Hassabis: Math Olympiad wins don’t prove AI’s general intelligence
Hassabis: Math Olympiad wins don’t prove AI’s general intelligence
AI leaders say solving complex math isn’t the benchmark for true intelligence; timelines for AGI loom as experts call for continuous learning and global cooperation.
London — Demis Hassabis, the head of Google’s AI efforts and DeepMind, has framed a measured view of progress in artificial intelligence. He argues that AI systems capable of solving advanced math problems, like those seen in international Olympiads, do not demonstrate true general intelligence. Such feats illustrate narrow strengths, but these systems can still make mistakes on basic math, and they’re essentially frozen after training, lacking the ongoing adaptability needed to handle new tasks.
Hassabis cautions that relying on isolated achievements risks overestimating AI capabilities. Real progress, he says, depends on developing systems that learn continuously and can plan for the future across diverse tasks, rather than optimizing a single objective in a single domain.
In a broader industry reflection, there is acknowledgment that the pace of AI advancement could usher in a significant shift. Experts describe the current moment as a threshold, with the potential for artificial general intelligence to emerge within the next five to eight years. That prospect has spurred calls for thoughtful governance, safety protocols, and international cooperation to manage societal impacts as AI capabilities expand.
The conversation underscores a balance between optimism about transformative possibilities and caution about managing risks. As researchers push for AI that learns over time and collaborates across domains, policymakers, businesses, and researchers are urged to prepare for rapid change while keeping safety and ethics at the forefront.
