但他也明确指出，真正的创新靠的是强化学习，不是蒸馏。从 DeepSeek、月暗、MiniMax 公开的论文来看，它们都用有相当完善的基础设施和优秀的人才，远非只靠小聪明小伎俩企图弯道超车的「小作坊」。
loop.call_soon_threadsafe(
。WPS官方版本下载对此有专业解读
Queries duration: PT9.64S | PT9.135S
Sounds promising, right? I spent some time testing it, but results were poor—high false positives and false negatives, no clear threshold for reliable classification. Plus, it suffers from high inference costs, poor cross-model generalization, difficulty deploying large models locally, and lack of access to closed-source model weights. All in all, it’s neither elegant nor reliable.