"通往超级智能的那条路——无非是不断训练大语言模型、喂更多合成数据、雇上几千人做后训练、再在强化学习上搞点新花样——在我看来完全是胡 扯，这条路根本行不通。" 近日，在一档名为《The Information Bottleneck》的访谈栏目中，主持人 Ravid Shwartz-Ziv 和 Allen Roush 与图灵奖得主、前 Meta 首席 AI 科学家 Yann LeCun 展开了一场近两小时的高质量对话，在访谈中，LeCun 解释了为什么会在 65 岁这个别人已经退休的年纪他还在创业，此外，他也对当前硅谷主流 的人工智能发展路径给出了罕见而尖锐的评价。 结束在 Meta 长达 12 年的职业生涯后，LeCun 正将个人学术声誉与职业"遗产"押注在一套截然不同的 AI 愿景之上。他直言，业界对大语言模型规模化的 执念，正在把人工智能引向一条看似高速、实则封闭的死胡同。 在 LeCun 看来，真正制约 AI 进步的关键，并不是如何更快地逼近"人类级智能"，而是如何跨越一个常被低估却极其困难的门槛——让机器具备"狗的智 能水平"。这一判断挑战了当前以语言能力和知识覆盖面为中心的评估体系。在他看来，现实世 ...

周末的上海张江科学会堂成了"遛娃圣地"：2025国际具身智能技能大赛在此举办，"硅基生命"被推到公 众眼前，完成爬坡、搬运、叠衣、陪护等任务。家长带娃前来想领略"黑科技"，却见证了颇多笑料：有 的机器人在执行任务时突然"一脸懵"愣在原地，有的在绕过障碍时把自己扳倒；有一台机器人刚出场就 重重摔伤，脑壳碎了一地……人们纷纷揶揄道，原来热度颇高的具身智能赛道，也不过如此。 在围观者的戏谑和唱衰中，有人发出"灵魂质疑"：以比赛的形式将行业的痛点和局限暴露出来，真的有 必要吗？笔者认为，现在"应摔尽摔"，未来才能大步向前。这些让人捧腹的机器人"翻车"瞬间，恰恰是 大赛设计理念的核心所在——将机器人从理想的实验室环境推向不可预测的真实世界，从而暴露其技术 瓶颈，为未来的发展指明方向。 近年来人工智能发展呈现出极具反差感的两面：网页端的大模型已能通过律师考试，而现实中的机器人 却只能像学龄前孩童一样蹒跚学步——这就是著名的"莫拉维克悖论"。人们一边感叹通用人工智能已经 来了，一边又诧异于机器人叠个衬衫"难于登天"。 如何才能让机器人具备真正的干活能力？首先要撕掉实验室"滤镜"，让机器人在真实碰撞中暴露弱点。 本次大赛上， ...

周末的上海张江科学会堂成了"遛娃圣地"：2025国际具身智能技能大赛在此举办，"硅基生命"被推到公众眼 前，完成爬坡、搬运、叠衣、陪护等任务。家长带娃前来想领略"黑科技"，却见证了颇多笑料：有的机器人在 执行任务时突然"一脸懵"愣在原地，有的在绕过障碍时把自己扳倒；有一台机器人刚出场就重重摔伤，脑壳碎 了一地……人们纷纷揶揄道，原来热度颇高的具身智能赛道，也不过如此。 在围观者的戏谑和唱衰中，有人发出"灵魂质疑"：以比赛的形式将行业的痛点和局限暴露出来，真的有必要 吗？笔者认为，现在"应摔尽摔"，未来才能大步向前。这些让人捧腹的机器人"翻车"瞬间，恰恰是大赛设计理 念的核心所在——将机器人从理想的实验室环境推向不可预测的真实世界，从而暴露其技术瓶颈，为未来的发 展指明方向。 近年来人工智能发展呈现出极具反差感的两面：网页端的大模型已能通过律师考试，而现实中的机器人却只能 像学龄前孩童一样蹒跚学步——这就是著名的"莫拉维克悖论"。人们一边感叹通用人工智能已经来了，一边又 诧异于机器人叠个衬衫"难于登天"。 证券时报各平台所有原创内容，未经书面授权，任何单位及个人不得转载。我社保留追 究相关 行 为主体 法律责任 ...

