DeepMind 執行長警告 AI 風險需全球合作

• Jagged Intelligence Problem: AI models can win gold medals at the International Mathematical Olympiad yet fail on elementary math questions, indicating uneven reasoning capabilities across domains[2] • World Models Development: DeepMind's Genie 3 and similar systems are learning physics intuition from video data, understanding phenomena like liquid flow and shadow casting—essential for AGI systems to plan and reason in physical environments[1] • Capability Gaps: Current systems lack true continual learning (models are 'frozen' after deployment), long-term memory, and genuine creativity required for scientific breakthroughs[1][2][3] • Foundation Model Limitations: While powerful for specialized problem-solving and scientific assistance, foundation models lack the creativity and judgment that distinguish exceptional scientists[3] • AlphaFold 2 Achievement: Hassabis's Nobel Prize-winning system (2024) can predict 3D protein structures for 200 million proteins, demonstrating AI's scientific potential[6]

The 5-8 year AGI timeline creates urgency for establishing international AI governance frameworks before systems achieve human-level general intelligence. India's positive stance on AI positions it as a potential global superpower in scientific innovation, while biosecurity and cybersecurity risks demand immediate defensive capability development. The emphasis on 'world models' suggests future AI systems will have enhanced physical reasoning for robotics and autonomous systems. Hassabis's focus on fixing jagged intelligence indicates the next phase of AI development will prioritize consistency and reliability over raw capability scaling, potentially reshaping how companies approach model training and deployment strategies.