问 HN：我们是否在强迫 LLM 成为状态机？
1 分•作者: kodiyak•6 天前
我正在构建一个客户服务平台，现在遇到了令人沮丧的瓶颈。那些“AI 智能客服”的演示和教程总是光鲜亮丽，但将用户混乱、非结构化的意图与僵化、事务性的内部流程连接起来的现实，却变成了一场充满边缘情况的噩梦。
感觉我 80% 的工程精力都花在构建防护措施上，以防止幻觉或灾难性的逻辑错误，而真正用于发布新功能的时间只有 20%。
对于那些真正参与其中的人（仅限生产级），我的问题是：
你们是否找到了一个在严格的业务确定性和 LLM 的概率性之间，真正可行的架构“最佳点”？
还是我们只是试图将一个随机的 token 预测器硬塞进一个有限状态机里，而从根本上说，这只是对关键任务工作流程不可持续的炒作？
我需要的是实战经验和现实反馈，而不是理论性的推销。
查看原文
I'm building a customer service platform and I've hit a wall of frustration. The "AI Agent" demos and tutorials are always sleek, but the reality of bridging messy, unstructured user intent with rigid, transactional internal processes has been a nightmare of edge cases.<p>It feels like I spend 80% of my engineering effort building guardrails to prevent hallucinations or catastrophic logic failures, and only 20% actually shipping features.<p>My question for those with actual skin in the game (production-grade only, please):<p>Have you found a legitimate architectural "sweet spot" between strict business determinism and the probabilistic nature of LLMs?<p>Or are we just trying to shoehorn a stochastic token predictor into acting like a Finite State Machine, and deep down, this is just unsustainable hype for mission-critical workflows?<p>I’m looking for war stories and reality checks, not theoretical pitches.