LLMs can't reach human-level intelligence: Meta AI chief Yann LeCun
Meta's Chief Artificial Intelligence Scientist, Yann LeCun, has expressed doubts about the ability of large language models (LLMs) to achieve human-like reasoning and planning capabilities. Despite their impressive performance in several tasks, he believes these models are fundamentally limited. In an interview with Financial Times, LeCun warned against over-reliance on LLMs in the pursuit of human-level intelligence, describing them as "intrinsically unsafe."
LeCun highlights fundamental limitations of LLMs
LeCun pointed out several fundamental limitations of LLMs. They include their lack of understanding of logic, absence of persistent memory, limited grasp of the physical world, incapacity for hierarchical planning, and inability to reason. He also stated that these models heavily depend on pre-existing training data and can only provide precise answers, within the scope of their training.
Advocating for new generation of AI systems
LeCun advocates for a shift in approach toward developing a new generation of AI-powered systems, that aim to imbue machines with human-level intelligence. He acknowledged the nature of this vision and estimated that it might take up to 10 years to realize. Despite investor concerns about immediate returns on AI investments, LeCun believes they are on the brink of a breakthrough in AI systems.
Meta's AI research focuses on "world modelling"
LeCun leads a team at Meta's Fundamental AI Research (FAIR) lab that is pursuing a different path called "world modelling." This approach aims to create AI systems capable of developing common sense and learning about the world in ways similar to human cognition. Despite the continued progress in LLMs by Meta and its competitors, LeCun remains committed to this alternative approach toward achieving artificial general intelligence (AGI).
LeCun envisions AI interaction through wearable technology
LeCun envisions developing AI agents that users can interact with through wearable technology, like augmented reality (AR) glasses and EMG bracelets. However, there is still skepticism among experts about the possibility of achieving human-level intelligence in AI systems. Despite these challenges, LeCun is committed to pushing the boundaries of AI research toward this ambitious goal, viewing it not just as a product design or technological problem, but as a fundamental scientific challenge.
