Jan Leike

Jan Leike wrote a PhD on the theory of general reinforcement learning under Marcus Hutter at the Australian National University, spent six months at Oxford's Future of Humanity Institute, and joined DeepMind's technical safety team in 2016. With OpenAI's Paul Christiano and Dario Amodei, Leike co-wrote the June 2017 paper that trained agents from people's choices between pairs of video clips, the method later called reinforcement learning from human feedback, and in 2018 set out a research agenda for aligning agents through learned reward models.

Leike moved to OpenAI in early 2021 to lead its alignment team, which applied that method to language: InstructGPT (January 2022) was the first large model OpenAI trained on human rankings, and ChatGPT followed on the same recipe. In July 2023 Leike and Ilya Sutskever launched the Superalignment team, promised 20 percent of OpenAI's secured compute to solve the alignment of superintelligence within four years. TIME named Leike to its first TIME100 AI list that September.

Leike announced a resignation on 15 May 2024, the day after Sutskever announced his departure, and two days later explained on X, "I have been disagreeing with OpenAI leadership about the company's core priorities for quite some time, until we finally reached a breaking point." The thread said the team had been "struggling for compute." Sam Altman replied that "he's right we have a lot more to do; we are committed to doing it," and Altman and Greg Brockman posted that OpenAI had raised awareness of AGI's risks and would keep doing safety research. OpenAI dissolved the Superalignment team within days. On 28 May Leike joined Anthropic to lead work on scalable oversight, weak-to-strong generalization and automated alignment research. On 8 May 2026 Leike announced a new research project at Anthropic, writing that "many things are needed to make AGI go well, and alignment is only one of them," and stepped back from running the alignment team.