这家名为 Advanced Machine Intelligence Labs （AMI Labs）的新公司，计划于明年一月正式亮相，目标估值 30亿欧元 （约247亿人 民币）。 克雷西 发自 凹非寺 量子位 | 公众号 QbitAI LeCun在Meta的Last Day还没来，新公司又被曝出更多细节。 前脚LeCun本人在播客当中宣布了新公司名称，现在融资和估值目标就被《金融时报》曝光了。 AMI Labs的研究方向，就是LeCun一直主推的"世界模型"，而且将走开源路线，老东家Meta也将与其保持合作。 另外，曝料也透露了AMI Labs的 CEO人选并非LeCun本人 ，而是他的一位老部下。 LeCun不当CEO 新公司AMI Labs定于2026年1月在巴黎正式启动，在Meta逐渐转向封闭生态的背景下，LeCun选择了他在学术界一贯坚持的开源路线。 而且在技术层面，AMI Labs选择了比主流的LLM更具挑战性的道路—— 死磕"世界模型" 。 因为在LeCun看来，基于自回归机制的LLM存在根本性的逻辑缺陷，它们只是在统计概率上预测下一个字符，并不真正理解物理世界的运行规 律。 为此，新公司将通过 ...

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 to achieve true AI capabilities [1][2]. 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 centered on language capabilities [2]. - He advocates for the development of "world models" that can understand and predict the world, contrasting with mainstream models that focus on generating text or images [2][8]. - LeCun's new company, AMI, aims to pursue this alternative technical route, emphasizing cognitive and perceptual fundamentals rather than merely scaling existing models [2][7]. Group 2: Research and Open Science - LeCun stresses the importance of open research, arguing that true research must be publicly shared and scrutinized to avoid the pitfalls of insular corporate environments [5][6]. - He believes that allowing researchers to publish their work fosters better research quality and motivation, which is often overlooked in many industrial labs [6]. Group 3: World Models and Learning - The concept of world models involves creating abstract representations of the world to predict outcomes, rather than relying on pixel-level predictions, which are ineffective in high-dimensional data [8][10]. - LeCun emphasizes that effective learning requires filtering out unpredictable details and focusing on relevant aspects of reality, which is crucial for developing intelligent systems [10][22]. Group 4: Data and Training - LeCun highlights the vast difference in data requirements between language models and video data, noting that video data is richer and more valuable for learning due to its structural redundancy [18][19]. - He argues that relying solely on text data will never lead to human-level intelligence, as it lacks the necessary complexity and richness found in real-world data [19][25]. Group 5: Future of AI and AGI - LeCun expresses skepticism about the concept of "general intelligence," suggesting it is a flawed notion and that true progress will be gradual rather than sudden [30][32]. - He predicts that achieving "dog-level intelligence" will be the most challenging part of AI development, with significant advancements expected in the next 5 to 10 years if no unforeseen obstacles arise [32][34]. Group 6: Industry Trends and Company Direction - LeCun's departure from Meta and the establishment of AMI reflect a desire to pursue a different technological path amid a trend of companies focusing on large language models [1][48]. - He notes that the competitive environment in Silicon Valley often leads to a monoculture where companies pursue similar technological routes, which can stifle innovation [48].