He first AI-based software engineer is a reality. Cognition Labs, a Silicon Valley startup, introduced Devin, an artificial intelligence capable of problem-solving, planning, and performing complex engineering tasks. The new developer is completely autonomous and can join work teams with other programmers.
In accordance with Cognition, Devin can perform tasks that require thousands of decisions.s, is able to remember the context of each step and learns to correct mistakes. Unlike Copilot It is not a wizard that offers code suggestions, but you can develop a project completely. Its creators have equipped it with typical tools such as a code editor, shell, and a browser within an isolated environment.
Cognition Labs has released a series of videos showing Devin's skills in action. In general terms, this AI can execute actions through simple text commands (prompts). Can ask you to develop a game, an app, or learn how to use an unknown technology. Devin also joins work teams and contributes to production repositories or helps find code bugs.
Perhaps the most terrifying thing is that Devin can train and refine other artificial intelligence modelsl. One example asks you to configure a fine tune for the Llama 7B model from a Github repository and a series of text instructions. When sending the prompt, the engineer virtual clone the repository, analyze the code and run the training. If an error is found, Devin studies the environment and finds a solution.
Devin's AI Doesn't Replace Human Software Engineers (For Now)
Something important to mention is that Devin is not a latent threat to the work of other programmers, at least not at this moment. Cognition evaluated its artificial intelligence through SWE-bench, a benchmark to solve real problems found in GitHub repositories. According to the first data, Devin correctly solved 13.86% of problems from start to finish.
Although the percentage is low, Devin outperforms language models such as Claude 2, GPT-4, SWE-Llama 13 GB and 7B and ChatGPT 3.5.
A curious detail about this artificial intelligence is that it can perform jobs on Upwork. Devin writes and debugs code, taking into account project instructions, so some novice programmers could benefit from additional income.
Devin is in development, so it is not available to all users. Cognition is accepting applications to test it in real environments, although being aware that it is not capable of replacing your staff of programmers.
In accordance with Bloomberg, Cognition is made up of engineers who have won gold medals in programming competitions. Unlike OpenAI and other giant companies, The work team is small and everyone operates remotely. His work with Devin has guaranteed them an investment of $21 million and the support of some industry leaders.