AGI是否需要世界模型？顶级AI专家圆桌论道，清华求真书院主办

Core Insights - The forum highlighted the current limitations of AI technology, particularly the reliance on correlation rather than causation in modeling [3][4] - Experts discussed the potential need for a shift from token prediction models to world models to better handle complex cognitive tasks [6][14] - The debate on AI's originality emphasized that significant scientific breakthroughs require more than just data accumulation, as AI lacks the ability to make transformative hypotheses [7][9][10] Group 1: AI Limitations and Causality - Experts pointed out that the lack of causal modeling capabilities restricts AI's effectiveness in natural sciences and mathematical modeling [3][4] - The discussion included the emergence of causal understanding in large models, particularly in logical reasoning and mathematical problem-solving [5] Group 2: Token Prediction and World Models - The forum questioned whether the next token prediction paradigm is sufficient for achieving general intelligence, suggesting a need for new modeling approaches [6] - Optimistic views were presented regarding the ability of language to describe complex systems, which could allow AI to learn through token prediction [6][14] Group 3: Originality and Scientific Progress - The consensus was that AI cannot achieve true originality, as significant scientific advancements stem from hypothesizing about unknown phenomena rather than mimicking known structures [9][10] - AI's potential lies in its ability to cover known boundaries and reconstruct combinations, which could significantly impact scientific research [12][13] Group 4: Computational Challenges and Innovations - The exponential growth in computational power required for large models was highlighted, with training costs reaching approximately $10 billion and needing 200,000 GPU cards [14] - Innovations in optical computing and algorithms aimed at low-precision models were discussed as potential solutions to current computational challenges [16][17] Group 5: Future Paradigms in AI - A vision for a distributed interactive learning system involving 1 million robots was proposed, which could lead to a new evolutionary path for AI beyond centralized training paradigms [18][20] - The forum concluded with a call for breakthroughs in both theoretical and system aspects of AI to address its fundamental issues [21]