‘I Actually Thought He Was Going to Hit Me,’ OpenAI’s Greg Brockman Says of Elon Musk

In August 2017, Greg Brockman and Ilya Sutskever gathered at Elon Musk’s self-described “haunted mansion,” a 47-acre, $23 million estate in Hillsborough, south of San Francisco, to discuss the future of OpenAI. Actor Amber Heard, Musk’s then-girlfriend, had served the group whiskey and then dashed off with a friend, Brockman, OpenAI’s cofounder and president, testified in federal court during the trial for Musk v. Altman on Tuesday.

Ahead of the meeting, Musk gifted Brockman and Sutskever, OpenAI’s cofounder and former chief scientist, new Tesla Model 3 cars. “It felt like he was buttering us up,” Brockman said on the stand. “He wanted us to feel indebted to him in some way.” Sutskever tried to reciprocate for the occasion. The amateur artist presented Musk with a painting of a Tesla. Musk and the other cofounders wanted to establish a for-profit arm to entice investors to give them billions of dollars to pay for compute. But Musk also wanted control of the company, and Sutskever and Brockman objected to granting the Tesla CEO what they believed would be a “dictatorship” over the future of AI development. They proposed having shared control.

After several minutes of deliberation, Musk rejected their offer. “He stood up and stormed around the table,” Brockman recalled. “I actually thought he was going to hit me, physically attack me.” Musk grabbed the painting, said he would cut off his funding of the nonprofit until Brockman and Sutskever quit, and left the room, according to Brockman’s testimony. But that night, Musk’s so-called chief of staff Shivon Zilis called Brockman and Sutskever “to say it’s not over,” Brockman testified. “There were discussions of futures that included us.”

The story of the heated negotiations emerged as Brockman wrapped up his testimony on Tuesday. To OpenAI, the events at the mansion are representative of repeated instances of erratic behavior by Musk that they believe undermine his arguments about the company. Musk contends his roughly $38 million in donations to OpenAI were abused by Brockman and others on the path to creating the $852 billion for-profit venture now known for services such as ChatGPT and Codex. Brockman, OpenAI CEO Sam Altman, and OpenAI deny any wrongdoing, and the jury in Musk v. Altman could begin deliberating on an advisory ruling as soon as next week.

After Tuesday’s testimony, William Savitt, an attorney for OpenAI, told reporters that what Brockman had learned in 2017 was how tough it can be to meet one’s heroes. Brockman admired and respected Musk’s business acumen, but his desire for control was absolute and concerning, Savitt said. Marc Toberoff, an attorney for Musk, told reporters that the true concern was Brockman’s motivations for sharing control, with his desire for wealth having faced scrutiny in court a day earlier.

For his part, Brockman offered another story on Tuesday to underscore why he thought Musk was not up to the task of controlling an AI company. Brockman recalled then-OpenAI researcher Alec Radford showing Musk an early version of an AI chatbot that didn’t generate responses that he liked. Musk “kept saying this system is so stupid, that a kid on the internet could do better,” Brockman said. Radford “was absolutely crushed” and “demoralized” to the point that he almost quit the AI research field altogether, Brockman said. Brockman and Sutskever “spent a lot of time” rebuilding his confidence. Musk’s inability to see the potential in the early technology—which eventually became the basis for ChatGPT—made him unfit to control OpenAI, in Brockman’s view. “You needed to dream a little bit,” Brockman said. And Musk hadn’t shown that he could.

Boardroom Fights

Brockman said Tuesday that he, Sutskever, and Altman considered voting Musk off the OpenAI nonprofit board as negotiations with him about a for-profit sibling company dragged on for months. They would meet again over whiskey at Musk’s mansion to discuss alternative funding options. There was agreement over what not to do, but little on what to do instead. But Brockman and Sutskever decided removing Musk felt “wrong,” Brockman testified. Eventually, Musk left on his own after deeming OpenAI was on a path of “certain failure,” according to an email he wrote in early 2018.

Zilis, then an adviser to both OpenAI and Musk, kept him informed about developments at the AI venture in the years to come. “She was proxy Elon in some ways,” Brockman said, referring to her as “a friend” who he had first met in 2012 or 2013.

But she eventually faced her own reckoning. Zilis joined OpenAI’s board in 2020 and gave birth to Musk’s twins in 2021. She told Brockman about the children. But he learned that Musk was the father only later through news articles. When Brockman confronted her, “she said it was via IVF and it was entirely platonic with Elon,” he testified. Several board members wanted to remove Zilis from the board, but Brockman said that he and Sutskever convinced them to let her stay because she helped manage Musk’s frustrations with OpenAI. “We actually had a board vote,” he said.

Zilis left the OpenAI board in 2023 after Musk launched the rival lab xAI.

It wasn’t the only board member conflict OpenAI faced. Brockman said he supported the removal, or at least partial recusal, of Quora CEO Adam D’Angelo from the board after the knowledge-sharing platform launched a chatbot in February 2023 that competed with OpenAI’s recently released ChatGPT. D’Angelo remains on the board today. Brockman also supported the removal of AI safety researcher Helen Toner from the board, he said, without specifying a reason. She resigned from her role in 2023 after she helped fire Altman in a move that backfired and quickly led to his reinstatement. D’Angelo and Toner didn't immediately respond to WIRED’s requests for comment on the testimony.

How the boardroom drama played out from Musk’s side will likely get fresh airing on Wednesday. Zilis is expected to take the witness stand to advance his case. Musk’s attorneys had asked her testimony not be livestreamed out of security concerns for her and the four children that she now shares with Musk. But US District Judge Yvonne Gonzalez Rogers found there was no “specific compelling concern” that justifies “cutting off the transmission.”

This is an edition of Maxwell Zeff’s Model Behavior newsletter. Read previous newsletters here.