LLMs Lose Performance After Launch. New Data Exposes the Pattern.

An independent researcher has published an interactive visualization revealing a hidden industry pattern: large language models systematically decline in performance weeks or months after their release. By tracking daily ELO ratings from the public LM Arena leaderboard, this tool shows that headline-grabbing benchmark announcements don't reflect real-world performance decay, likely caused by aggressive quantization or additional filters applied post-launch. For users and organizations evaluating LLMs, this underscores that launch-day metrics are unreliable predictors of sustained quality.