First AI software engineer: Devin writes code and can generate complete programs

Unlike ChatGPT or GitHub Copilot, Cognition Labs' AI software developer "Devin" should not only generate code snippets, but also be able to program and debug independently based on text input. To this end, Devin has its own Linux command line, code editor and web browser, allowing it to work independently as an AI software developer.

In a short video, Cognition CEO Scott Wu demonstrates Devin's functionality using a development task. In the demo, the "first AI software developer", as Cognition calls Devin, is asked to investigate the performance of different LLaMA language models. The AI starts with a flowchart, then reads the API documentation in the browser, and then fixes an error in the generated code. The final result is presented by Devin on a live website.

As it works, Devin reports its progress via text prompts and can respond directly to change requests. "With Devin, engineers can focus on more interesting problems and engineering teams can strive for more ambitious goals," Cognition promises in the blog post about its AI developer Devin.

In other videos, Cognition employees present software created with Devin, ranging from a test environment for a program, to an interactively developed "Game of Life", to the training of another AI by AI software developer Devin. According to Cognition, Devion should be able to solve about 14 percent of problems on its own, which is significantly higher than other AI models.

In the video presentation, Cognition CEO Scott Wu also offers interested developers the opportunity to send problems from their own software development to Devin. It is not yet possible to use Devin directly, and Devin is the first product from the start-up company Cognition.