Author: Techub News Compilation

Introduction

In March 2025, Meta's Chief AI Scientist and Turing Award winner Yann LeCun participated in an in-depth interview with science communicator and former EU digital ambassador Dr. Matt Keçi. As one of the founders of deep learning, LeCun not only reviewed the ups and downs of the field over the decades but also analyzed the fundamental limitations of current AI systems represented by large language models (LLMs), outlining his vision for the evolution of the next generation of AI. In a time when AI technology is rapidly evolving and the dynamics of competition and cooperation are increasingly complex, the thoughts of this top scholar provide a crucial perspective for understanding the future direction of AI development.

Summary

• Current AI (especially LLMs) can manipulate language proficiently but are "very stupid" in understanding the physical world, persistent memory, reasoning, and planning, and are far from the level of human or animal intelligence.
• Achieving human-level AI hinges on enabling machines to learn "intuitive physics" efficiently through sensory inputs like babies or animals, which requires new architectures, such as his proposed JEPA (Joint Embedding Predictive Architecture).
• Open source and open research are the cornerstones of the rapid development in the AI field over the past decade; closed systems will hinder overall progress; the future of AI lies in global collaboration rather than zero-sum competition.
• True autonomous robots (including Level 5 autonomous driving) depend on AI systems that can understand the physical world and possess common sense, which may be a prerequisite for the impending "decade of robotics."
• The emotions of AI may stem from "expectation" of action outcomes rather than hardcoded anger or jealousy; consciousness currently lacks a clear definition and should not become the focus of research.

From the Wave of Deep Learning to the "Stupid" State of AI

Yann LeCun reviewed the two waves of deep learning: the first occurred from the late 1980s to the mid-1990s, when silence ensued due to limitations in computational power and data; the second has started explosive growth around 2013 and became widely known due to the landmark review paper he published with Geoffrey Hinton in Nature in 2015. However, in the face of the fervor caused by models like ChatGPT today, LeCun gave a calm and even sharp assessment: "Current AI systems are very stupid in many ways." He pointed out that we are misled by their fluent language capabilities into thinking they are smart, but in reality, they cannot understand the physical world, lack the persistent memory of humans, and cannot truly reason and plan—features that are core to intelligent behavior.

LeCun emphasized that language appears complex precisely because it is a sequence of discrete symbols that are relatively simple to process. In contrast, the quantity of information received about the physical world through sensory inputs like vision is massive and continuous, making it much more challenging to understand and predict this world. This is referred to as "Moravec's Paradox": it is relatively easy for computers to complete abstract thinking tasks (like playing chess), but extremely difficult to equip them with the physical intuition and action capabilities of a baby or a kitten. He noted that the total amount of text data used to train a large language model (about 10^14 bytes) is equivalent to what a toddler can receive through visual experiences in about four years. This suggests that solely relying on text training will never achieve human-level AI; we must enable AI to learn to understand the real physical world.

The Blueprint for Next-Generation AI: Understanding Physics, Memory, Reasoning, and Emotion

So, how can we build AI that understands the physical world? LeCun introduced new paths his team is exploring at Meta's FAIR lab and New York University. They aim to design a novel AI system that is still based on deep learning but has broader capabilities. The key abilities of this system include:

• Understanding the physical world: Master intuitive physics and be able to predict the consequences of one's actions.
• Persistent memory: Having a memory system similar to that of humans.
• Reasoning and planning: Capable of hierarchical planning, breaking complex goals into executable sequences of subtasks.

LeCun particularly pointed out that the reasoning method of current LLMs (generating and filtering extensively in "token space") is inefficient and not like that of humans. Humans simulate and reason in a mental "world model," predicting the outcomes of different action paths to formulate plans. Building such a world model is a significant challenge.

