When was OpenAI founded, by whom, and what did it set out to do?

If you use GPT-3 through an app, or have read about OpenAI's Codex model writing code this month, you have met an organization that is not quite six years old. OpenAI introduced itself to the world on 11 December 2015 with a short blog post titled "Introducing OpenAI." It described a new kind of lab: a research company organized as a non-profit, with a mission framed around all of humanity rather than shareholders.

Knowing how OpenAI started helps explain many of the debates around it today, from how it releases its models to how it pays for the enormous computers they need.

What was announced, and who was involved

The post was signed by Greg Brockman, Ilya Sutskever and "the OpenAI team." Its opening line put the structure first: "OpenAI is a non-profit artificial intelligence research company." The goal was to push digital intelligence forward in whatever way would do the most good for humanity as a whole, without the pressure to earn a financial return.

The founding group looked like this:

Role	People
Research director	Ilya Sutskever, a deep learning researcher who had been at Google
Chief technology officer	Greg Brockman, formerly CTO of the payments company Stripe
Co-chairs	Sam Altman and Elon Musk
Other founding members (researchers and engineers)	Andrej Karpathy, Durk Kingma, John Schulman, Trevor Blackwell, Vicki Cheung, Wojciech Zaremba, Pamela Vagata
Advisors	Yoshua Bengio, Pieter Abbeel, Alan Kay, Vishal Sikka, Sergey Levine
Funders	Elon Musk, Sam Altman, Greg Brockman, Peter Thiel, Reid Hoffman, Jessica Livingston, YC Research, Infosys, Amazon Web Services

Together the funders committed $1 billion, though the post said the lab expected to spend only a tiny fraction of that in its first few years. The organization was based in San Francisco. According to accounts of its early days, it was first run out of Brockman's living room; Brockman had met with Yoshua Bengio to draw up a list of top researchers, and nine of them joined as the first employees that December.

What OpenAI said it would do

The post made the case in three steps.

AI had become surprisingly capable. Deep learning, an approach studied for decades, had recently started winning across many problems, including recognizing objects in images, translating languages and recognizing speech. Instead of hand-coding a new algorithm for every task, researchers designed networks that shaped themselves to the data.
The stakes could become very high. Today's systems were narrow, the post said, but they might eventually reach human performance on almost every intellectual task. Such systems could benefit society enormously, or damage it badly if built or used carelessly.
So a different kind of institution was needed. Because no one could predict when human-level AI might arrive, it would be important to have a leading lab that put a good outcome for everyone ahead of its own interests.

To that end, OpenAI promised openness. Its researchers would be strongly encouraged to publish papers, blog posts and code; any patents would be shared with the world; and it would collaborate freely with other institutions. The post also said AI should extend what individual people want to do, and that its benefits should be spread as widely and evenly as safety allows.

Try it: Read the original post yourself through the Internet Archive's snapshot of "Introducing OpenAI" (it takes about three minutes). Write down the three promises you think matter most, then compare them with how OpenAI released GPT-2 and GPT-3. Which promises does it still keep in the same form?

Why it mattered

In 2015, much of the most visible deep learning research came from large technology companies, Google above all, which had bought the London lab DeepMind the year before. OpenAI's launch put a large, well-funded lab outside those companies, with a public mission about long-term safety and broad benefit. It attracted strong researchers even though it did not match the salaries that Google or Facebook paid.

Its early work reflected the open approach. In April 2016 it released a public beta of OpenAI Gym, a free toolkit for reinforcement learning research (teaching software agents through rewards), and in December 2016 it released Universe, a platform for training agents on games and websites. In September 2016 it moved into the Pioneer Building in San Francisco's Mission District.

How has OpenAI changed since its founding?

Quite a lot, and the changes are themselves part of the story. Elon Musk left OpenAI's board in 2018. In 2019, OpenAI created a new company, OpenAI LP, under the control of the non-profit. It is a "capped-profit" company: investors can earn returns, but only up to a limit of 100 times their investment. OpenAI said the change would let it raise the money needed for large-scale computing and offer employees equity like its competitors. Microsoft then invested $1 billion, and OpenAI's systems moved to Microsoft's Azure cloud. Some researchers have argued that the shift fits poorly with the original non-profit vision.

The promise of openness has also been reinterpreted. In February 2019, OpenAI announced GPT-2 but initially withheld the full model, citing the risk of misuse, and released it in stages over that year. GPT-3, described in May 2020, was offered not as a download but through an API, with access granted by application. This month, OpenAI began offering Codex, a GPT-3 descendant trained on code, through the same API with a waitlist; Codex also powers GitHub's Copilot preview.

What has stayed constant is the mission statement, which still speaks of benefiting all of humanity. Whether the current structure serves that mission better than the 2015 design is one of the most discussed questions in AI today.