Google co-founder Sergey Brin has returned to hands-on work at the firm and called on staff to double down on making Gemini AI models competitive, according to Reuters.

Citing two anonymous sources, the newswire reported that Brin has taken more of an active role in Gemini’s development over the past few months and in the immediate aftermath of Anthropic launching Claude Mythos urged Google DeepMind staff to work harder and faster to ensure Google keeps up with frontier competitors.

Last week, Sir Demis Hassabis stepped down as the chief executive of Google DeepMind and took on a new role as chief scientist at Alphabet. His exit puts Google DeepMind under the control of newly-appointed SVP Koray Kavukcuoglu, who formerly held the role of chief technology officer at the research lab and chief AI architect at Google.

The leadership shakeup also saw other senior figures leave including chief scientist Jeff Dean, Sanjay Ghemawat, Oriol Vinyals, and Quoc V. Le. All have been key to Google’s AI efforts from Google Translate to its custom silicon for AI training.

The Financial Times reported that the shift consolidated power away from Google DeepMind’s London HQ to California, as both Kavukcuoglu and Brin work from Google’s Mountain View HQ and there are now fewer senior figures with a direct say over Gemini

Brin founded Google with Larry Page in September 1998 and served as president first of the Search giant and later of its parent company Alphabet, until his retirement in 2019.

In 2023, Brin returned from retirement to work on AI at Alphabet in response to the launch of ChatGPT which the firm’s chief executive Sundar Pichai has since admitted caught it off guard. In April, the Information reported that Brin was leading a “Strike Team” aimed at improving Gemini’s coding capabilities, as models by Anthropic and OpenAI took the lead in the space.

Coding capabilities are seen as a key profit-driver for AI firms, as a core offering for enterprise customers. Anthropic and OpenAI offer their most capable coding models, Claude Mythos and OpenAI’s GPT-5.6 Cyber, as gated models available only to trusted organisations.

---