Regarding emotion and consciousness, LeCun proposed an interesting viewpoint: future advanced AI systems may possess "emotions," but these are not programmed anger or jealousy, but stem from "expectations" of action outcomes. When the system predicts that its actions will successfully achieve goals, it generates positive signals akin to "happiness"; conversely, it produces negative signals. This is an intrinsic drive based on a predictive mechanism. As for consciousness, he believes that there is currently no clear definition or measurement method, and that excessive focus on it may be akin to historically getting entangled in questions like "why do we see the image as right-side up when the retinal image is inverted."

Evolving Learning Paradigms and the World Model Architecture JEPA

LeCun outlined three major paradigms of machine learning: supervised learning, reinforcement learning, and self-supervised learning. He noted that reinforcement learning is inefficient (not suitable for training autonomous driving due to crashing thousands of times), while the LLMs that have achieved huge success in recent years primarily rely on self-supervised learning (like predicting masked words in text). However, applying this self-supervised learning, which is suitable for discrete text, directly to the continuous, high-dimensional physical world (like video prediction) has failed. This is because it cannot effectively represent the probability distribution for predicting future video frames.

To this end, LeCun proposed JEPA (Joint Embedding Predictive Architecture) as a solution. The core idea of JEPA is: not to predict in the original input space (like pixels) but to enable the system to learn the abstract representation of the input and make predictions within that abstract representation space. This avoids the system wasting resources predicting unpredictable details (like all the details of a painting on the wall after the camera moves) and instead focuses on higher-level, more essential changes. JEPA is a macroscopic architecture that can include modules like Transformers. It does not directly replace Transformers but provides a more feasible framework for processing information from the physical world.

Open Source Collaboration, the Future of Robotics, and Industry Insights

In light of the competition in the AI field, especially the discussions triggered by the recent launch of the high-performance open-source model DeepSeek by Chinese teams, LeCun firmly defended the value of open source and open research. He emphasized that open source benefits the entire field and the world at large, and is the fundamental reason behind the rapid progress of AI over the past decade. The widespread use of open-source frameworks like PyTorch is evidence of this. He believes that competition at the foundational research methods level is not a zero-sum game but a global collaboration. Good ideas come from all over the world, and an open environment can accelerate progress in the entire field. Regarding rumors of Microsoft's multi-billion-dollar "Stargate" computing project, LeCun noted that its investment scale is on par with Meta and Microsoft's current inputs, and that most of the investment will go towards fulfilling the inference needs of billions of users rather than training.

When discussing robotics, LeCun believes the next decade will be the "decade of robotics," but its realization hinges on breakthroughs in AI's understanding of the physical world. Current humanoid robot demonstrations are impressive, but their level of intelligence is still far from sufficient to tackle the complexities of the real world. He specifically commented on Elon Musk's predictions regarding Tesla's autonomous driving, bluntly stating "we must stop believing his statements in this regard," as his predictions have failed to materialize for eight consecutive years. The difficulty of Level 5 autonomous driving lies not in the sensors but in AI's lack of understanding and common sense regarding the physical world.

Regarding Europe's role in the AI competition, LeCun pointed out that Europe possesses top talent but faces challenges in regulatory clarity (for example, the visual capabilities of his Meta smart glasses have not been launched in Europe due to regulatory uncertainty). He believes that Europe will continue to play an important role.

Reflection, Legacy, and Future Directions

When asked if he has any regrets in his career, LeCun candidly stated that he may have realized the importance of self-supervised learning too late and wishes he had worked harder to maintain the community's interest in neural networks back in the day to avoid the emergence of an "AI winter." He speculated that Geoffrey Hinton's regret may lie in spending too much time searching for learning algorithms different from backpropagation that the brain may use, only to ultimately find that backpropagation itself is already very effective.

As the inventor of convolutional neural networks (CNNs), LeCun's legacy has permeated daily life, from mobile image recognition, voice assistants to automatic emergency braking systems in cars, all backed by CNNs. Today, he is focusing on the more cutting-edge "world model," aiming to endow AI with true physical understanding and common sense. At the end of the interview, he also expressed positive expectations for the application of AI in fields like healthcare (such as breast cancer diagnosis) and serves as an advisor to related startups, believing that deep learning will bring great benefits to human health.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。