Alex Wang“没资格接替我”！Yann LeCun揭露Meta AI“内斗”真相，直言AGI是“彻头彻尾的胡扯”

Core Viewpoint - Yann LeCun criticizes the current AI development path focused on scaling large language models, arguing it leads to a dead end and emphasizes the need for a different approach centered on understanding and predicting the world through "world models" [2][3]. Group 1: AI Development Path - LeCun believes the key limitation in AI progress is not reaching "human-level intelligence" but rather achieving "dog-level intelligence," which challenges the current evaluation systems focused on language capabilities [3]. - He is establishing a new company, AMI, to pursue a technology route that builds models capable of understanding and predicting the world, moving away from the mainstream focus on generating outputs at the pixel or text level [3][9]. - The current industry trend prioritizes computational power, data, and parameter scale, while LeCun aims to redefine the technical path to general AI by focusing on cognitive and perceptual fundamentals [3][9]. Group 2: Research and Open Science - LeCun emphasizes the importance of open research, stating that true research requires public dissemination of results to ensure rigorous methodologies and reliable outcomes [7][8]. - He argues that without allowing researchers to publish their work, the quality of research diminishes, leading to a focus on short-term impacts rather than meaningful advancements [7][8]. Group 3: World Models and Planning - AMI aims to develop products based on world models and planning technologies, asserting that current large language model architectures are inadequate for creating reliable intelligent systems [9][10]. - LeCun highlights that world models differ from large language models, as they are designed to handle high-dimensional, continuous, and noisy data, which LLMs struggle with [10][11]. - The core idea of world models is to learn an abstract representation space that filters out unpredictable details, allowing for more accurate predictions [11][12]. Group 4: Data and Learning - LeCun discusses the vast amount of data required to train effective large language models, noting that a typical model pre-training scale is around 30 trillion tokens, equating to approximately 100 trillion bytes of data [20]. - In contrast, video data, which is richer and more structured than text, offers greater learning value, as it allows for self-supervised learning due to its inherent redundancy [21][28]. Group 5: Future of AI and General Intelligence - LeCun expresses skepticism about the concept of "general intelligence," arguing it is a flawed notion based on human intelligence, which is highly specialized [33][34]. - He predicts that significant advancements in world models and planning capabilities could occur within the next 5 to 10 years, potentially leading to systems that approach "dog-level intelligence" [35][36]. - The most challenging aspect of AI development is achieving "dog-level intelligence," after which many core elements will be in place for further advancements [37]. Group 6: Safety and Ethical Considerations - LeCun acknowledges the concerns surrounding AI safety, advocating for a design approach that incorporates safety constraints from the outset rather than relying on post-hoc adjustments [43]. - He argues that AI systems should be built with inherent safety features, ensuring they cannot cause harm while optimizing for their objectives [43][44].