DeepMind chief outlines limits of current AGI systems

Artificial general intelligence remains a future ambition rather than a present reality, according to Google DeepMind chief executive Demis Hassabis. Speaking at an AI summit in New Delhi, he said current systems still fall short of matching human-level intelligence in several vital areas.

Hassabis identified three key limitations. Existing AI models lack continual learning, meaning they cannot update their knowledge dynamically once deployed. Instead, they rely on static training completed before release, preventing them from adapting to new contexts or personalising responses over time.

Long-term planning is another weakness. While advanced models can handle short-term reasoning tasks, they struggle to plan strategically over extended periods, as humans do.

Consistency also remains an issue, as systems may perform exceptionally well in complex domains but make unexpected errors in simpler tasks.

Despite these shortcomings, Hassabis has previously suggested that genuine AGI could emerge within the next five to ten years. For now, however, he maintains that present systems have not yet reached that threshold.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!