Pay packages in the AI industry are some of the most generous in tech. Salaries in the U.K. government… not so much.
So Jade Leung took a big pay cut when, in October 2023, she quit her job at OpenAI to become chief technology officer to the U.K. government’s new AI Safety Institute (AISI). Established in late 2023, the U.K. AISI has in a short time become the leading government body, anywhere in the world, dedicated to testing the safety of the most powerful AI systems.
[time-brightcove not-tgx=”true”]The move was worth it, Leung says, because of the opportunity to shape the way that AI systems are governed. “You really want a public interest body that is genuinely representing people to be making those decisions,” she says. “By definition, they need to be [taken] outside of the organizations developing” AI models.
Leung’s role at the U.K. AISI has largely been focused on building it into a body that can carry out state-of-the-art safety research. Alongside staffing up the nascent organization, her key responsibility is designing and overseeing evaluations that can test whether AI models pose the risk of being used to facilitate cyber, biological, or chemical attacks. The organization has also been developing a more speculative set of tests designed to check whether AI systems have the ability to escape the control of their designers. The U.K. government has negotiated early access to new models from leading AI companies to carry out safety testing on them; the AISI has completed two such tests already, one of them on Anthropic’s Claude 3.5 Sonnet model. “We are currently, I would say, the best single place that you could go to to have your models evaluated across a range of frontier AI risk areas in a fairly robust and in depth way,” Leung says.
