There’s no shortage now of LLMs that help engineers write and debug code, but a new startup seems to have taken things up a notch.
US-based Cognition has launched what it calls the world’s first AI software engineer. Named Devin, the AI can write, debug and deploy code to create functioning websites and products. Its creators say that its coding performance far surpasses existing state-of-the-art LLMs like GPT-4 and Gemini.
The startup had been working in stealth, and was formally incorporated just two months ago. Its founders are Scott Wu, Steven Hao, and Walden Yan, who’re all coders themselves. Cognition’s team of 10 specializes in coding — between them, they’ve won 10 gold medals are international coding competitions. They now seem to have used this background to help create an AI that codes.
And Devin, which Cognition says is the world’s first AI software engineer, does really seem to live up to its billing. Devin requires just a prompt to get started — users can ask it to build a website, or a software program. Devin then comes up with a plan to systematically implement the software, and tackles each part one by one. To help it operate, Devin has its own command line, its own code editor, and even its own browser. After receiving the prompt, Devin gets to work, creating files and writing code. It debugs its own code as well — in the demo, Devin is shown to automatically add a print statement when it runs into an error, and then proceeds to fix the problem. Early users have reported being able to build whole websites and simple games in 5-10 minutes.
This appears to be a step up from what AI was capable of so far. LLMs like ChatGPT and Google Gemini are able to write code, but users will need to deploy and implement it to get a functioning website or product. Devin, on the other hand, is able to create finished products from a simple prompt while the user simply watches. Also, Devin’s coding skills seem to be superior to other LLM-based approaches — it was able to solve 13.86% of open issues on Github, compared to 4.8% for Anthropic’s Claude and 1.8% for GPT-4.
Unsurprisingly, Devin has found itself some serious backers. PayPal co-founder and noted contrarian Peter Thiel, along with a bunch of other investors, have already put in $22 million into the company. It remains to be how how Devin fares, but the idea of an AI software engineer that can plan, execute and deploy code is something that can completely upend the field software engineering as we know it.