Google DeepMind CEO Demis Hassabis said the company is focusing on rigorous testing and responsible development as AI moves toward greater intelligence.
These efforts are critical because the transition to artificial general intelligence, or AGI, carries significant technical and societal risks that require proactive management.
During public appearances in 2024, including Google I/O and the Athens Innovation Summit, Hassabis said AI could unlock breakthroughs in fields such as energy and medicine [1, 2]. He said that while the progress is rapid, the industry must address hard technical questions regarding how these systems actually learn and reason [3, 5].
Hassabis highlighted the limitations of current large language models, suggesting a need for "world models" to achieve deeper understanding [5]. He said that the development of truly intelligent systems requires more than just pattern recognition, and it requires the ability to solve complex, novel problems autonomously.
Recent milestones illustrate this trajectory. The AlphaProof Nexus system solved nine [1] open Erdős problems [1]. These specific mathematical challenges were 56 [1] years old [1].
Despite these achievements, Hassabis maintained a cautious perspective on the current state of the technology. While some reports suggest AGI is coming soon [3], he said that solving these long-standing math problems does not yet constitute the arrival of AGI [1]. He said that the path to general intelligence involves passing rigorous tests that current systems have not yet fully cleared [3].
Hassabis said the goal is to ensure that as these systems become more capable, they remain aligned with human interests and safety standards [4]. This approach involves a balance between pushing the boundaries of what AI can do and maintaining a strict framework for its deployment [2, 4].
“AI to unlock breakthroughs in fields such as energy and medicine”
The distinction Hassabis makes between high-level problem solving and AGI suggests that Google DeepMind views 'intelligence' as a multi-dimensional threshold rather than a single achievement. By emphasizing the need for world models over simple language prediction, the company is signaling a shift in research toward systems that understand physical or logical causality, which is a prerequisite for safe and reliable deployment in critical infrastructure.
