OpenAI 推出自定义 GPTs

2023 年 11 月 6 日的 DevDay 上，做一个“自己的 ChatGPT”看起来只需要几步。写下指令，上传资料，选择浏览、代码解释器等能力，必要时再用 Actions 连接外部 API。

GPT Builder 把名称、说明、知识和工具收成一个可保存、可分享的对象。创建按钮离用户很近，训练模型所需的机器则完全不在这条路径上。

GPTs 回应了一种真实需求：很多人并不想开发完整应用，只想让 ChatGPT 稳定围绕一组资料和规则工作。教师可以放入课程材料，团队可以放入内部说明，个人可以保存反复使用的工作方法。定制从“每次重新写一遍提示”变成一个有名字的产品对象，而且不需要重新训练模型。

这种轻量也划出了能力边界。上传文件不等于建立了可靠知识系统，Actions 能调用 API 也不等于权限、鉴权和失败补偿已经处理好。一个 GPT 可以看起来非常专用，实际只有一段薄指令和几份未经整理的文件。配置层缩短了搭建时间，没有替创建者回答它为谁服务、数据多久更新、错误由谁发现。

早期创建面向 ChatGPT Plus 与 Enterprise 用户，并非所有人同时获得入口。更关键的是，创建与分发没有在同一天完成。原计划 11 月上线的 GPT Store 延至 2024 年 1 月开放。没有商店时，GPT 主要通过链接分享；商店出现后，新的难题变成发现、审核、排名和推荐。会做出一个助手，与让合适的人长期找到并使用它，是两套产品能力。

门槛降低后，同质内容很快堆积。许多 GPT 只在名称和提示上做轻微变化，用户很难在拥挤货架中判断质量。商店可以组织目录，也把流量分配权交给平台。创建者面对的不再只是“能否做出来”，还包括能否形成稳定场景、能否持续维护资料，以及推荐机制是否愿意把它展示给别人。

GPTs 没有把每个用户都变成软件公司，却让“配置一个助手”成为普通产品动作。它留下的反差也很清楚：几分钟就能点亮一张新卡片，真正困难的是几个月以后，那张卡片是否仍有准确的资料、清楚的权限，以及一个愿意再次打开它的人。

At DevDay on 6 November 2023, making “your own ChatGPT” appeared to require only a few steps. Write instructions, upload material, select capabilities such as browsing or Code Interpreter, and connect an external API through Actions when needed.

GPT Builder folded a name, description, knowledge, and tools into an object that could be saved and shared. The create button was close to the user; the machines required to train a model were absent from the path.

GPTs answered a real demand. Many people did not want to develop a complete application. They wanted ChatGPT to work consistently around a set of documents and rules. A teacher could provide course material, a team could add internal guidance, and an individual could preserve a repeated method. Customization moved from “rewrite the prompt every time” into a named product object, without retraining the model.

The lightness also exposed the boundary. Uploading files did not create a dependable knowledge system. Connecting an Action did not resolve authorization, authentication, or compensation after failure. A GPT could look highly specialized while containing only a thin instruction and a few disorganized files. The configuration layer shortened setup. It did not answer who the assistant served, how often its data changed, or who would notice an error.

Early creation access targeted ChatGPT Plus and Enterprise users rather than everyone at once. More importantly, creation and distribution did not arrive together. The GPT Store, originally planned for November, opened in January 2024. Without the Store, a GPT traveled mainly through a shared link. With the Store, discovery, review, ranking, and recommendation became new problems. Building an assistant and helping the right people find and repeatedly use it were two separate product capabilities.

As the barrier fell, similar artifacts accumulated quickly. Many GPTs varied little beyond a name and a prompt, leaving users to judge quality on a crowded shelf. A store could organize the catalog and also give the platform power over traffic. A creator now had to ask not only whether an assistant could be made, but whether it addressed a durable use case, whether its material would be maintained, and whether recommendation systems would show it to anyone.

GPTs did not turn every user into a software company. They made “configure an assistant” an ordinary product action. The contrast they left is sharper: a new card can light up in minutes; the difficult part is whether, months later, that card still contains accurate material, intelligible permissions, and a person who wants to open it again.