👀 Looks like Alibaba's AI coding agents had their 'oops' moment in a real-world stress test. Over 233 days of continuous code evolution, it turns out maintaining code is the Achilles' heel of AI models. Enter SWE-CI, focusing on long-term code stability over flashy one-time fixes. With 75% of AI models faltering during maintenance, only Claude Opus 4 managed to keep its head above a 50% zero-regression rate. While HumanEval and SWE-bench ask 'does it work right now?', SWE-CI probes 'will it still work after 6 months of changes?'. Alibaba's EvoScore adds a twist by penalising code that falls apart after initial success. The home truth? AI can churn out code really well, but a bigger challenge is making the code better over time. In AI's defence, maintaining code is like herding cats - if the cats were continuously changing.

#TechLife #BuildingWithAI #AI