Demis Hassabis on Gemini 3, world models, and the AI bubble

Demis Hassabis was noticeably relaxed when he joined our virtual call from London.

It was the day before the release of Gemini 3, and Googlers all the way up to Sundar Pichai had been hyping the imminent model drop on X like they worked at OpenAI. Hours after our chat, Hassabis himself jumped in and posted the following in the middle of the night: “It’s nearly 3 here, my favourite part of the night shift… locked in... 💪🚀”

After some early fits and stumbles in the AI race, Google has been unquestionably locked in this year. By all early accounts, Gemini 3 Pro is the best model at just about everything. Meanwhile, the Gemini app has more than 650 million monthly users; over 2 billion people use Gemini via Search AI Overviews each month; and 13 million developers use Gemini in their products. Atop all this is Hassabis, who leads the company’s AI efforts as CEO of Google DeepMind.

During our interview, Hassabis addressed the previous shortcomings Google set out to address with Gemini 3, why the Gemini app has gained so much traction this year, why he’s spending a lot of time on world models, how Google evaluates compute bottlenecks, and his view of the AI bubble. “I feel like we’re in a really strong position to come out on top,” he told me when asked about the bubble discourse, “whether it carries on and there’s no bubble or there is one and there’s some retrenchment.”

The following conversation has been edited for clarity: