How Hasabis led Deepmind to a counterattack: from the Gaia project to Gemini 3 anti-OpenAI

At the end of 2025, Google released Gemini 3, the first time that a new model of OpenAI had been levelled over core indicators. The emergence of this situation marked a critical turning point for Deepmind, after years of strategic confusion and technological pursuit.

The scientific idealism of Hassabis contrasts sharply with the commercial pragmatism of OpenAI’s leader, Sam Otman. Sebastián Malabi, a chronic contact with Hasabis, revealed that Khashabis’ public assessment of Ottman was quite direct: "Otman seeks power, and I pursue knowledge and science." However, this sense of value superiority is not a substitute for a positive competition at the technological and product levels.

Hassabis, since the creation of Deepmind, has placed enhanced learning at the heart of the research system, treating language as a source of data rather than as an independent research direction. He believes that universal artificial intelligence (AGI) can eventually be achieved by training intelligent people in a simulated environment – a concept that has given rise to a costly Gaia project – which, after two or three years, has failed to produce tangible results.

At the same time, Google self-researching the efficiency of TPU chips began to emerge, laying the foundation for a facility-level follow-up counter-surveillance. During this phase, Hasabis himself demonstrated leadership that had not previously been fully recognized – a transition from a purely scientific role to an integrated management that integrated thousands of people's research and development teams and coordinated engineering and scientific relationships.

The publication of Gemini 3 was an acceptance of the results of this adjustment period. The model exceeded GPT-4 partially in terms of key indicators such as reasoning, multimodular understanding, long context windows. Marabi quoted the words of Hassabis: "Gemini is now in the lead." When he said that, his smile "can't hide it."

In his dialogue with Malabi, he clearly identified himself as a scientist rather than an entrepreneur. He admitted that he did not have the same business outreach and capital capacity as Otman, but believed that his strengths were: adherence to AGI’s long-term objectives, commitment to interdisciplinary research methods, and the computational and talent resources available on Google platforms. He judged that it would take five to ten years to achieve AGI, and one or two theoretical breakthroughs at the level of Transformer, which Deepmind still considered to be the best platform to achieve.

In terms of industry patterns, Deepmind’s backlash has indirectly inspired the AI industry in China: in the era of large models, arithmetic infrastructure, synergic optimization of models and data, and close collaboration between research and engineering remain the central determinants of competitive success. Although Hasabis is home to scientists, his strategic patience and organizational ability to lead DeepMind in the transition from research institutions to integrated AI laboratories is perhaps more interesting than any single technological breakthrough.

(Composite titanium media report, 28 March 2026)