在周二于Meta伦敦举行的一次活动中，图灵奖得主、Meta AI部门负责人Yann LeCun对通用人工智能（AGI）的可能性表达了深度的怀疑。LeCun表示，人类智能的普遍性是不存在的，因此追求“人类水平的AI”更为实际。他批评了当前的AI系统在推理、规划、持久记忆和理解物理世界等四个关键认知领域的局限性。
LeCun特别点出了大型语言模型（LLMs）的不足，尽管它们在语言生成上表现出色，但对现实世界的理解却十分有限。他引用自动驾驶汽车的安全性问题、家用机器人的家务处理能力以及智能助手的单一任务执行能力，作为AI应用局限性的实例。
他还指出，人类通过与环境的直接互动来学习和理解世界，这与LLMs仅依赖文本数据的方式大相径庭。LeCun估计，一个四岁孩子通过感知和体验所获取的信息量，甚至超过了目前最大的LLMs处理的数据量的50倍。这一观点强调了AI在模拟人类智能道路上面临的巨大挑战。
LeCun的言论引发了对当前AI研究方向的反思，暗示了在追求更为先进的人工智能技术时，需要更注重模拟人类的学习方式和理解深度，而不仅仅是扩大模型的规模或处理文本数据的能力。
英语如下：
**News Title:** “Turing Award Winner Yann LeCun: Large Models Fall Short of Human Intelligence, AI Faces Four Cognitive Challenges”
**Keywords:** Turing Award, Large Models, Limitations of AI
**News Content:**
### Yann LeCun, Turing Award Winner, Cautions: Large Models Can’t Reach Human Intelligence, AGI Concept Questioned
During an event at Meta’s London office on Tuesday, Yann LeCun, a Turing Award recipient and head of Meta’s AI division, expressed deep skepticism about the feasibility of artificial general intelligence (AGI). LeCun argued that human intelligence is not universal, making the pursuit of “human-level AI” a more realistic goal. He criticized the limitations of current AI systems in four key cognitive domains: reasoning, planning, long-term memory, and understanding the physical world.
LeCun specifically highlighted the shortcomings of large language models (LLMs), despite their prowess in language generation. He pointed out their limited understanding of the real world, using examples such as safety concerns with autonomous vehicles, the capabilities of household robots, and the single-task execution of AI assistants to illustrate AI’s application limitations.
He also noted that humans learn and understand the world through direct interaction with their environment, a stark contrast to LLMs, which rely solely on textual data. LeCun estimated that a four-year-old child acquires information through perception and experience at a rate 50 times greater than the largest LLMs currently process. This underscores the significant challenges AI faces in emulating human intelligence.
LeCun’s remarks prompt reflection on the direction of current AI research, suggesting that in pursuing more advanced AI technologies, greater emphasis should be placed on mimicking human learning methods and depth of understanding, rather than merely expanding model size or text-processing capabilities.
【来源】https://thenextweb.com/news/meta-yann-lecun-ai-behind-human-intelligence
Views: 1