Open Source Devin Achieves 12.29% on SWE Bench Test, Led by John and @jyangballin

The development and release of an open source version of Devin, utilizing GPT4, has marked a significant advancement in AI and coding, led by John and his team, including contributions from @jyangballin. Achieving a 12.29% success rate on 100% of the SWE Bench test set, it has shown impressive results compared to Devin's 13.84% on 25% of the test set. This advancement demonstrates the potential for further generalization and application in coding, especially in handling GitHub repo issues. The SWE-Agent, a key component of this development, allows users to interact with, read the source code, and understand its operational logic. It has been highlighted for its capability to interpret code with better reasoning, additional skills such as chart plotting and CSV file analysis, higher reliability, and the ability to test its output, as noted by @e2b_dev. The Agent-Computer Interface (ACI) design is deemed critical for the success of AI agents, akin to the importance of Human-Computer Interaction (HCI) in human-computer effectiveness. The SWE-Agent is readily usable on any GitHub issue, marking a step forward in AI research assistance, with @stuhlmueller explaining the significance of this development for @elicitorg on the latest @CogRev_Podcast.