Yann LeCun, as a pioneer in the field of AI, has always focused on advancing the fundamental progress of AI rather than short-term commercial interests. Below is his in-depth reflection on current technological trends, industry landscape, and personal choices:

1. The Core Significance of World Models
Yann LeCun believes that the ultimate goal of AI is to create intelligent agents capable of understanding, predicting, and acting in the world. World Models are key to achieving this goal. He criticizes the limitations of current mainstream large language models (LLMs) and generative AI (such as text/image generation):

  • Misconception of Pixel/Text Generation: Generating pixels or text only reflects superficial statistical correlations and fails to grasp the underlying physical laws of the world.
  • Prediction Should Occur in Abstract Spaces: True intelligence requires temporal prediction in abstract representation spaces (rather than pixel/text spaces), for example, Wayve’s representation space prediction model in autonomous driving.
  • Opposition to “Large Models Are the Future”: He points out that large language models perform poorly in handling continuous, high-dimensional, and noisy data (such as video and physical perception), while Silicon Valley’s “compute + data” competition may hinder the birth of truly breakthrough technologies.

2. Criticism of the Silicon Valley Technological Path
Yann LeCun directly states that the AI ecosystem in Silicon Valley has serious issues:

  • Technological Homogenization: All major players (OpenAI, Meta, Google, etc.) are concentrated on the large language model赛道, squeezing innovation.
  • “Large Model Brainwashing” Culture: Many people blindly believe that expanding model scale and piling up data can achieve super intelligence, but he argues this neglects the real complexity of the physical world.
  • Risk of Path Dependence: If all resources are concentrated on a single technological path, the industry may miss disruptive breakthroughs, such as models based on physical dynamics or neuro-symbolic systems.

3. Strategic Choices in Founding a New Company
Yann LeCun chose to establish operations in Paris, New York, and other locations instead of Silicon Valley, with core reasons including:

  • Avoiding Suppression of Alternative Technological Paths: The “large model culture” in Silicon Valley may make it difficult for different technological paths (such as world models) to survive.
  • Promoting Open Basic Research: He firmly believes that open and accumulated basic research is the core driving force for long-term AI development, hence founding a global company AMI to attract engineers within Silicon Valley who share his views.
  • Mission-Driven Approach: Enhancing global intelligence is an “inherently correct thing,” and he aims to create more valuable tools for humanity through world models.

4. Evaluation of Other Companies

  • Meta’s FAIR Lab: Yann LeCun acknowledges FAIR’s shift toward application-oriented research but believes this weakens the openness of basic research, also one of the reasons for his departure.
  • Wayve: He recognizes their attempt to build abstract representation spaces in autonomous driving but points out their representation spaces still rely on reconstruction training rather than true dynamic understanding.
  • NVIDIA and Sandbox AQ: He supports their exploration of large quantitative models (handling continuous high-dimensional data), which aligns closely with his views.
  • Google’s Dreamer Series: He commends their exploration in the world model direction but regrets that Hafner has started a company, leading to research interruptions.

5. Personal Reflections and Scientific Historiography
Yann LeCun admits he once missed publishing ideas like the backpropagation algorithm, but he emphasizes that the emergence of scientific ideas is the result of a complex chain. The concept of world models is not new; it was already applied in control theory and aerospace engineering in the 1960s. The real challenge lies in transforming theory into working systems.

6. Vision for the Future of AI
Yann LeCun’s path declaration reveals two directions for the AI industry:

  • Short-Term Path: Silicon Valley’s “compute + data” competition, pursuing product implementation and commercial value.
  • Long-Term Path: Basic research based on cognition + perception, pursuing technological breakthroughs and the essence of intelligence.
    He believes that only diverse perspectives and open exploration can prevent the industry from falling into path dependence, and world models may become the key to achieving truly intelligent agents.

Conclusion
Yann LeCun’s commitment is not only a technological choice but also a profound understanding of the essence of AI. His actions remind us that the future of AI does not lie in scaling up but in whether it can truly understand and predict the world. Regardless of the final outcome, his mission-driven spirit and courage provide the industry with valuable diverse perspectives and showcase the初心 of scientists—to let technology serve a more intelligent future for humanity.

