Demis Hassabis, co-founder and CEO of Google Deepmind, will speak on February 26, 2024 at Mobile World Congress, the largest annual rally in the telecom industry in Barcelona, Spain.
Pau Barrena | AFP | Getty Images
London – Artificial intelligence is still in some way to suit humans at any task, but it’s only a matter of time before that becomes a reality. Google Deepmind.
During a briefing at Deep Mind’s London office on Monday, Demis Hassavis said he believes artificial general information (AGI) (smart or smarter than humans) will begin to appear in the next five or ten years.
“I think today’s systems are very passive, but there’s still a lot we can’t do. But over the next five to ten years, many of these abilities will start to come to the forefront and move towards what’s called artificial general information,” Hassavis said.
Hassavis defined AGI as “a system that can demonstrate all the complex capabilities that humans can do.”
“We’re not there yet. These systems are very impressive with certain things. But there are other things they can’t do yet. There’s quite a bit of research work before that,” Hassavis said.
Hassavis doesn’t just suggest that AGIs will take some time to appear. Last year, CEO of China’s leading technology company Baidu Robin Lee said he saw AGI as “more than a decade away” and pushed back the predictions of excitability from some of his peers about what this breakthrough is happening in a much shorter time frame.
Still time to go
Hassabis’ predictions push the timeline back to some degree compared to what his industry peers have been sketching.
Dario Amodei, CEO of AI startup humanity, told CNBC in January at the World Economic Forum in Davos, Switzerland, that he saw it as “better than almost all humans” that will appear in the “next two or three years.”
Other tech leaders have seen AGI arrive even earlier. Cisco Chief Product Officer Jeet Patel believes that examples of AGI could emerge soon this year. AI “has three major phases,” Patel told CNBC in an interview at the Mobile World Congress event in Barcelona earlier this month.
“There’s the basic AI we’re experiencing now. Then there’s artificial general information where cognitive abilities come across human abilities. And there’s what they call super intelligence,” Patel said.
“I think we’ll see meaningful evidence that AGI is being played in 2025. We haven’t talked about a distance in years,” he added. “I think Super Intelligence will be like in a few years at best.”
Artificial superintelligence, or ASI, is expected to arrive after the AGI and outweigh human intelligence. But when such a breakthrough occurs, “No one really knows,” Hassavis said Monday.
Last year, Tesla CEO Elon Musk predicted that AGI would likely be available by 2026, but Openai CEO Sam Altman said such systems could be developed in a “rationally near future.”
What do you need to reach AGI?
Hassavis said the main challenge regarding achieving artificial general information is to bring today’s AI systems to the point of understanding context from the real world.
It is possible to develop systems that can break down problems in the gaming realm, such as complex strategy board games, and autonomously complete tasks, but bring such technology into the real world is even more difficult.
“The problem is, how quickly we can generalize planning ideas and the actions of the kind of agents, planning and reasoning, and in addition to something like a world model, a model that allows us to understand the world around us, we can generalize it to working in the real world,” Hassabis said. ”
“And I think we’ve made good progress in the world model over the last few years,” he added. “So the question is, what is the best way to combine it with these planning algorithms?”
Hassabis and Thomas Kurian, CEOs of Google’s Cloud Computing Division, said the so-called “multi-agent” AI systems are a technological advancement that has gained a lot of traction behind the scenes.
Hassavis said there is a lot of work going on to reach this stage. One example he mentioned is Deepmind’s job, which allows AI agents to find a way to play the popular strategy game “Starcraft.”
“We’ve done a lot of work in the past when you have an agent society or a league of agents.
“When we think about agent-to-agent communication, that’s what we do too to enable agents to express ourselves… What are your skills? What tools do you use?” Kurian said.
“All of these are elements that need to be able to ask questions to agents, and once they have that interface, other agents can communicate with it,” he added.
