Why Large Language Models Aren’t Truly Intelligent: LeCun’s View on AGI

Media hype around artificial general intelligence (AGI) is rampant, with daily headlines claiming that machines may soon surpass human cognition. In the past month, three tech leaders made bold predictions: Nvidia CEO Jensen Huang expects AGI within five years, "AI father" Ben Goertzel predicts three years, and Elon Musk has suggested the end of 2025.

However, not everyone shares this optimism. Yann LeCun, Meta’s chief AI scientist and Turing Award laureate, doubts the existence of AGI, arguing that human intelligence is far from general. He emphasizes four fundamental cognitive challenges that current AI systems cannot yet solve: reasoning, planning, persistent memory, and understanding the physical world.

LeCun explains that without these capabilities, AI applications remain limited and error‑prone—autonomous driving is still unsafe, household robots struggle with basic chores, and virtual assistants can only perform simple tasks. These deficits are especially evident in large language models (LLMs), which rely heavily on text and lack a deeper form of human knowledge.

According to LeCun, the apparent fluency of LLMs can be misleading; they appear intelligent because of language proficiency, yet their grasp of reality is shallow. He likens this to a dead‑end, noting that scaling up text data will not bridge the gap to human‑level intelligence.

Goal‑Driven AI as an Alternative

LeCun proposes a "goal‑driven" AI architecture that learns through interaction with the world via sensors and video, building a "world model" that predicts the consequences of actions. For example, moving a chair left or right changes the environment, and the system updates its internal memory accordingly, enabling it to plan multi‑step tasks.

He is confident that, over time, machines will surpass human intelligence, but stresses that this will not happen imminently and certainly not as soon as some, like Musk, claim.