Core Viewpoint - Yann LeCun, a prominent AI scientist at Meta, is leaving the company to start a new venture focused on advanced machine intelligence, diverging from Meta's current investment in large language models (LLMs) [1][36][38]. Group 1: Departure and New Venture - Yann LeCun announced his departure from Meta after 12 years, stating that the company will be a partner in his new startup, although Meta will not be an investor [1][36]. - LeCun's new company will focus on teaching AI to understand the physical world rather than developing LLMs like ChatGPT [3][36]. Group 2: Critique of Large Language Models - LeCun has been a vocal critic of LLMs, arguing that they have reached their limits and lack true understanding of the physical world, memory, and multi-step reasoning capabilities [6][8]. - He believes that LLMs are merely token generators and do not possess the reasoning abilities necessary for true intelligence [6][20]. Group 3: The Concept of World Models - LeCun advocates for the development of "world models," which he believes are essential for achieving true machine intelligence, as they allow for understanding and interaction with the physical world [12][22]. - He emphasizes that human-like intelligence requires more than just language processing; it necessitates the ability to interact with and learn from the environment [35][36]. Group 4: Industry Implications - The AI industry is heavily focused on LLMs, which LeCun describes as a "black hole" that absorbs resources and attention, hindering progress in other areas of AI research [8][40]. - LeCun's departure and criticism of LLMs may signal a shift in the AI landscape, as he suggests that the next major breakthroughs will come from alternative approaches like world models [12][40].

Core Insights - Generalist, a company founded by Pete Florence, has released a new embodied foundation model called GEN-0, which can scale predictably with the growth of physical interaction data [1][4] - The company aims to create universal robots, focusing initially on the dexterity of robots [4][5] Company Overview - Generalist was co-founded by Pete Florence, Andrew Barry, and Andy Zeng, with a team that includes experts from OpenAI, Waymo, and Boston Dynamics [4] - Early investors include Spark Capital, NVIDIA, and Bezos Expeditions, although the investment amounts remain undisclosed [3] Model Features - GEN-0 is based on high-fidelity raw physical interaction data and employs a multi-modal training approach [5] - A key feature of GEN-0 is "Harmonic Reasoning," allowing the model to think and act simultaneously, which is crucial for real-world applications [6][7] Scaling and Performance - The model exhibits a "phase transition" point in its intelligence capacity, indicating that larger models are necessary to absorb complex sensory-motor data [8][10] - Models with 1 billion parameters struggle to absorb diverse data, while those with 6 billion parameters show strong multi-task capabilities [10][11] - Models with over 7 billion parameters can internalize large-scale pre-training data and quickly adapt to downstream tasks [12] Scaling Law - GEN-0 demonstrates a clear Scaling Law, where increased pre-training data and computational resources lead to predictable improvements in downstream performance [15] - The company has developed a predictive formula to determine the optimal data allocation for specific tasks [15][16] Data Quality and Diversity - The training dataset for GEN-0 consists of 270,000 hours of real-world manipulation trajectories collected from diverse environments, significantly larger than existing datasets [16][18] - The quality and diversity of data are more critical than sheer volume, allowing for the creation of models with different characteristics [18] Industry Context - The field of embodied intelligence is still in its early stages, with various companies exploring foundational models [19] - Despite the presence of numerous top-tier companies, the technology landscape remains fragmented, and commercial applications are limited [19][20] Future Prospects - The advancements in Scaling Law and model capabilities suggest a promising future for the commercialization of embodied intelligence [20] - Chinese entrepreneurs have a competitive advantage in this field due to a mature hardware supply chain and rich data sources [21]

