Yang Likun, Chief AI Scientist at Meta: It is unfounded to worry that AI threatens human survival. The current mainstream models are not even as good as cats.

As artificial intelligence (AI) technology continues to advance, and AI chatbots like ChatGPT that can directly communicate with humans are born, more and more people are not only enjoying the convenience brought by AI, but also beginning to worry whether AI will weaken human traits and autonomy in the future, or even pose a threat to human survival. However, according to a recent report in The Wall Street Journal, Yann LeCun, a Turing Award winner and the chief scientist leading Meta's AI research, stated that although the current AI models have many uses, they do not even reach the intelligence level of pets, let alone surpass humans. Yann LeCun pointed out that the common feline animals usually have a mental model of the physical world, persistent memory, a certain degree of reasoning ability, and planning ability, but the current AI models, including those developed by Meta itself, have not yet reached this level. LeCun further explained that the problem with the current AI models is their design approach. Models like ChatGPT and other chatbots using large language models (LLMs) are only predicting the next word in the text content, which is due to their large memory capacity, but this is not reasoning, it just appears to be reasoning. The output content is merely a repetition of the training data they have received, so no matter how large the AI system is built, the current approach will not bring human-level Artificial General Intelligence (AGI) to humans. However, LeCun also stated that in the future, humans may still be able to create AI models with human-level characteristics, but it will take decades. In the future, when people communicate with AI systems, smart glasses, or anything else, these AI systems will need to have human-level characteristics, truly have common sense, and behave like human assistants. But creating an AI model with such capabilities will take decades, and the current mainstream approach is simply not capable of achieving this goal. It is worth mentioning that perhaps to realize this vision, LeCun is leading the Meta Fundamental AI Research team to study how to establish a model whose learning method is similar to that of small animals, which will build a model of the world through the received visual information.