DeepMind, the AI-focused division of Google, has been at the forefront of significant advancements in artificial intelligence technologies. Many of these innovations have been integrated into various Google products, while others are still under development and awaiting broader rollout.

One prominent project poised for improvements is Gemini. In a recent interview, Demis Hassabis, CEO of Google DeepMind, shared insights into the company’s future direction in AI.

During an interview on 60 Minutes, Hassabis discussed the fascinating features of Project Astra, which is an advanced multimodal conversational AI experience. Notably, the AI was able to recognize the interviewer from a previous meeting, a testament to its advanced capabilities.

As Project Astra’s features rolled out in Gemini, the chatbot recently acquired multimodal abilities, allowing it to interpret visual information in real-time. However, the capacity to remember past interactions is not yet available in the current version of Gemini.

Hassabis revealed that developers are working on a version with enhanced memory capabilities that can recall previous conversations, potentially bringing a “10-minute memory” feature to Gemini Live in the future. Hassabis also hinted at the AI assistant’s development towards becoming more “agentic,” stating that DeepMind is training Gemini not only to observe the world but to interact with it actively.

This would enable tasks such as making online purchases or booking tickets. While these developments are exciting, he reiterated that true artificial general intelligence (AGI) is still several years away, with a more nuanced understanding of the world expected by 2030.

When questioned about self-aware AI systems, Hassabis acknowledged the theoretical possibility of such technology but clarified that current systems do not possess self-awareness. He emphasized that while the advancement of AI might lead to some forms of self-awareness, it is not a primary objective for DeepMind.

He elaborated that consciousness would need to be understood within the context of the different substrates—biological versus artificial—that underpin humans and machines.