In the latest interview, Demis Hassabis, CEO of Google DeepMind, said that artificial general intelligence (AGI) is expected to usher in the first form in the next five to ten years. However, he also pointed out that there are still many important challenges in technology.
Hassabis was interviewed by the media at DeepMind's London office. He stressed that although today's AI systems perform well on specific tasks, they still lack many basic capabilities. He defines AGI as “a system that can demonstrate all the complex capabilities of human beings.” In his opinion, the key to implementing AGI is to make AI systems truly understand the physical world.
Hassabis mentioned that despite the progress made by researchers in creating capabilities for autonomous planning and problem solving, migrating these capabilities into real-life scenarios remains a difficult challenge. “The key is how quickly we can summarize planning ideas and proxy behaviors and apply them to the real world, while also building models of the world around us,” he said.
When talking about world models, Hassabis believes that although researchers have made some progress in this field recently, finding the best way to effectively combine world models with planning algorithms is still a key problem. Similarly cautious attitude to Hassabis is Yann LeCun, the chief AI researcher of Meta, who is also actively exploring his own development ideas.
Hassabis' latest view is consistent with what he said in August 2024, when he believed that current AI capabilities are often overestimated while the technology's long-term potential is underestimated. According to median expectations of Metaculus forecast platform users, AGI will be available around 2030.
At present, there are more and more doubts about the way AGI is coming in the AI industry. According to a new survey, most AI researchers believe that large-scale language models (LLM) alone cannot achieve AGI. OpenAI recently changed its previous view, believing that the bursting capabilities of AI models will not directly lead to a rapid breakthrough in AGI, but rather regards the development of AGI as a gradual process of continuous evolution. Microsoft CEO Satya Nadella questioned AGI's predictions, calling this view a "meaningless benchmark hacker" and hoped that AI development would focus on providing real economic benefits.
Key points:
AGI is expected to be achieved in the next five to ten years, but the technical challenges remain.
Current AI systems perform well on specific tasks, but lack the complex capabilities of humans.
The industry's view on how AGI is implemented is gradually becoming cautious, believing that large-scale language models cannot be relied on.