Deepmind plans to release Gemini 4 within months

Koray Kavukcuoglu, the new head of Google Deepmind, announced that the team plans to release Gemini 4 well before the end of the year. He made this statement during his first appearance at The Information’s AI Agenda Live Summit. Currently, Gemini 4 is in an early post-training phase, meaning the model is already being tested internally in an environment called Antigravity. Key to this progress are the rapid iterations following the release of the first post-training version, allowing Google to position itself against competitors and refine its product offerings.

Kavukcuoglu emphasizes that the focus is shifting from the quest for Artificial General Intelligence (AGI) to developing reliable agents. This approach reflects a new priority within Deepmind, where safety and reliability are paramount. Unlike former CEO Demis Hassabis, who focused on AGI for years, Kavukcuoglu believes it is crucial to build agents that are actually trustworthy.

Meanwhile, competitors such as Anthropic and OpenAI are rapidly advancing, putting pressure on Google to respond quickly. Kavukcuoglu is confident that his team, which he describes as "insanely strong," can catch up. However, it remains unclear whether Gemini 4 can make the impact needed to place Google back at the top of the AI market. Additionally, Google has relocated its development activities to the Bay Area and the company underwent significant restructuring, altering Deepmind’s role within Google.

The focus on new models will also play a crucial role in chip development within Google. By selling TPUs directly to customers, Google hopes to position itself closer to the advancements of providers like Nvidia. The changes within the team and the greater emphasis on product development and infrastructure indicate that Google is significantly adjusting its strategy in light of the evolving AI market.

Read the full article from AI Insider.