Link preview
把临床决策从 LLM 手里拿走
让 LLM 做治疗，它会凭感觉即兴选技术；而提示词约束不住——一条足够情绪化的消息就能把模型带过它的指令。所以我们把决策移出了模型：纯 Python 逐轮裁定，LLM 只负责给状态打分和写措辞。本文讲架构、实测延迟，以及这套刚性设计的代价。 Hamo AI · hamoai.tech
An LLM asked to do therapy will improvise a technique based on vibes, and prompts don't hold — a sufficiently emotional message talks the model past its instructions. So we moved the decision out of the model. Pure Python adjudicates every turn; the LLM only scores state and writes words. Here's the architecture, the measured latencies, and what the rigidity cost us.
Comments