With the advent of ChatGPT, the slightly obtrusive term “large language model” (LLM) became the darling of the spotlight. From star startups such as OpenAI and Anthropic to companies such as Google, Microsoft, and Meta, they are continuing to advance LLM's capabilities and commercial application prospects.

However, Yann LeCun (Yann LeCun), chief scientist of artificial intelligence at Meta, believes that the current LLM route cannot lead to AGI and is very dangerous, showing that the industry is divided over the AI development roadmap.

Yang Likun: LLM has major flaws and cannot achieve real AGI

Recently, in an interview with the media, Yang Likun bluntly pointed out that the current LLM technology has major flaws such as “extremely limited ability to understand logic,” “inability to model the physical world,” “inability to form lasting memory,” and “inability to perform hierarchical planning and reasoning,” and that simply pursuing the development of LLM is “inherently unsafe” and cannot achieve real AGI.

LeCun pointed out that although the existing LLM excels in natural language processing, dialogue interaction, text creation, etc., it is still only a “statistical modeling” technique. By learning the statistical rules in data, it is essentially not capable of “understanding” and “reasoning”.

But until this year, tech giants, including OpenAI and Google, saw LLM as a key step towards AGI. OpenAI CEO Sam Altman has stated many times that the GPT model is an important breakthrough in the direction of AGI.

What Yang Likun advocates is the so-called “World Modeling” (World Modeling) method, which is the ability to gradually learn to recognize the world through observation and experience, just like humans, to form “common sense” and ultimately achieve AGI.

Yang Likun believes that the “world model” is closer to true intelligence than just learning the statistical characteristics of data. Take the human learning process as an example. As children grow up, they perceive the world more through observation, interaction, and practice, rather than simply being “injected” with knowledge. Yang Likun's “world model” route is an attempt to allow AI to experience this kind of self-learning process through simulating and completing media such as video and audio.

However, he also acknowledged that achieving a “world model” is not an easy task, and that this ambitious goal may take 10 years to achieve. Industry experts are also skeptical about this.

Aaron Cullota, a computer science professor at Turan University, pointed out that “common sense” has always been a pain point in the development of artificial intelligence. Teaching AI models “causal relationships” is not an easy task, and it is easy to “fail unpredictably.” Previously, Meta employees also questioned Yang Likun's “world model” concept being vague; it was more like a gimmick.

Meta needs to quickly commercialize AI, but the chief scientist frequently breaks it down

Essentially, Yang Likun's differences with companies such as OpenAI and Google over the AI development path also reflect the contradictions and differences between tech giants in investing in the AI field.

For Meta, which spent a lot of money to buy GPUs, the primary mission is to quickly launch commercial AI products in the short term. After the failure of the metaverse, the company's achievements in AI exploration are critical to Meta's future. Moreover, Zuckerberg is facing tremendous profit pressure. Last month, when Zuckerberg vowed to make Meta “the world's leading artificial intelligence company,” Meta once evaporated its market value of nearly 200 billion US dollars, reflecting investors' concerns about the return on investment in artificial intelligence.

But Yang Likun, who is the chief scientist, values exploratory research and long-term value more than short-term income. He has publicly claimed that there are major problems with the LLM route, and he is also punching his own company in the face — Meta is also one of those companies that spend huge sums of money to develop “dangerous LLM.”

Outsiders have long speculated that there may be some kind of delicate conflict between Meta CEO Zuckerberg and Yang Likun. For example, at the end of last year, Zuckerberg divested the “AI commercialization” work from Yang Licun and formed an independent “GenAI” team, which directly divided the power of Yang Likun's “academics.”

Overall, Meta is still the main beneficiary of the AI wave. So far this year, Meta's stock price has surged 35%.