Google DeepMind CEO Demis Hassabis has described today's most advanced AI systems as "jagged intelligences" - models that dazzle in one domain yet stumble on basic problems when approached from a different angle. The flaw, he said, is a dealbreaker on the road to artificial general intelligence.
"They're really amazing at certain things when you pose the question in a certain way, but if you pose a question in a slightly different way they can actually still fail at quite elementary things," Hassabis told Harry Stebbings on the 20VC podcast, "Hvylya" reports. "A general intelligence shouldn't be that sort of jagged."
Beyond inconsistency, Hassabis pointed to several missing capabilities that separate today's models from true AGI. Continuous learning - the ability to absorb new information after training is complete - remains unsolved. "People haven't quite figured out yet how to integrate new learning into the existing systems that you spent months training," he said. The human brain, he noted, does this "very elegantly" through mechanisms like sleep reinforcement and memory consolidation.
Long-term planning is another gap. Current systems struggle to reason across extended time horizons - years rather than minutes - something the human mind handles routinely. Hassabis also flagged the need for better memory architectures, calling today's long context windows "a bit brute force."
Despite these shortcomings, Hassabis said DeepMind is ahead of schedule in most areas. Interactive world models like Genie, he said, would have amazed him if shown five or 10 years ago. He maintained there is "a very good chance" of reaching AGI within five years - a timeline consistent with the 20-year prediction he and co-founder Shane Legg made when they launched DeepMind in 2010.
Also read: why OpenAI scored an F on safety after dissolving its alignment teams.