Demis Hassabis, the visionary behind AlphaGo's groundbreaking success in 2016, is once again at the forefront of AI innovation with Google's latest offering: the Gemini AI model. According to Hassabis, Gemini represents a monumental leap forward not just for Google but for the entire field of artificial intelligence. This multimodal AI model, unveiled recently by Google, promises to chart an unexplored path in AI, potentially unlocking major breakthroughs.

Hassabis, drawing from his background as both a neuroscientist and a computer scientist, expresses a long-standing ambition to create a new generation of AI models inspired by human sensory experiences. In an exclusive interview with WIRED preceding the official announcement, he shares that Gemini is a significant step towards realizing this vision. Described as "multimodal," Gemini distinguishes itself by its ability to process information in various formats, encompassing text, audio, images, and video.

The initial iteration of Gemini is now accessible through Google's chatbot Bard, offering users a glimpse into the model's capabilities. Google has ambitious plans for Gemini, with an even more potent version, Gemini Ultra, slated for release next year. Google claims that Gemini Ultra surpasses GPT-4, the model powering ChatGPT, across several benchmark tests. Videos released by Google showcase Gemini's prowess, tackling tasks involving complex reasoning and seamlessly integrating information from disparate modalities.

Hassabis notes a departure from the conventional approach to multimodality, where separate modules are trained and then combined. He implies a distinction from OpenAI's technology, stating that such an approach is suitable for some tasks but lacks the depth required for complex reasoning in multimodal space.

The genesis of Gemini can be traced back to the end of 2022 when Google faced challenges to its AI leadership position from OpenAI's ChatGPT. In response, Google rapidly launched Bard, revamped its search engine, and introduced PaLM 2 as a competitor to ChatGPT. Hassabis, elevated to lead a consolidated AI division, unveiled plans for Gemini at Google's developer conference, I/O, in May, marking the culmination of seven months of accelerated development.

Beyond language proficiency, Gemini aims to address the challenge of grounding AI in physical reality. Hassabis envisions combining Gemini with robotics to enable physical interactions with the world, emphasizing the inclusion of touch and tactile feedback for true multimodality. Google had previously taken steps in this direction with AI models like Gato and projects like RT-2, showcasing the capacity to perform tasks from playing games to manipulating physical objects with robotic arms.

In the evolving landscape of AI, concerns about safety have taken center stage. Hassabis emphasizes Google's commitment to rigorous safety testing, extending to the most potent version, Gemini Ultra, slated for release in the coming year. The company is actively addressing potential issues related to bias, toxicity, and overall safety, reflecting an ongoing commitment to responsible AI development.

As the race for AI supremacy intensifies among tech giants, debates surrounding the risks associated with AI models have gained prominence. Hassabis, involved in a UK government initiative earlier in the year, highlights Google's proactive stance in addressing risks, emphasizing ongoing efforts to ensure the safety of AI technologies. As Google unveils the initial version of Gemini, the AI community anticipates a new era in artificial intelligence marked by multimodal capabilities and enhanced safety measures.