Core Insights - The core prediction is that by 2030, robots capable of autonomously managing entire households will emerge, driven by the "robot data flywheel" effect [1][11]. Group 1: Robot Development and Implementation - Robots are expected to be deployed faster than autonomous driving and large language models due to their ability to quickly obtain clear feedback from the physical world [2]. - The clear technological path involves an integrated model of "vision-language-action," allowing robots to understand tasks and plan actions autonomously [3]. - Real-world applications in small-scale settings are prioritized over large-scale simulations to leverage precise data feedback [4]. Group 2: Emerging Capabilities and Challenges - "Combination generalization" and "emergent abilities" will lead to significant advancements in robot technology, enabling robots to transition from specific tasks to general household capabilities [5]. - Current challenges in robot development include response speed, context memory length, and model scale, but these can be addressed by combining existing technologies [6]. - The rapid decrease in hardware costs has lowered the entry barrier for AI entrepreneurs, allowing small teams to quickly iterate and validate market needs [7]. Group 3: Future Vision and Timeline - The ultimate goal for robots is to execute long-term, high-level tasks autonomously, requiring advanced capabilities such as continuous learning and problem-solving [10]. - The "flywheel effect" will accelerate robot capabilities as they perform useful tasks and gather experience data [11]. - Predictions suggest that within one to two years, robots will start providing valuable services, with fully autonomous household management achievable in about five years [11]. Group 4: Comparison with Other Technologies - The development of robots may progress faster than large language models and autonomous driving due to the unique nature of their interaction with the physical world [12][13]. - Robots can learn from clear, direct human feedback in physical tasks, contrasting with the challenges faced by language models in extracting effective supervisory signals [12]. Group 5: Learning and Data Utilization - Robots benefit from embodied intelligence, allowing them to focus on relevant information while learning from vast amounts of video data [20][21]. - The ability to generalize and combine learned skills will be crucial for achieving general intelligence in robots [23][25]. Group 6: Systemic Challenges and Solutions - The "Moravec's Paradox" highlights the difficulty of replicating simple human tasks in robots, emphasizing the need for physical skill development over memory expansion [26][27]. - Future advancements will require addressing the trade-offs between reasoning speed, context length, and model scale [28][29]. Group 7: Hardware and Economic Factors - The cost of robotic hardware has significantly decreased, enabling broader deployment and data collection for machine learning [33]. - The economic impact of automation will enhance productivity across various sectors, necessitating careful planning for societal transitions [34]. - Geopolitical factors and supply chain dynamics will play a critical role in the advancement of robotics, emphasizing the need for a balanced ecosystem [35].

Core Insights - The article discusses the evolution and challenges of embodied intelligence, highlighting the distinction between "problem-solving" AI and "practical" AI, with the latter focusing on real-world interactions and learning through sensory experiences [1][3] - It emphasizes the need for embodied intelligence to overcome significant hurdles in understanding, associating, and interacting with the environment, which are essential for robots to function like humans in real-world scenarios [3][5] Group 1: Challenges in Embodied Intelligence - Embodied intelligence must adapt to unstructured real-world environments, requiring advanced computational capabilities to handle dynamic and unpredictable situations [5][6] - The development of higher cognitive strategies that integrate multiple sensory inputs is crucial for robots to understand and interact with their surroundings effectively [6][7] - Robots need to surpass traditional static data processing models to achieve a deeper understanding of dynamic changes and relationships in their environment [6][12] Group 2: Technological Components - The perception layer of embodied intelligence is vital for converting chaotic physical stimuli into understandable digital signals, relying on multimodal sensor fusion and dynamic environment modeling [8][10] - The cognitive layer processes raw data from the perception layer, employing hierarchical decision-making and world model construction to enable robots to learn from experiences [12][14] - The action layer ensures robots can execute tasks safely and effectively, utilizing bio-inspired drive technologies and human-robot collaboration safety designs [16][18] Group 3: Current Limitations and Future Directions - Current embodied intelligence models struggle with task completion rates in non-training scenarios, with a success rate of only 65% for tasks like object grasping [17] - Energy consumption and high costs remain significant barriers to the widespread adoption of humanoid robots, with typical models having a battery life of less than 2 hours and costs exceeding 500,000 yuan [18][19] - Research is focused on optimizing energy efficiency and reducing costs through new battery technologies and domestic production of core components [21][22] Group 4: Future Trends - The integration of multimodal large models is a key future direction, enabling robots to understand natural language commands and adapt quickly to new tasks with minimal samples [23][24] - Lightweight hardware innovations, such as bio-inspired muscle drive technologies, are expected to enhance performance while reducing costs [23][24] - The trend of virtual-physical collaborative evolution will allow robots to train in simulated environments, significantly improving their task execution capabilities in real-world settings [24][25]

