Empire of AI: Dreams and Nightmares in Sam Altman's OpenAI

On October 31, 2023, Toner, McCauley, and D’Angelo met to discuss next steps. They resolved to fire Altman and replace him with Murati as interim CEO. They conducted an investigation and learned that Altman routinely lied or dissembled and that several employees “described Altman’s behaviors as abuse and manipulation” (361). They also felt he did not have the experience to lead a “mature” and growing company. On November 16, Murati confirmed her support for their plan.

On Friday, November 17, the board publicly announced their plan. They were shocked by the immediate pushback from OpenAI employees and leadership. Murati declined to publicly support the coup attempt. Sutskever also folded in the face of ongoing pressure, feeling that continued criticism would result in the end of OpenAI. Elon Musk weighed in by sharing an open letter from, presumably, former members of the Safety team who felt that Altman and Brockman did not take AI “safety” seriously. By Monday, the board recognized that it had lost, and Altman was reinstated as CEO. The Safety team felt “betrayed by the board” (372).

Two weeks later, Altman held an all-hands meeting to discuss a project of Sutskever’s, who was notably absent from the meeting. They had developed an algorithm called Q* that was capable of more efficient inferences than traditional AI models. Although OpenAI described Q* as a “breakthrough,” it refused to release any of its data for testing.

On March 8, 2024, the board’s investigation into Altman ended with the conclusion he had not done anything that warranted his firing.

After Altman’s reinstatement, the Doomers grew increasingly concerned about OpenAI’s plans to “create an AI chip company” (377) and a digital assistant modeled after the Scarlett Johannsen character in the Spike Jonze movie Her (2013). In 2023, with a breakthrough in voice capabilities, OpenAI debuted an audio AI system called Omni that used GPT-4o, the newest iteration of the model.

Meanwhile, Altman’s public relations approach became increasingly cutthroat. Hao notes, “if Altman was being brazen and boastful, most likely something wasn’t going well” (383). OpenAI was facing increasing competition from rival AI companies like Anthropic and Google. He was also facing lawsuits from publishers for stealing their text for training data and pressure from Microsoft to earn more money. Employees felt that the company’s growing insularity was antithetical to its original goal of benefiting humanity.

After the GPT-4o demonstration, OpenAI announced that Sutskever was leaving the company. Soon after, the leader of the Superalignment project, Jan Leike, left to join Anthropic. Leike’s departure was seen as the continued minimization of “safety” within OpenAI. On May 17, an EA-affiliated Vox journalist, Kelsey Piper, reported that researcher Daniel Kokotajlo, who had worked in Safety, had been pressured to sign an non-disclosure agreement (NDA) when he left the company or face losing his equity. The documents even stated that if he spoke out after leaving OpenAI, the company would “claw back” his vested equity. Learning about this clause prompted employee outrage. On May 20, actor Scarlett Johansson sued OpenAI for using her voice for the GPT-4o system. In a meeting on May 22, OpenAI leadership claimed, unconvincingly, that it was unaware of the “claw back” clause and assured employees they were going to fix it. As the controversies mounted, Murati, Brockman, and Pachocki begged Sutskever to return to the company, but the next day they retracted their offer. Instead of reflecting during this moment of “Omnicrisis,” Hao argues that OpenAI doubled down and expected it all to blow over.

Hao argues that Altman perverts values like openness in the furtherance of his “empire.” She argues that he does this by “centralizing talent” around a vision, centralizing resources while overriding regulation, and keeping the mission “vague” so it can be directed in whatever way the “centralizer” desires. For instance, the meaning of AGI keeps changing depending on the needs of the company.

In May 2024, pressure for greater transparency within OpenAI and the AI industry more broadly continued. Tenured executives, including Murati, began to leave the company. Sutskever started his own company, Safe Superintelligence. OpenAI’s scaling efforts hit a wall. In October 2024, OpenAI’s investors instituted a new caveat that they “could demand their money back if the company did not convert into a for-profit in two years” (405). At the end of 2024, OpenAI announced they would convert into a “for-profit public benefit corporation” with an attached non-profit arm with shares in the for-profit. Altman remained publicly confident about the future of OpenAI and its development of “superintelligence.”

In the Epilogue, Hao describes a Māori AI language project she argues is an example of an alternative to OpenAI’s “imperial” form of AI. The te reo Māori language is considered endangered due to decades of its repression by English colonialists in New Zealand. Indigenous couple Peter-Lucas Jones and Keoni Mahelona turned to AI to help preserve and promote the te reo language. They used recordings from Te Hiku radio, a Māori radio station, and made new recordings of Māori elders to capture as much of the language as they could. Then, with the consent of those they had recorded, they trained a limited AI on that dataset to be used to teach and promote the language. They have an agreement to only share the model with pre-vetted licensed vendors which will benefit Māori people, and they run the model on local databases. It is a “small, specialized model that excels at one thing” in contrast with OpenAI models which seek to “[hoover] up as much data as possible” and centralize it (412).

Hao argues that AI need not be built along “imperial” lines like OpenAI; rather, that it can be built on a smaller scale that benefits people without exploiting them. Google critic Timnit Gebru and her coauthor Alex Hanna have continued to work on decentralizing AI through their Distributed AI Research Institute (DAIR). In Kenya, data workers like Mophat Okinyi continue to advocate for better labor rights for data workers in the developing world. In Uruguay, activists like Daniel Pena continue to fight against data centers that exploit the local natural resources. Hao presents these as examples of anti-imperial approaches to AI. She draws on queer AI researcher Ria Kalluri’s framework of challenging centralization of knowledge, resources, and influence throughout the AI sector. She argues that the focus should not be on “the theoretical rogue AI harms of Doomerism, but the existing real-world harms” (419).

Hao argues that there should be more transparency into how AI works, better labor protections for AI workers and those whose creative works are used as data, and better education to demystify AI.