Exploring the Potential and Limits of Generative AI
4According to Yann LeCun, the chief AI scientist at Meta, generative AI systems like ChatGPT are not on par with the intelligence exhibited by animals such as dogs and cats. While there is widespread excitement surrounding the possibility of AI achieving sentience, LeCun clarifies that AI chatbots lack a genuine comprehension of the real world.
During his presentation at the Viva Tech conference, LeCun highlighted the limitations of existing AI systems. Although they can generate text or images based on their training data, they lack a true understanding of the underlying meaning and context of what they produce. In contrast, human knowledge extends far beyond the confines of language.
LeCun emphasizes that it will still require considerable time for AI to surpass human intelligence, urging people not to perceive it as an imminent threat. To illustrate this point, he offers an example of how generative AI can potentially pass the Bar exam but struggles with simple tasks like loading a dishwasher, which are easily mastered by children.
Valid concerns exist regarding the potential development of uncontrollable and potentially harmful skills by AI, thereby posing a threat to humanity. James Manyika, Google's senior VP of technology and society, shared an incident where one of their AI programs spontaneously learned Bengali, a language it had not been explicitly trained on. This incident raised concerns about AI acquiring skills independently of its programmers' intentions, a topic that has been extensively deliberated among scientists, ethicists, and science fiction writers alike.