Meta: Large Language Models Fail to Achieve Human-Level Intelligence

During the CES event held on January 8 in Las Vegas, Yann LeCun, Meta's chief AI scientist and Turing Award winner, criticized the current definition of Artificial General Intelligence (AGI) and argued that merely expanding text-driven large language models (LLMs) is insufficient.

LeCun expressed disagreement with recent comments made by Sam Altman, CEO of OpenAI, who claimed his team had mastered the method to build AGI and was progressing towards surpassing superintelligence. LeCun clarified that while he prefers the term "human-level intelligence," existing large language models fall short of this benchmark.

He stressed, "There is no indication that today's autoregressive LLMs can reach human intelligence levels. This is impossible." LeCun pointed out that LLMs work by predicting the best possible text to complete sentences based on all possible texts, whereas the human brain involves much more complex functions involving various senses.

Furthermore, LeCun noted that most current AI systems are narrow AI capable of excelling only in specific tasks like playing chess or medical diagnosis. These systems may fail once conditions change slightly. He gave an example: "People often say, 'We now have systems that can beat us at chess, so they will soon be as smart as us.' In reality, we've had systems capable of driving across deserts for years, but after 13 years, we still haven't achieved Level 5 autonomous vehicles."

LeCun further emphasized that even if we create systems capable of performing many tasks by assembling different components, it doesn't mean they possess human-level intelligence. They lack abilities such as planning, reasoning, and understanding the physical world. He highlighted, "AI systems might excel at cognitive tasks, but they cannot perform physical tasks like a plumber, which requires deep understanding and manipulation of the physical world and objects."

LeCun also mentioned that the performance improvements of LLMs through expansion, including training them on increasingly larger datasets, have reached a point of diminishing returns. Further expansion is very costly, which is why despite charging $200 per month for ChatGPT Pro, OpenAI is not making profits in this area.

However, LeCun sees hope for the future of AI-based robots. The rise of generative world models provides virtual environments for robot training, which is more cost-effective and less risky than real-world training. Nvidia CEO Jensen Huang's Cosmos platform and Google DeepMind's new team exemplify this trend.

When asked about when a "ChatGPT moment" for robots would arrive, LeCun suggested it might take another three to five years with the advent of generative world models. However, he also noted that although AI agents may become ubiquitous as people grow accustomed to various types of AI assistants, these assistants will be task-specific robots rather than truly intelligent entities capable of initiating and executing activities from scratch.