WHY THIS MATTERS IN BRIEF
In time we’ll see a fully autonomous AI software developer, at first it will be okay, then it will be great.
Love the Exponential Future? Join our XPotential Community, future proof yourself with courses from XPotential University, read about exponential tech and trends, connect, watch a keynote, or browse my blog.
A new Artificial Intelligence (AI) model has triggered unease among developers because of its astounding ability to write complex code, then scan any errors that may arise in compilation and automatically correct them – just like a human programmer would. The model, dubbed ‘Devin’ is developed by AI startup Cognition.
Backed by bigwigs like Peter Thiel’s Founders Fund, former Twitter exec Elad Gil, and Doordash co-founder Tony Xu, Cognition has secured $21 million in funding. And while AI coding assistants like those on Stackoverflow and Github have been around for a while, including OpenAI’s celebrated Copilot, Devin purportedly raises the bar by taking on end-to-end development responsibilities.
If Cognition’s claims hold water, Devin could mark a shift in the world of AI-assisted coding. Rather than playing second fiddle to human developers, this AI seems primed to operate as a self-sufficient, dare we say ‘edging on autonomous,’ software engineer in its own right. According to the startup’s founder and CEO Scott Wu, Devin operates within a secure sandbox, planning and executing complex engineering tasks through common dev tools like code editors and web browsers.
All a human needs to do is feed Devin instructions via a chat interface. From there, the AI dynamically maps out a solution, gets its hands dirty writing the actual code, fixes bugs along the way, tests its work, and keeps the user updated in real-time. If the programmer spots any issues, they can simply message Devin to course-correct.
See DEVIN in action
Wu demonstrated Devin’s impressive range in a blog post, from deploying web apps and websites to fine-tuning large language models using GitHub repos.
Perhaps its biggest feat, however, is its performance on the SWE-bench test which evaluates AI’s ability to resolve real open-source software issues from GitHub. Devin could solve 13.86% of these cases entirely independently compared to figures like 4.8% for Claude, 3.97% for a different AI called SWE-Llama, and 1.74% for GPT-4.
While Devin remains under wraps for now, Cognition hopes to make it available to select customers soon. The company seems to view coding as just the start too, suggesting it could leverage its core “long-term reasoning and planning” advances to create AI workers for other domains.