In a dimly lit auditorium at Google‘s annual I/O developer conference, Demis Hassabis took the stage with an ambitious vision — a glimpse into the future of artificial intelligence. Hassabis, the founder of DeepMind and leader of Google’s AI efforts, has been working towards this moment for years. But four or five years ago, a crystallizing realization struck him: the world was on the cusp of having a universal AI assistant, “multimodal, with you all the time.”

“It’s that helper that’s just useful. You get used to it being there whenever you need it,” said Hassabis, describing his vision akin to the Star Trek Communicator or the sentient AI from the movie Her. And at I/O, he unveiled an early version of what could become that ubiquitous AI companion — a project dubbed Astra.

In a stunning demo video that Hassabis insists was not doctored, an Astra user in Google’s London office seamlessly conversed with the AI system. With simple voice commands, Astra identified components of a speaker, located the user’s missing glasses, reviewed code, and handled other tasks with striking real-time proficiency. The demo showcased Astra’s potential as a multimodal AI assistant that can perceive the world, understand context, and assist with nearly any task.

“Our history in agents is longer than our generalized model work,” Hassabis explained, pointing to the groundbreaking AlphaGo system from nearly a decade ago that mastered the ancient game of Go. In his vision, some AI agents will be ultra-simple tools, while others could evolve into true collaborators and companions. “I think it may even be down to personal preference at some point, and understanding your context.”

Hassabis believes Astra represents a significant step towards realizing the dream of a real-time AI assistant. While the underlying language model, Gemini 1.5 Pro, provides the technological backbone, the true challenge was optimizing the system for speed and latency. “Without that, the usability isn’t quite there,” Hassabis said. For six months, the DeepMind team has been laser-focused on accelerating Astra’s performance to deliver a seamless user experience.

Much of Google’s AI announcements at I/O centered around making the company’s latest large language model, Gemini, more accessible and conversational. Gemini Live, for instance, is a voice-only assistant that enables back-and-forth dialogues, allowing users to interrupt or refer back to earlier parts of the conversation. A new Google Lens feature lets users initiate web searches by narrating a video. These capabilities are made possible by Gemini’s large context window, which Hassabis says is crucial for natural interactions with AI assistants.

Interestingly, Google’s rival OpenAI seems to share a similar vision, demoing a product akin to Gemini Live shortly after Hassabis’s presentation. The two tech giants are increasingly vying for the same AI territory, each offering their own take on how these assistants might change lives and be used in the future.

But precisely how these AI agents will function and integrate into daily life remains uncertain, even for Hassabis. One area of focus for Google is using Gemini to plan trips, with a new tool that allows users to collaborate with the AI on building vacation itineraries. Hassabis envisions phones and augmented reality glasses as key devices for these agents but acknowledges “there is probably room for some exciting form factors.”

Astra is still an early prototype, representing just one potential interface for interacting with Gemini. The DeepMind team continues researching how to optimally combine multimodal models and strike the right balance between massive general models and more specialized ones.

While the current era of AI development obsesses over model sizes and benchmarks, Hassabis believes we’ll soon be asking different questions — about what these assistants can achieve, how they accomplish tasks, and how they can truly enrich our lives. The technology may be imperfect, but it’s advancing rapidly, ushering in a new age of AI that could redefine our relationship with technology.