Translation

杨立昆（Yann LeCun）作为AI领域的先驱者，其观点和行动始终围绕着推动AI的本质进步，而非短期商业利益。以下是他对当前技术趋势、行业格局以及个人选择的深度思考：

1. 世界模型的核心地位
杨立昆认为，AI的终极目标是构建能够理解、预测和行动的智能体，而世界模型（World Models）是实现这一目标的关键。他批评当前主流的大语言模型（LLM）和生成式AI（如文本/图像生成）的局限性：

  • 像素/文本生成的误区：生成像素或文本只是表层统计相关性，无法理解世界的底层物理规律。
  • 预测应在抽象空间进行：真正的智能需要在抽象表征空间（而非像素/文本空间）中进行时间预测，例如Wayve在自动驾驶中构建的表征空间预测模型。
  • 反对“大模型即未来”：他指出，大语言模型在处理连续、高维、噪声数据（如视频、物理感知）时表现不佳，而硅谷的“算力+数据”竞赛可能阻碍真正突破性技术的诞生。

2. 对硅谷技术路线的批判
杨立昆直言，硅谷的AI生态存在严重问题：

  • 技术单一化：所有巨头（OpenAI、Meta、Google等）扎堆在大语言模型赛道，导致创新被挤压。
  • “大模型洗脑”文化：许多人盲目相信扩大模型规模、堆砌数据就能实现超级智能，但他认为这忽视了物理世界的真实复杂性。
  • 路径依赖风险：若所有资源集中于同一条技术路线，行业可能错失颠覆性突破，例如基于物理动力学的模型或神经符号系统。

3. 创办新公司的战略选择
杨立昆选择在巴黎、纽约等地布局而非硅谷，核心原因包括：

  • 避免技术路线被压制：硅谷的“大模型文化”可能让不同技术路线（如世界模型）难以生存。
  • 推动开放基础研究：他坚信基础研究的开放和沉淀是AI长期发展的核心动力，因此创办全球性公司AMI，吸引硅谷内部认同其观点的工程师。
  • 使命驱动：提升全球智能总量是“内在正确的事”，他希望通过世界模型为人类创造更有价值的工具。

4. 对其他公司的评价

  • Meta的FAIR实验室：杨立昆承认FAIR转向应用导向，但认为这削弱了基础研究的开放性，也是他离开的原因之一。
  • Wayve：认可其在自动驾驶中构建抽象表征空间的尝试，但指出其表征空间仍依赖重建训练，而非真正的动力学理解。
  • NVIDIA与Sandbox AQ：支持其探索大型定量模型（处理连续高维数据），与他的主张高度一致。
  • 谷歌的Dreamer系列：肯定其在世界模型方向的探索，但遗憾哈夫纳已创业，导致研究中断。

5. 个人反思与科学史观
杨立昆坦言，自己曾因未及时发表反向传播算法等想法而被他人抢先，但他强调科学思想的产生是复杂链条的结果。世界模型的概念并非新，早在1960年代的控制论和航天工程中已有应用，真正的挑战在于将理论转化为可工作的系统。

6. 对AI未来的愿景
杨立昆的路线宣言揭示了AI行业的两大方向：

  • 短期路径：硅谷的算力+数据竞赛，追求产品落地和商业价值。
  • 长期路线：基于认知+感知的基础研究，追求技术突破和智能本质。
    他相信，只有多元视角和开放探索才能避免行业陷入路径依赖，而世界模型可能成为通向真正智能体的关键钥匙。

结语
杨立昆的坚持不仅是技术选择，更是对AI本质的深刻理解。他的行动提醒我们：AI的未来不在于规模的堆砌，而在于能否真正理解并预测世界。无论最终结果如何，他的使命感和勇气为行业提供了宝贵的多元视角，也为我们展现了科学家的初心——让技术服务于人类更智慧的未来。

Reference:

https://www.youtube.com/watch?v=7u-DXVADyhc

<
Previous Post
Google发布Gemini 3.0 Flash
>
Next Post
Anthropic-AI vending machine agent