OpenAI 成立

2015 年 12 月，一家新的 AI 研究机构宣布成立，名字简单直接：OpenAI。发起名单星光熠熠——Sam Altman、Elon Musk、Ilya Sutskever、Greg Brockman，加上一批来自各大实验室的顶尖研究者。它宣称的使命只有一句话，却分量十足：以开放的方式推进 AGI，让全人类受益。初始承诺资金约 10 亿美元。

OpenAI 的成立背景，是 2015 年前后 AI 圈弥漫的复杂情绪。深度学习的势头已经清晰，但很多人担心两件事：AI 技术可能被几家巨头垄断，成为商业工具；或者 AGI 在不被理解的情况下失控。OpenAI 想提供的答案，是一个不受商业利益支配、以安全和开放为优先的研究力量。这份理想主义，是它最初的立身之本。

早期 OpenAI 的开放姿态是实打实的。它开源了强化学习工具 Gym，开源了机器人训练环境 Universe，论文与代码公开程度在业界领先。研究方向上，它一边做深度强化学习——比如训练 Dota 的 OpenAI Five——一边在语言模型上投入。2018 年发表的 GPT-1，就诞生在这个「什么都研究、什么都开源」的阶段。

但转折来得很快。2018 年，Elon Musk 因与团队理念分歧离开；2019 年，OpenAI 成立营利实体，微软投资 10 亿美元。曾经的「非营利+完全开放」路线，被调整为「非营利母体+有上限利润的营利体」结构。GPT-2 的发布也开始采用分阶段公开而非全量开源。这些变化在当时引发不少讨论——「OpenAI 还 Open 吗」成为公众议题。

后来的故事已经众所周知：GPT-3、ChatGPT、GPT-4、GPT-5 系列让 OpenAI 站上生成式 AI 的中心，估值一路冲到数千亿美元。「让全人类受益」的宏大承诺，与商业公司的现实经营之间，始终存在张力。但无论外界如何评价，一个事实无法回避：2015 年那间小小的非营利实验室，确实改变了 AI 的进程。

今天回看 OpenAI 的成立，它最值得记住的，或许不是后来的任何一款产品，而是那个最初的姿态——一帮人相信，AGI 应该被以安全和开放的方式追求，而且值得为此投入十亿美元级别的资源。这个姿态后来被现实不断修正，但它的确把「AI 安全」与「AGI」两个词，从学术圈的小众讨论，变成了全世界的公共话题。

In December 2015 a new AI research institution announced itself with a plain name: OpenAI. The founding list glittered—Sam Altman, Elon Musk, Ilya Sutskever, Greg Brockman, plus top researchers from major labs. Its stated mission was one line with heavy weight: advance AGI in an open way that benefits all of humanity. Initial pledged funding was around $1 billion.

OpenAI's founding context was the complex mood of the AI community around 2015. Deep learning's momentum was clear, but many worried about two things: AI might be monopolized by a few giants and become a commercial tool, or AGI might spiral out of control before it was understood. OpenAI's answer was to be a research force free of commercial interests, prioritizing safety and openness. That idealism was its original identity.

Early OpenAI was genuinely open. It open-sourced the reinforcement-learning toolkit Gym and the robot-training environment Universe; papers and code were public to a rare degree. In research it ran both deep reinforcement learning—training Dota's OpenAI Five—and language models. The GPT-1 paper in 2018 was born in this "research everything, open-source everything" phase.

The turning point came fast. In 2018 Elon Musk left over philosophical differences; in 2019 OpenAI created a for-profit entity and Microsoft invested $1 billion. The former "nonprofit plus full openness" path was adjusted into a structure of a nonprofit parent plus a capped-profit arm. GPT-2 also began releasing in stages rather than fully open. These changes sparked debate—"is OpenAI still open?" became a public question.

The rest is well known: GPT-3, ChatGPT, GPT-4, and the GPT-5 line put OpenAI at the center of generative AI, with valuations climbing into the hundreds of billions. The grand promise of "benefiting all humanity" has always coexisted in tension with running a real company. But no matter how one judges it, one fact stands: the small nonprofit lab of 2015 did change the trajectory of AI.

Looking back at OpenAI's founding, what deserves remembering is less any later product and more the original posture—a group of people believing AGI should be pursued safely and openly, and that it was worth committing billion-dollar-scale resources. Reality kept revising that posture, but it did push the two words "AI safety" and "AGI" from niche academic discussion into worldwide public conversation.