A conversation with Demis Hassabis: Exploring the nature of intelligence and the future of AI

In a recent conversation with Demis Hassabis, CEO of DeepMind, this blog summarises the some of the most fascinating aspects of intelligence and the future of artificial intelligence. Hassabis, with his unique blend of neuroscience and AI expertise, provided compelling insights into the fundamental questions about human and machine intelligence.

Is Intelligence One Skill or Thousands of Subskills?

“Intelligence is so broad and what we use it for is so generally applicable,” Hassabis shared. This broad nature suggests that there must be high-level common themes in how the human brain processes information. While the brain certainly has specialized areas for different functions, Hassabis believes in underlying principles that connect these specialized abilities, much like a master plan or blueprint that guides us across various situations.

LLMs and Their Surprising Specializations

Discussing large language models (LLMs), Hassabis highlighted how feeding these models domain-specific data often results in asymmetric improvements — models get better in particular areas while also improving in general reasoning. “For example, improving models at coding can enhance their general reasoning abilities,” Hassabis said…