Winston Weinberg was a first-year associate at O'Melveny & Myers, a securities and antitrust litigator, when he was handed a landlord-tenant case in a field he did not know. His roommate in Los Angeles, Gabriel Pereyra, a research scientist who had worked at Google DeepMind and Meta, showed him GPT-3. Weinberg's first use of it, he has said, was running a Dungeons & Dragons game for friends.
Then he pointed it at the law.
They had cold-emailed Sam Altman and OpenAI's general counsel, and the pitch reached OpenAI's leadership on a July 4 call.
They wrote a long prompt about California tenant statutes and ran it on 100 questions from a public legal-advice forum. Three attorneys reviewed the answers without knowing a model wrote them. On 86 of the 100, at least two said they would send the answer to a client unedited.
That got them a call. They had cold-emailed Sam Altman and OpenAI's general counsel, and the pitch reached OpenAI's leadership on a July 4 call. The OpenAI Startup Fund wrote a $5 million cheque in late 2022. The name came from Harvey Specter, the lead character of the TV drama Suits.
Weinberg has been candid about how foreign it all was. He had no tech friends, did not grow up in San Francisco and did not know how fundraising worked. What he knew was how lawyers think. Early sales used public litigation briefs: Harvey showed the lawyers who had written them how to argue against their own work. Allen & Overy began a trial in November 2022 that reportedly reached 3,500 lawyers.
Here is the part that rarely gets told. In early 2024, with Harvey valued around $700 million, Weinberg pursued a merger with a more traditional company. It fell apart because Harvey would have taken on too much debt to pay for it. He has recalled a week spent asking what the company's future should be.
He has since made failure a working rule. He calls it "a very good way to learn", describes the job as "destroying your ego 24/7", and says he points out about 15 failures a day to his own staff, who must re-earn their roles every six months.
The numbers kept moving. A $715 million valuation in December 2023 became $3 billion in February 2025, $8 billion in December 2025 and $15.5 billion in September 2026. Revenue passed $100 million ARR in August 2025 and was about $190 million at the end of the year.
The lesson sits in the near-miss. The deal that nearly happened would have bought size with debt. The company that exists instead was built on a test a stranger could grade, and on a founder who says out loud how often he is wrong.