Google Presents Gemini Upgrades Aimed at Advancing Toward General AI

At Google’s 2025 I/O event, CEO of DeepMind Demis Hassabis highlighted the company’s advancements in artificial intelligence, particularly through its Gemini models. These models are increasingly showing capabilities such as reasoning, agency, and modeling the physical world—key components in the pursuit of Artificial General Intelligence (AGI). Google unveiled new versions of Gemini Flash and Gemini Pro, with the latter outperforming rivals in benchmark tests. A new feature, Deep Think, enhances simulated reasoning.

Practical applications of Gemini’s intelligence were introduced, including Mariner, an AI agent for the Chrome browser that can perform tasks like online shopping, and Astra, a virtual assistant capable of perceiving and interacting with the real world through smart devices. Google also announced plans to adapt its core search engine with AI-enhanced tools like AI Mode and AI Overviews.

Hassabis believes that achieving AGI could still take 5 to 10 years and will require mastering inventiveness and environmental understanding—skills essential for real-world applications such as humanoid robots. DeepMind is collaborating with robotics companies and experimenting with coding agents like AlphaEvolve and game-based learning environments to enhance AI creativity and reasoning.