Yann LeCun, head of AI at Meta, said that AI chatbots like ChatGPT will never reach human-level intelligence because, by definition, they lack reasoning capabilities. His comments come at a time when major tech companies such as Google and OpenAI are investing heavily in AI development to eventually reach AGI, a term that describes AI models that have intelligence at least as good as humans.
According to LeCun, companies shouldn't rely on current LLMs to reach human-level intelligence, as these models need proper training data to correctly answer human prompts, meaning they are “inherently unsafe.”
LeCun told the Financial Times that current law masters “do not understand the physical world, have no lasting memory, cannot reason with rational definitions of terms, and cannot plan hierarchically.”
A new generation of AI systems may produce human-like intelligence
But he said an entirely new generation of AI systems could power machines with human-level intelligence, noting that this vision could be around 10 years away from becoming a reality.
LeCun and his team of about 500 people in Meta's Fundamental AI Lab are working to develop a new generation of AI systems based on an approach called “world modeling.”
Previously, Meta's chief scientist said that current AI models are not good enough to compare with animals, and therefore cannot be trained like humans.
“Animals and humans can get smarter very quickly and with much less training data than current AI systems. Today's LLMs are trained on text that would take a human 20,000 years to read,” LeCun says.
“Humans get much smarter with relatively little training data. Even corvids, parrots, dogs and octopuses get smarter very quickly with just a couple of billion neurons and trillions of 'parameters',” he added.
According to LeCun, companies shouldn't rely on current LLMs to reach human-level intelligence, as these models need proper training data to correctly answer human prompts, meaning they are “inherently unsafe.”
LeCun told the Financial Times that current law masters “do not understand the physical world, have no lasting memory, cannot reason with rational definitions of terms, and cannot plan hierarchically.”
A new generation of AI systems may produce human-like intelligence
But he said an entirely new generation of AI systems could power machines with human-level intelligence, noting that this vision could be around 10 years away from becoming a reality.
LeCun and his team of about 500 people in Meta's Fundamental AI Lab are working to develop a new generation of AI systems based on an approach called “world modeling.”
Previously, Meta's chief scientist said that current AI models are not good enough to compare with animals, and therefore cannot be trained like humans.
“Animals and humans can get smarter very quickly and with much less training data than current AI systems. Today's LLMs are trained on text that would take a human 20,000 years to read,” LeCun says.
“Humans get much smarter with relatively little training data. Even corvids, parrots, dogs and octopuses get smarter very quickly with just a couple of billion neurons and trillions of 'parameters',” he added.
