The world is still some time away from solving intelligence or building artificial general intelligence (AGI), Google Deepmind Co-founder and CEO Demis Hassabis said on Wednesday. “We’re not there yet. But we are at a threshold moment where AGI is on the horizon in 5-8 years,” Hassabis said in a conversation with Balaraman Ravindran, professor and researcher at IIT Madras and head of Centre for Responsible AI, at the AI Impact Summit.
The Nobel laureate observed that his measure for solving intelligence is a system that can exhibit all the cognitive capabilities that humans can including in terms of creativity and long-term planning. However, while today’s systems are impressive, they have flaws in that they are not always consistent across different tasks which we would want from a general system.
Jagged Paradox
“When I look at the current systems and what is missing from them being a kind of a general intelligence, I would say things like continual learning, so learning after being trained and put out in the world,” Hassabis observed.
Today’s systems, he explained, are trained and “kind of frozen” and put out in the world. “But what we would like is for them to continually learn online from experience and from the context they are in, maybe personalised to the situation,” the Deepmind CEO said.
He pointed out that the systems also do not have the capability to do long-term coherent planning. “But I think one of the biggest issues is consistency. They are kind of like jagged, good at certain things but poor at others,” Hassabis said, adding that humans are not jagged in that way. According to him, today’s general foundational models, for instance, are very poor at chess which is not what we would expect.
AIso, while AI is very useful as a scientific tool or assistant but what separates great scientists from good scientists is creativity. “So far, today’s systems don’t have that capability. They can be the ultimate tools for science in the future. In the next 10 years we will enter a new era for scientific discovery,” Hassabis observed. The opportunities from AGI, the most transformative technology ever, will be incredible, especially in the fields of science and medicine and can revolutionise drug discovery.
Defensive Superiority
In his view the the top two risks relating to artificial intelligence (AI) in the near term are bio risks and cyber risks. “We need to make sure that the cyber defences are more powerful than the offences,” the Deepmind CEO said, calling for international collaboration. “Technology will impact everyone and it is hard to understand all the vulnerabilities,” he added.
Hassabis spoke of risks relating to bad actors whether humans or nation states and also to agentic systems which are more autonomous, calling for guardrails. He countered a view that the countries with resources were the ones which were enjoying the benefits of AI, saying that the leading foundational models are available, quite soon, to other countries as well. The foundational models, he said, are a critical part of AGI solutions and it makes sense to use these models.