ChatGPT 发布

ChatGPT 的发布文章没有为它安排盛大舞台。2022 年 11 月 30 日，OpenAI 的开头只说：我们训练了一个可以用对话方式交互的模型，叫 ChatGPT。研究预览免费，旁边是一条试用链接。没有倒计时，没有参数量口号，也没有要求读者先读完一篇论文。

界面做的减法近乎彻底。用户输入自然语言，模型回答；不对，可以接着指出，不必把全部背景重写一遍。对话历史让“解释简单一点”“换一种写法”成为完整指令。此前，同类能力多半藏在 API、研究演示或自管状态的脚本里；这一次，GPT-3.5 与 InstructGPT 路线上的成果，被装进人人熟悉的消息往来。用户不必先理解 token、上下文窗口或 RLHF，也能开始使用；这三个词却已经写在训练与产品设计里。

开放当天没有权重下载，也没有要求信用卡。账号与地区政策决定能否进入；速率与容量限制决定排队与拒绝。知识截止日期、无法联网（首日）、安全策略的误拒与漏拒，与能力宣传写在同一产品页上。后来的增长故事常被拿来定义这一天；发布页上的内容是界面、示例错误、局限清单与反馈入口。

发布页上的代码示例保留了两面。用户贴出 Go 代码，怀疑 channel 有问题；模型语气流畅、步骤完整，关键观察却错了——原代码已有 defer close。错误没有藏进附录，而是与产品一起出现在介绍里。OpenAI 也直接列出：听来可信但错误的答案、过度冗长、含混时爱猜测、安全拒绝既有漏网也有误伤。点赞、点踩把公众交互接回改进回路。研究预览名副其实：公众既是用户，也是持续标注的一端。

它的日常性正是事件核心——一个会犯错、却足够配合的模型停在浏览器里，等人发下一条消息。发布页已写出它会答错；后来发生的事，是无数人仍愿意再问一次，因为继续追问的成本低到几乎可以忽略。

没有盛大发布会，反而让产品更像日常工具：打开网页，打一行字。教育者、程序员、写作者与好奇者在同一会话里试问、反驳、重写。增长神话会来，但 11 月 30 日的页面上还没有那些曲线——只有试用链接、局限清单，以及一个已经会答错的示例。把事件写准确，就要把“不好看”的部分留在正文里。

ChatGPT’s launch post did not build a grand stage. On 30 November 2022, OpenAI began simply: we trained a model that can interact in a conversational way, called ChatGPT. Free research preview; a try link beside the text. No countdown, no parameter slogan, no demand that readers finish a paper first.

The interface’s subtraction was nearly total. Users typed natural language; the model answered; if wrong, they could point it out without rewriting all context. Conversation history made “explain more simply” and “try another phrasing” complete instructions. Similar capability had mostly lived in APIs, research demos, or scripts that managed state; here, work along the GPT-3.5 and InstructGPT path entered a form everyone already knew—message exchange. Users need not first understand tokens, context windows, or RLHF to begin; those three terms were already written into training and product design.

Launch day brought no weight download and no credit-card requirement. Account and regional policy decided entry; rate and capacity limits decided queues and refusals. Knowledge cutoffs, no live web (day one), and safety policies that both over- and under-refused sat on the same product page as capability claims. Later growth stories often define the day; the page itself held interface, sample errors, limitation lists, and feedback entry points.

A code sample on the page kept both faces. A user pasted Go, suspecting a channel bug; the model’s tone was fluent, steps complete, the critical observation wrong—the original already had defer close. The error was not buried in an appendix; it appeared with the product. OpenAI also listed, plainly: answers that sound true but are false or absurd; over-verbosity; guessing under ambiguity; safety refusals that miss and that over-block. Thumbs up and down wired public interaction back into the improvement loop. Research preview was literal: the public was both user and ongoing annotation.

Everydayness is the event’s core—a model that errs yet cooperates enough sits in a browser, waiting for the next message. The launch page already said it would be wrong; what followed is that countless people still asked again, because the cost of another turn was almost nothing.

No grand stage made the product feel more like a daily tool: open a page, type a line. Educators, programmers, writers, and the merely curious tried, pushed back, and rewrote inside one session. Growth myths would arrive, but the 30 November page did not yet hold those curves—only a try link, a limitations list, and a sample already wrong. To write the event accurately is to keep the unflattering parts in the body.