Artificial intelligence is advancing to such an extent that numerous people are starting to question—and are truly worried—about whether AI is poised to take over their roles. Though this might be accurate, in some cases Meta AI Chief Scientist Yann LeCun contends that we haven’t reached human-like AI yet. As one of the “three godfathers of AI” alongside Geoffrey Hinton and Yoshua Bengio, his opinions hold significant influence. He argues that despite numerous advances, AI cannot match humans because it currently lacks four key attributes. Researchers aim to incorporate these qualities into systems to develop so-called general-purpose AI—a term used within the field—but achieving this remains elusive as there’s no quick path to replicating human and animal cognition.
LeCun suggests that the four essential human attributes missing from current AI models, particularly large language models (LLMs), include comprehending the physical environment, maintaining long-term memory, reasoning effectively, and devising plans—especially through hierarchical planning. This doesn’t mean businesses developing these models aren't considering such capabilities; rather, their efforts so far have focused mainly on augmenting existing models with additional functionalities to close this gap.
To comprehend the physical world, you develop an independent vision system. Then, integrate this into the large language model [LLM]. As for memory, you can employ Retrieval-Augmented Generation [RAG], add some form of associative memory atop it, or simply expand the size of your model," he elaborated. Notably, RAG, shorthand for Retrieval Augmented Generation, was introduced by Meta—specifically researchers like LeCun—to improve how LLMs incorporate outside information into their outputs.
LeCun has rejected present "workarounds" and endorsed an alternate strategy employing what he refers to as "world-based models." Elaborating on this notion, LeCun explained, "At time T, you possess certain knowledge about the condition of the world; then, imagining a potential action, your world model forecasts how the situation will change based on that action taken." Thus, abstraction stands out as crucial for AI systems to effectively predict the countless and unforeseeable scenarios encountered in reality with accuracy akin to human-level understanding.
Meta is currently investigating this strategy through what they call "V-JEPA," introduced in early February. This tool uses a non-generative model designed to learn by forecasting obscured segments within videos. As explained, the core concept revolves around not predicting individual pixels directly but rather training a system to generate an abstract interpretation of the video content. The aim here is for these abstractions to enable more accurate predictions while filtering out unpredictable elements.
LeCun further clarified that grasping hierarchical structures is essential when trying to comprehend our surroundings—a capability lacking in many contemporary artificial intelligence frameworks today.
LeCun has consistently maintained faith in AI achieving human-level intelligence, acknowledging that it requires considerable time. In contrast to statements from individuals like Elon Musk—who asserted that AI might outsmart an individual person by 2025 and surpass all humankind collectively by 2029—LeCun has sought to alleviate worries regarding potential AI dominance over humanity with the assertion, “AI isn’t akin to a spontaneous occurrence that suddenly turns perilous.”
Stream English Language Television Channels Live
Watch Hindi Channels Live Television