OpenAI Codex is the AI model and set of APIs behind GitHub Copilot, the version control tool’s much-awaited AI-powered pair programmer (currently in Technical Preview) right inside your code editor.
Currently, OpenAI is working with a number of companies and development partners to build out the model and APIs so that they can eventually be made available to the public (for a fee).
Related: How Much Does OpenAI Cost?
How Was OpenAI Codex Trained?
They say an AI model is only as good as the dataset it was trained with. If that’s true, then set your expectations high for OpenAI Codex!
OpenAI Codex is based on GPT-3, OpenAI’s third-generation autoregressive language model trained on 570 GB of text from the Internet (pretty much the entire publicly available World Wide Web).
In addition to the GPT-3 base, OpenAI Codex was trained on 54 million public software repositories on GitHub, Venture Beat reports, resulting in a total of 159 GB of programming training data.
How Does OpenAI Codex Work?
Think of OpenAI Codex as a pair programmer who you collaborate with via an API. It parses your natural language input and outputs functional source code.
You can ask it to solve basic programming tasks for you, or you can use it as an AI-powered programming assistant that will auto-complete your code. As long as it understands the task, there’s a good chance it can output a working solution.
If you’re the curious type of person who doesn’t want to wait to get access to the private beta—and who does!—you can watch a live demo that the OpenAI Codex team recently conducted right now:
Of course, this model has quirks.
It won’t always understand what you’re asking it to do. And it will probably take a few tries to get to the solution you’re looking for.
Will OpenAI Codex Replace Developers?
If you’re worried about losing your job because you’ve been replaced by an AI programmer, don’t be. These tools are far from being able to replace humans in programming and are, at least in the foreseeable future, assistants rather than competitors.
As Thomas Smith, who got to test the OpenAI Codex private beta, eloquently puts it for IEEE Spectrum, “The day when a non-coder can sit down with Codex, write up a spec sheet, and crank out a working piece of software is still far away.”
With that being said, as AI models like OpenAI Codex get better and better at churning out code for basic programming tasks, it’s not hard to see how entry-level programming won’t have the employer demand or career appeal it has today.
It is not inconceivable that today’s DevOps Engineers, for example, could become tomorrow’s AI Model Operators. Software will become ever more complex, and the line between human- and AI-developed systems will get increasingly blurry.
How Can I Get Access to OpenAI Codex?
OpenAI Codex is currently in private beta. If you would like to gain access to OpenAI Codex, you must complete a questionnaire and be placed on the waiting list.
Apart from your name and email address, the questionnaire asks you how you plan to use OpenAI Codex, if there are any specific ideas you’d like to implement with it, and whether you’d like to provide a link to your Github profile.
Although not all questions are mandatory, we would like you to fill them out and answer them as accurately as possible.