Core Viewpoint - The article discusses the contrasting perspectives of AI pioneers Geoffrey Hinton and Hans Moravec on the future of artificial intelligence, likening AI to either a domesticated tiger or a human offspring, with implications for human civilization and evolution [1][3]. Group 1: Perspectives on AI Development - Hinton and Moravec, contemporaries in the AI field, represent different approaches: Hinton focuses on neural networks and learning capabilities, while Moravec emphasizes embodied intelligence and evolutionary processes [3][7]. - Moravec predicts that universal robots will surpass human intelligence between 2030 and 2040, as computational power continues to grow [4][5]. - The evolution of robots is expected to progress from basic learning to human-like reasoning, reflecting a gradual transformation of intelligence [5][6]. Group 2: Moravec's Paradox - Moravec's paradox highlights that human reasoning requires minimal computational resources, while perception and motor skills demand significant resources, challenging common intuitions about AI capabilities [9][12]. - The paradox suggests that the advanced perceptual and motor skills developed over millions of years of evolution are deeply embedded in human genetics, while abstract reasoning is a relatively recent development [8][11]. - This paradox serves as a reminder of the complexities in developing robots that can truly replicate human-like perception and action [13][14]. Group 3: Current State of Robotics - The article critiques the current state of humanoid robots, suggesting that many demonstrations are misleading and do not reflect true capabilities, as they often lack genuine environmental perception [14][15]. - Training robots to perform complex tasks is significantly more challenging than training them for simple, pre-programmed movements, emphasizing the need for advanced perception and interaction with the physical world [15][17]. - The distinction between "blind gymnasts" and robots capable of perception and action illustrates the current limitations in robotics research [15][16]. Group 4: Future Implications - The potential for AI to surpass human intelligence raises questions about the future relationship between humans and intelligent machines, with Moravec suggesting that robots may inherit human civilization [19][20]. - Hinton's views on AI's potential risks have evolved, indicating a belief that AI can be developed to be both intelligent and benevolent, though Moravec expresses skepticism about humanity's ability to control this evolution [18][19].

Group 1 - The article discusses the importance of intuitive and embodied intelligence, emphasizing that true understanding comes from experience rather than abstract reasoning [1][39][84] - It highlights the concept of "world models" in AI, which aim to enable machines to understand and interact with the physical world in a more human-like manner [23][76][84] - The text draws parallels between human cognitive processes and AI development, suggesting that both rely on a form of non-verbal, intuitive understanding [17][29][72] Group 2 - The article references the limitations of current AI systems in understanding the physical world compared to human capabilities, particularly in spatial reasoning and perception [18][22][25] - It discusses the evolution of intelligence, noting that human cognitive abilities have been shaped by millions of years of evolution, which AI is still trying to replicate [21][75] - The piece concludes with the notion that as AI develops its own "taste" through embodied experiences, it may reach a level of understanding that parallels human intuition [72][84][85]

Core Insights - The article discusses the evolution of AI models using games as a testing ground, highlighting the recent achievement of Google's AI model Gemini 2.5 Pro in independently completing the original Pokémon game, which has reignited interest in AI capabilities [4][30]. Group 1: AI Development and Gaming - AI has been tested through games for nearly a decade, with notable milestones including AlphaGo's victory over human players in Go and DeepMind's success in games like DOTA2 and StarCraft II [2][3]. - The use of games as a benchmark for AI intelligence remains prevalent, as demonstrated by Gemini's recent accomplishment, which was celebrated by Google's CEO and DeepMind's head [4][5]. Group 2: Challenges in AI Learning - The Moravec's paradox suggests that tasks perceived as easy for humans can be significantly more challenging for AI, which is exemplified by Gemini's achievement in Pokémon [6][7]. - The process of AI learning in games like Pokémon is complex, requiring the AI to develop its own understanding and strategies without predefined rules or guidance [16][17]. Group 3: Comparison of AI Models - Anthropic's Claude 3.7 struggled to progress in Pokémon, achieving only three badges after a year of iterations, while Gemini completed the game with approximately 106,000 actions, significantly fewer than Claude's 215,000 actions [11][30]. - The differences in performance between Claude and Gemini are attributed to their respective frameworks, with Gemini's agent harness providing better input processing and decision-making capabilities [34][35]. Group 4: Implications for AI Research - The ability of AI to navigate and complete games like Pokémon indicates its potential for independent learning and problem-solving in real-world scenarios [37][38]. - The choice of Pokémon as a training ground reflects the game's themes of growth, choice, and adventure, paralleling the journey of AI in understanding complex rules and environments [39][40].