Why can I not reproduce the exact output my model gave in production?
机会
Even at temperature zero, the same prompt sent to the same model twice can return different tokens. A January 2026 arXiv paper traced the root cause to dynamic batching: GPU reduction kernels change their internal arithmetic tree when batch size changes, and floating-point addition is not associative, so two requests processed under different batch sizes take different numerical paths and diverge. Serving infrastructure has no obligation to record or expose the batch context that produced a given output. This makes it impossible to replay a failed inference, write a regression test that pins exact behavior, or reconstruct what a model actually said during a production incident. The only solution demonstrated so far imposes a 61 percent throughput cost.
为什么重要
Deterministic replay is the minimum bar for treating a model call as an auditable computation rather than a black box you trust but cannot inspect.
我如何评估机会
机会评分是我的个人判断，而非量化指标：痛苦程度、发生频率，以及当前解决方案的匮乏程度。分数越高，意味着我认为越值得去构建。
出现时造成的痛苦程度。
人们实际遇到它的频率。
当前针对它的优质工具有多匮乏。