Demis Hassabis, CEO of Google DeepMind, says that reaching artificial general intelligence or AGI—a fuzzy term typically used to describe machines with human-like cleverness—will mean honing some of the nascent abilities found in Google’s flagship Gemini models.
Google announced a slew of AI upgrades and new products at its annual I/O event today in Mountain View, California. The search giant revealed upgraded versions of Gemini Flash and Gemini Pro, Google’s fastest and most capable models, respectively. Hassabis said that Gemini Pro outscores other models on LMArena, a widely used benchmark for measuring the abilities of AI models.
Hassabis showed off some experimental AI offerings that reflect a vision for artificial intelligence that goes far beyond the chat window. “The way we’ve ended up working with today’s chatbots is, I think, a transitory period,” Hassabis told WIRED ahead of today’s event.
Hassabis says Gemini’s nascent reasoning, agentic, and world-modeling capabilities could enable much more capable and proactive personal assistants, truly useful humanoid robots, and eventually AI that is as smart as any person.
At I/O, Google revealed Deep Think, a more advanced kind of simulated reasoning for the Pro model. The latest AI models can break down problems and deliberate over them in a way that more closely resembles human reasoning than the instinctive output of standard large language models. Deep Think uses more compute time and several undisclosed innovations to improve upon this trick, says Tulsee Doshi, product lead for the Gemini models.
Google today unveiled new products that rely on Gemini’s ability to reason and take action. This includes Mariner, an agent for the Chrome browser that can go off and do chores like shopping when given a command. Mariner will be offered as a “research preview” through a new subscription plan called Google AI Ultra costing a hefty $249.99 per month.
Google also showed off a more capable version of Google’s experimental assistant Astra, which can see and hear the world through a smartphone or a pair of smart glasses.
As well as converse about the world around it, Astra can now operate a smartphone when needed, for example using apps or searching the web to find useful information. Google showed a scene in which a user had Atra help look for parts needed for bike repairs.
Doshi adds that Gemini is being trained to better understand how to preempt a user’s needs, starting with firing off a web search when this might be useful. Future assistants will need to be proactive without being annoying, both Doshi and Hassabis say.
Astra’s abilities depend on Gemini modeling the physical world to understand how it works, something Hassabis says is crucial to biological intelligence. AI will need to hone its reasoning, agency, and inventiveness, too, he says. “There are missing capabilities.”
Well before AGI arrives, AI promises to upend the way people search the web, something that may affect Google’s core business profoundly.
The company announced new efforts to adapt search to the era of AI at I/O (see WIRED’s I/O liveblog for everything announced today). Google will roll out an AI-powered version of search called AI Mode to everyone in the US and will introduce an AI-powered shopping tool that lets users upload a photo to see how an item of clothing would look on them. The company will also make AI Overviews, a service that summarizes results for Google users, available in more countries and languages.
Shifting Timelines
Some AI researchers and pundits argue that AGI may be just a few years away—or even here already depending on how you define the term. Hassabis says it may take five to 10 years for machines to master everything a human can do. “That’s still quite imminent in the grand scheme of things,” Hassabis says. “But it’s not tomorrow or next year.”
Hassabis says reasoning, agency, and world modeling should not only enable assistants like Astra but also give humanoid robots the brains they need to operate reliably in the messy real world.
DeepMind is currently collaborating with Apptroniks, one humanoid maker. A number of other companies, including big players like Tesla and startups such as Agility, Figure AI, and 1X are also building humanoids and touting their usefulness for factory and warehouse work. The ways these robots can be used is, however, very limited because they lack general intelligence.
“What is missing from robotics is not so much the robot itself, but its understanding of its physical context,” Hassabis says, adding that this is especially true for a home robot that would need to operate in complex and unfamiliar environments. In March, Google introduced Gemini Robotics, a version of its model capable of operating some robots.
Hassabis says that AI must become more inventive, too, if it is to imitate human intelligence faithfully. “Could [today’s models] invent general relativity with the knowledge that Einstein had in 1900? Clearly not,” he says.
Google is currently exploring ways to coax greater inventiveness out of AI models. The company recently unveiled AlphaEvolve, a coding agent capable of coming up with new algorithms for longstanding problems.
Hassabis says it may be possible to expand this creativity to areas beyond math and coding by having AI play games inside realistic 3D worlds.
This would represent something of a return to DeepMind’s roots, since the company made its name developing AI programs capable of playing video and board games. “You won’t be surprised to learn that I’m keen on games again as a testing ground for that,” Hassabis says.
Hassabis says AI may learn the same way that the board-game programs AlphaGo and AlphaZero learned to play chess and Go, although this will involve more ambitious world modelling. “You want a world model instead of a game model,” he says. “We think that’s critical for AGI to really understand the world.”
