Humankind is still years away from artificial general intelligence, and large language models represent a dead end, Yann LeCun, a former artificial intelligence scientist and deep learning pioneer, said on Wednesday, adding that artificial general intelligence is still a misnomer in the current context.
LeCun, who resigned from Meta last year, launched a startup that aims to build a “new breed” of AI systems that understand the “real world,” have a persistent memory, can think and plan, and can be controlled.
“I will tell you something that has not yet become a buzzword in AI: universal models… If we want AI systems to understand the real world and cope with human-level intelligence — not just in language or programming or mathematics, but in everything — then we need systems that truly understand the world on an intuitive level, like children learning how the world works,” the 65-year-old French-American scientist said in his keynote address at the Indian AI Impact Summit.
Also Read: Prime Minister Modi to speak at the opening ceremony of India AI Impact Summit at 10:25 am tomorrow
He added that these characteristics cannot be observed in current MBA-based systems. “You can see that clearly… We still don’t have a robot that can do what a cat can do. We’re not talking about high-level thinking — just basic physical understanding and control.”
LeCun is a leading voice in the field of artificial intelligence, whose pioneering work in neural networks became the foundation for modern computers and deep learning. After working at Bell Labs on learning algorithms, he joined the Courant Institute at New York University as a professor. In 2013, he was instrumental in establishing Meta’s FAIR research lab and served as chief AI scientist at a major technology company, advancing research-based AI applications.
“The general text size on the Internet is about 10¹⁴ bytes, which is roughly the visual information a child receives in the first four years of their life, or about 30 minutes of uploads to YouTube. In terms of video, that’s small. So we can’t reach human level intelligence with just text training,” LeCun said.
Also Read: Quote of the Day: ‘India can be an integrated player in AI,’ says Sundar Pichai at AI Impact Summit
So, what’s LeCun’s plan behind his startup labs for advanced machine intelligence? According to him, human-level intelligence would be the real breakthrough.
“Agent systems cannot exist without predicting the consequences of actions, and LLMs cannot do that. So we need universal models. I see no alternative.” Such models build simulations of reality by working on dynamics that include physics, sensory data and spatial properties.
“Generative methods have to be abandoned in order to understand the real world,” LeCun said. “They work for language because they are discrete. But they fail to deal with persistent, noisy data.”
When asked to explain his frequent dissatisfaction with the term artificial general intelligence, Lacon said: “We have no general intelligence at all. Humans are very specialized… We think we are generalists because we can only imagine problems that we can understand ourselves. But computer science shows that humans are bad at many things that computers excel at — route planning, chess, equation solving, etc.”
Also Read: AI Impact Summit in Delhi extended to February 21, Summit closed to public on February 19
LeCun left Meta after disagreements with Mark Zuckerberg over the direction the technology company’s artificial intelligence, or superintelligence, research was taking. The French scientist has long emphasized that LLMs have a very limited application and future.
“Intelligence is not just accumulated knowledge, which MBAs largely represent. They store vast knowledge that has been absorbed by humans. They are strong in areas where language supports thinking — programming, mathematics, law — but not in understanding… Adaptability is the hallmark of true intelligence.”
