. You’ll also have access to the code that powers these tools, so you can build your own applications.
Codex gives you access to an interactive training environment where you can create your own machine learning models using cutting-edge technology like reinforcement learning and TensorFlow. You can explore these tools by reading tutorials or by trying out examples as you go through them. You’ll also have access to the code that powers these tools, so you can build your own applications.
Codex gives you access to an
Today we’re announcing OpenAI Codex, a program that recognizes and rewards anyone who creates content in the AI community.
Codex is a way for AI researchers and enthusiasts to get recognition for their work. It’s also a way for people to learn more about AI, and a means of discovering new people and ideas.
To participate, you can upload your own content (like blogs, code, or tutorials) to codex.openai.com and earn Codex Points by sharing your content with others. If you’re interested in machine learning but don’t know where to start, you can use Codex to find people who are working on exciting things related to your interests. Once you find someone interesting, you can follow them on Codex and see what they’re posting.
If you have any questions or comments, please reach out at @OpenAICodex on Twitter or email us at codex@openai.com.
Today we’re announcing a new platform for advanced AI research. Our goal is to give researchers around the world better tools, faster hardware, and more compute resources to make breakthroughs in artificial intelligence.
Codex is now open to everyone: it allows you to run your own experiments, train your own agents, and even host your own games. We believe this will accelerate the development of AI systems that act in the real world.
Codex runs in the cloud on top of hundreds of machines, each with four GPUs, making it possible to train deep reinforcement learning (RL) models far faster than previously possible. For example, using Codex you can train a model that plays Montezuma’s Revenge (a 1980s Atari game) as well as a professional human gamer using only two hours of gameplay experience, as compared to the months or years of experience required by previous algorithms. Note that these results are not yet published.
Codex allows running large-scale experiments, and we plan to release another blog post soon with more details about our computational infrastructure and its capabilities.