AI Summit: Systems can win Math Olympiad gold but still make mistakes at elementary maths, says Google DeepMind’s Demis Hassabis

Google DeepMind CEO Demis Hassabis in conversation with B Ravindran, IIT Madras

• Google DeepMind CEO says AGI is not yet achieved
• Current AI lacks consistency and continual learning
• AI still struggles with creativity and long-term planning

Did our AI summary help?

Google DeepMind CEO Demis Hassabis has said “we are still not there” as he talked about the path ahead for artificial general intelligence (AGI), systems that can exhibit all the cognitive capabilities of a human.

"Today's systems are very impressive, but they still have many flaws and things they can't do, including consistency across different tasks that you would want from a general system," Hassabis said on February 18 in a fireside chat at the India AI Impact Summit in New Delhi. The session was moderated by B Ravindran, head of the department of data science and AI at IIT- Madras.

The 49-year-old Nobel laurate identified consistency as one of the biggest issues. "They are like jagged intelligence. They are very good at certain things, but they are very poor at other things, including sometimes the same thing," Hassabis said.

For instance, today's models can get gold medals in the International Math Olympiad by solving really hard problems yet can make mistakes in elementary maths if a question is posed in a certain way.

According to Hassabis, the missing link for true general intelligence is continual learning.

"Today, we train the systems, but then they are frozen and put out into the world," he said "What you would like is for those systems to continually learn online from experience and the context they're in, maybe get personalised to the situation and the tasks that you have for them. Today's systems don't do that."

Hassabis also pointed out that current models struggle with long-term, coherent planning. "They can plan over the short term but over the longer term, like the way that we can plan over years, they don't really have that capability at the moment," he said.

Another key test is creativity, he said. "AI is very useful as a scientific tool assistant, or for solving specialised areas like AlphaFold does with protein structures but what separates great scientists from good scientists is their creativity and their taste for what's the right question, what's the right hypothesis?" he said.

"It's much harder to come up with the right question and the right hypothesis than it is to solve a conjecture. I would call that the highest level of creativity and so far, today's systems don't have that capability."

Hassabis, however, believes that these will be solved in time. "We're going to enter a new golden era for scientific discovery in the next 10 years," he said.