Just now, OpenAI released the most powerful programming agent Codex! The new feature is actually the originator of Microsoft Copilotb?
OpenAI Codex - the initiator of a new era of AI programming, subverting traditional coding methods.
Core content:
1. The release of OpenAI Codex and its ability to process programming tasks in parallel
2. Codex's performance in actual programming cases and intelligent identification of bugs
3. The historical origins of Codex and Microsoft Copilot and their technological evolution
OpenAI dropped a bombshell at the end of this week! A research preview of the new AI programming agent Codex appeared.
This is OpenAI's most powerful AI programming agent to date. It is no exaggeration to say that this is hiring the most programming AI to help us write code.
One of the biggest highlights of Codex is that it can handle multiple programming tasks in parallel.
OpenAI once said in a media briefing, "Codex is a cloud-based AI software engineering agent that can process multiple tasks in parallel and run securely and independently on a computer that belongs to it."
How powerful is Codex? Let’s take a look at a real example in today’s live broadcast:
The guy asked Codex to help find bugs in the document, and let three Codex work simultaneously to identify and correct grammatical errors, spelling errors, consistency of variable default parameters, and other issues.
After troubleshooting, the agent discovered that there seemed to be a slight inconsistency in setting the timeout. One time we set it to 120, and another time it was set to 60. The Codex assistant suggested a modification for him. The guy was shocked and decisively chose to adjust the parameter to 120.
President Greg Brockman also participated in the live broadcast. Although Ultraman did not show up this time, he tweeted his support immediately.
He then added in his retweet:
It’s amazing and exciting that one person was able to create so much software with tools like this. “You just do things” is one of my favorite memes; I didn’t expect it to be applied to AI itself and its users in such a significant way so soon.
However, the development of this Codex agent may have been earlier than expected - it is actually the predecessor of Microsoft Copliot!!!
Codex was the brains behind early GitHub Copilot
By the way, you may be familiar with the name Codex - it is the "brain" behind the early GitHub Copilot.
The original Codex was born in 2021. At that time, it was just a model that could translate natural language into code, and it was connected to OpenAI's then-primary API. Microsoft's GitHub Copilot was started with it, featuring an "auto-completion" programming assistant, embedded in IDEs such as VS Code, and became popular.
However, although that version of Codex set off a wave in the industry, it also had many shortcomings: the code was often written incorrectly, it was prone to security risks, and it could also lead to strange associations or misjudgments due to deviations in the training corpus.
Despite this, its influence is enough to turn "AI writing code" into a real product track. Today, "Codex" is no longer the name of an old model, but has been upgraded to represent a complete intelligent product line.
It is worth mentioning that GitHub Copilot has switched to the GPT-4 model since March 2023 and upgraded to Copilot X, which has deeper integration capabilities and contextual understanding capabilities.
Codex has only seen the light of day today, and has walked its own path, evolving into a more autonomous, agent-like AI programming assistant.
Codex has powerful programming capabilities, but is more concise and practical.
Let’s talk about the Codex itself.
Codex is powered by codex-1, a version of OpenAI's o3 AI reasoning model that is optimized for software engineering tasks. OpenAI says that the code generated by codex-1 is "cleaner" than o3, executes instructions more accurately, and runs tests until the code passes.
The Codex agent runs in a sandbox virtual computer in the cloud. By connecting to GitHub, the Codex environment can preload your code repository. OpenAI said that this AI programming agent can complete tasks such as developing simple functions, fixing bugs, answering questions about the code base, and running tests in 1 minute to 30 minutes.
OpenAI said that Codex can handle multiple software engineering tasks at the same time and will not restrict users from using local computers or browsers during operation.
Codex is rolling out to ChatGPT Pro, Enterprise, and Team users starting today. OpenAI says users will initially have “ample access,” but in the coming weeks the company will set rate limits on the tool. An OpenAI spokesperson told TechCrunch that users will then have the option to purchase additional credits to continue using Codex.
OpenAI plans to open access to Codex to ChatGPT Plus and Edu users soon.
AI tools for software engineers (also known as vibe coders) have been on a roll in recent months! The CEOs of both Google and Microsoft said that about 30% of their company’s code is now generated by AI. In February, Anthropic launched its own smart coding agent tool, Claude Code, and in April, Google updated its AI coding assistant Gemini Code Assist with more “smart agent-like” capabilities.
Last words
OpenAI has just acquired browser startup Windsurf, and is continuing to invest in Codex agents in the programming field. It is not difficult to see that they are paving the way for an "AI full-stack development environment."
Codex can currently simulate development environments, run tests autonomously, and iterate code in the cloud. Will it directly connect the front-end and back-end development links in the next step? Perhaps we will soon see the moment when Codex goes online in Windsurf.
