The remarkable capacity of a new AI model to create complex code, then identify and automatically fix any compilation errors, much like a human programmer, has raised concerns on tech Twitter. The AI company Cognition creates the “Devin” model.
Cognition has raised $21 million in funding and is supported by prominent figures such as Tony Xu, a co-founder of Doordash, Elad Gil, a former executive at Twitter, and Peter Thiel’s Founders Fund. Even while there have long been AI coding assistants, such as OpenAI’s highly regarded Copilot, Devin is said to raise the standard by handling end-to-end programming duties. Devin might signal a shift in the realm of AI-assisted coding if Cognition’s assertions prove to be true.
This AI seems ready to take the lead as a self-sufficient software engineer rather than taking the back seat to human workers. Scott Wu, the CEO and creator of the company, claims that Devin works in a secure environment, organizing and completing challenging technical tasks using standard development tools like code editors and web browsers.
Devin only needs to be instructed by a person via a chat interface. After that, the AI develops a solution on the fly, writes the actual code, fixes mistakes as it goes, tests it, and provides real-time updates to the user. The developer may easily notify Devin to make any necessary corrections if they see any errors.
In a blog post, Wu demonstrated Devin’s enormous breadth, demonstrating his ability to construct websites and online applications as well as fine-tune huge language models using GitHub repos. Its performance on the SWE-bench test, which evaluates AI’s ability to fix actual open-source software issues from GitHub, may be its greatest achievement, however. Devin was able to completely autonomously solve 13.86% of these cases, compared to 4.8% for Claude, 3.97% for SWE-Llama, a competing AI, and 1.74% for GPT-4.
Although Devin is currently under wraps, Cognition plans to soon make it accessible to a limited number of customers. The business seems to view coding as only the beginning, suggesting that it may be able to use its foundational advances in “long-term reasoning and planning” to create AI workers in other domains.
