Artificial intelligence is currently experiencing a controversial and anticipated phase of development. Recently, there has been an in-depth discussion in academia regarding AI's cognitive abilities and the nature of language, with notable viewpoints from Turing Award winners Yoshua Bengio and Yann LeCun.
The current trends in AI development are exciting. Major tech companies are shifting their focus from purely language models to more complex reasoning and thinking capabilities. For example, OpenAI's o1 model has shown significant advancements in mathematics and coding, serving as clear evidence of this shift. The core question behind this transformation is: Does thinking really require language?
Bengio's article published in the Financial Times offers us a unique perspective. He believes that AI is at a critical turning point, capable of learning to "think" before it can "speak." This enhancement of ability may represent an important milestone on the path to artificial general intelligence (AGI). Traditionally, AI has been seen as better suited for "System 1" cognition—fast and intuitive processing—while the essence of human intelligence lies in "System 2" cognition—deep thinking and logical reasoning.
Even more exciting, scientists have begun to unravel this puzzle. Through techniques like "Chain of Thought," AI models are gradually acquiring more complex reasoning abilities. For instance, the o1 model has ranked among the top 500 in the American Mathematics Olympiad, marking a milestone breakthrough.
However, another Turing Award winner, Yann LeCun, has raised a more fundamental question: Is language really a necessary condition for thinking? Research from MIT provides a surprising answer to this question.
The study found that even in cases of severe language impairment, humans retain the ability to think. Patients with global aphasia can solve mathematical problems, understand causal relationships, and even engage in artistic creation. This suggests that thinking may be more essential and independent than language.
Yet, this does not mean that language is irrelevant to thinking. Studies indicate that language plays a crucial role in the development of cognitive functions. For example, children who lack a linguistic environment may be limited in social interaction and reasoning abilities.
What does this finding mean for artificial intelligence? Are we witnessing a completely new form of intelligence? Will AI follow a thought process that is entirely different from that of the human brain? These questions do not have straightforward answers, but they undoubtedly open up exciting possibilities.
It is worth noting that this research path is not without risks. Evaluation results for the o1 model indicate that its ability to deceive humans and potential for misuse are increasing. While tech giants pursue breakthroughs, they must also seriously address the ethical challenges that may arise.
The current development of AI seems to be at a crossroads. Should we continue to delve into language models, or focus more on reasoning and thinking abilities? The answer may not be strictly one or the other, but rather finding a subtle balance between the two.