@aigclink: 最近在参与国内某头部医院检验科训练一个垂直细分医疗模型，分享一些个人观点，给AI+医疗领域的从业者一些建议（觉得有用拿走、觉得没用底下留言喷）： 1、医疗领域最核心的是数据，国内的公立及三甲医院基本上数据不出医院的，所以细分领域AI落地有…

最近在参与国内某头部医院检验科训练一个垂直细分医疗模型，分享一些个人观点，给AI+医疗领域的从业者一些建议（觉得有用拿走、觉得没用底下留言喷）：

1、医疗领域最核心的是数据，国内的公立及三甲医院基本上数据不出医院的，所以细分领域AI落地有非常多的机会，几乎每个科室有可以训练一个小模型来替代相关场景

2、通用医疗模型，国内百川的模型已经做得很好了，基础医疗初诊基本上没太大问题了，而且逐步接近中级，大家最认可的是证据链这个东西。这个是很多头部医院医生的反馈。我自己亲自试了下，确实挺不错的，慢性病、各种医疗的小问题，比好大夫、丁香园等应用上的地方医院医生要强，至少在综合性判断上效果要比一般医生强。

3、医院使用大模型最大的障碍是本地化成本问题，因为数据不能出医院所以无法用云，本地化就得涉及成本硬件，这个是阻碍很多医院深入使用大模型的核心原因。

4、短期内，医院为软件付费意愿很弱，搭配硬件更有机会拿到医院的直接付费，所以与医疗硬件厂商绑定是AI软件类公司的一条出路。

5、医疗AI有非常多可做的方向，数据伦理是最大的门槛，进入后数据标注也是最大的成本，这种数据需要专业医生来标注才行，普通人标注没啥鸟用，所以0-1很关键，几乎没有捷径。

#ai医疗 #百川医疗

Recently participating in training a vertically specialized medical model for the laboratory department of a leading domestic hospital. Sharing some personal insights and suggestions for practitioners in the AI + healthcare field (take it if you find it useful, leave a comment below to roast if you don’t):

1. Data is the core of the medical field. In China, public and top-tier (Tier 3A) hospitals basically keep their data within the hospital, so there are tons of opportunities for AI to land in niche areas—almost every department can train a small model to replace relevant scenarios.

2. For general medical models, Baichuan’s model has already done a great job domestically. Basic medical initial consultations are pretty much no longer a big issue, and it’s gradually approaching intermediate level. What everyone values most is the evidence chain. This is feedback from doctors at many leading hospitals. I tried it myself, and it’s indeed pretty good—better than local hospital doctors on apps like Haodf and Dxy for chronic conditions and various minor medical issues, at least in terms of comprehensive judgment, the effect is stronger than average doctors.

3. The biggest barrier to hospitals using large models is the localization cost issue. Since data can’t leave the hospital, cloud services can’t be used, so localization involves hardware costs—this is the core reason hindering many hospitals from deeply adopting large models.

4. In the short term, hospitals have very weak willingness to pay for software alone. Pairing with hardware offers a better chance to get direct payments from hospitals, so binding with medical hardware vendors is one viable path for AI software companies.

5. There are tons of directions to explore in medical AI. Data ethics is the biggest threshold; once you’re in, data annotation is the biggest cost. This kind of data requires professional doctors for annotation—annotations by regular people are pretty much useless. So getting from 0 to 1 is crucial; there are almost no shortcuts.

#ai医疗 #百川医疗

部分医生的反馈，当然也不一定代表所有医生

是，伦理是第一道门槛

大多数his给医院弄的自己都不知道是啥情况，压根就拿不出接口啥的

有人把一整个营销团队——CRO、文案、SEO、增长、定价、GTM、策略——打包成了 47 个 AI skill，装上之后，你的 Claude Code 就能当半个 CMO。

所有 skill 的地基是 product-marketing：任何其他 skill 动手前，都先读它，拿到“你的产品、受众、定位“。skill 间互相引用：copywriting cro ab-testing、customer-research → copywriting / cro / competitors，等于把一个真实营销团队的“共享上下文 + 分工协作SOP复刻进了 agent。同时，每个 skill 又都带 references/（方法库）和 evals/（真的做了评测）——是生产级的，不是攒一堆提示词。

几个特别的skill：

1、marketing-council：模拟一个“顾问董事会“，给你多个专家视角； 2、marketing-loops：让 agent 按周期自动跑的循环营销工作流（又是 loop / 自运行那套）； 3、ai-seo：专门优化“被 LLM 引用、出现在 AI 生成的答案里“——非常应景； 4、programmatic-seo：模板 + 数据批量出页。

“用这套 skill 当你自主 AI CMO“的 agent，装上这套skill，就能替代外包一半的营销活。

结合上周还在谈的“设计品味能不能打包成 skill“（hallmark、emilkowalski），这周整个营销职能也被系统化装进了 agent，两个结合就能跑出小的MVP。

#MarketingSkills #ClaudeCode #AI营销 #AgentSkills #增长