Open AI unveiled the main points of the AI ​​coding agent ‘CODEX’. This version is the extent of coping with coding operations instructed in parallel, but he emphasized that it is going to upgrade to AI colleagues in the long run.
sixteenth (local time) artificial intelligence (AI) coding agent Details of’Codex ‘It has been released.
Codex is a next -generation coding agent that may handle various development tasks at the identical time, resembling code writing, bug modification, refactoring, testing, and answering code related questions. The work runs individually in an independent cloud sandbox environment with a pre -contained user’s code repository.
It really works based on ‘Codex-1’, which is optimized for ‘O3’ for software engineering.
Open AI said that the Codex-1 training goal is to “adjust the outcomes based on human coding preference and standards.” In comparison with O3, he explained that an individual creates a cleaner patch that may be integrated into a regular workflow immediately.
The model has been trained to reflect a full request, and has the power to accurately execute the code by repeating the code until the test passes. Full Request is a proposal to merge recent code changes into the unique source code.
The user can access the codex through the chat GPT sidebar, and you may ask for a piece by pressing the ‘Code’ button or asking the code base with the ‘ASK’ function. You may read and modify files within the code repository, and you may as well run test harnesses, leaners, and sort checkers. Working time is around 1 to half-hour, and users can monitor the progress in real time.
Codex runs the code change within the sandbox environment after completion of the duty, and provides the method transparently through quoted materials resembling terminal logs and test printouts. After reviewing the outcomes, you may create a GitHub Full Request or integrate it directly into the local environment.
The codex can work more efficiently with the agents.md file within the user’s code storage. This can be a text file just like readme.md that guides the codex project structure, test command, and code standards. Like human developers, performance is improved when there may be a development environment and a transparent document.
It recorded the best performance (SOTA) in SWE-Bench, a software engineering evaluation indicator, and proved its ability to perform complex development work at the identical level.
Safety and security are also key design aspects. The codex has been trained to obviously reject the request for a malicious code, and it’s supported by supporting the opportunity of abuse while supporting legitimate use cases by strengthening safety evaluation and policy.
It really works within the isolated cloud environment with the Web blocked, and it’s not possible to access external web sites or APIs. This ensures thorough control and verification of user codebase.
Actually, the Codex CLI (CODEX CLI), a version of the Codex terminal, was also unveiled. The open source tools may be utilized in the local environment of ‘O3’ and ‘O4-Mini’ within the local environment, and are specialized in repeated work automation, code Q & A, and fast modification.
The lightweight model ‘Codex-Mini-Latest’ is ready because the default model of Codex CLI, and has enhanced the power to reply. It is usually available within the API and is provided at a price of $ 1.50 per token input and $ 6 per million tokens. As much as 75% discount on prompt caching can be applied.
The Codex CLI is robotically arrange with a Chat GPT account, and the API secret is robotically set, and a professional and plus user also offers $ 50 and $ 5 free API credits, respectively.
Open AI said, “Codex is only the start,” he said. “It just isn’t just an automation tool, but a goal to become an AI collaboration with the event team.”
On this version, you may handle multiple software engineering work at the identical time, and give attention to designing it for users to run other tasks during running. “Asynchronous multiple agent workflow can be a possible way for engineers to generate top quality code,” he said.
“You may see the 2 interaction modes, ultimately, real -time pairing and work delegation.” It’s explained that developers will find a way to offer guidelines throughout the work, collaborate on strategies, and get a preliminary update on the progress.
Meanwhile, Lex Asabor Open AI developer posted on X (Twitter), “Codex is at 3:15 am and is great enough to make a revised draft at the corporate’s codebase.”
That is the announcement that the unauthorized correction of the code was held at 3:15. In other words, Codex can be useful for the unauthorized modification of the code, and is a sniper of CEO and XAI, CEO of Illon Musk.
By Park Chan, reporter cpark@aitimes.com
