↗ https://reddit.com/r/wallstreetbets/comments/1rofy6p/ai_coding_agents_failed_spectacularly_on_new/
Alibaba tested AI coding agents on 100 real codebases, spanning 233 days each. The agents failed spectacularly.
Turns out passing tests once is easy. Maintaining code for 8 months without breaking everything is where AI collapses.
SWE-CI is the first benchmark that measures long-term code maintenance instead of one-shot bug fixes. Each task tracks 71 consecutive commits of real evolution.
Extremely bearish for AI coding use cases.
https://x.com/alex\_prompter/status/2030331477918126286