Meta AI Chief: "LLM Has Not Reached Human Level... Developing New AI"

"LLM Does Not Think Like Humans"
'Segye Modeling' Developed... "Next-Generation AI Pinnacle"

Meta Vice President and Chief AI Scientist Yann LeCun stated that the large language models (LLMs) powering current generative AI such as ChatGPT will not reach human intelligence levels.

On the 22nd (local time), Vice President LeCun said in an interview with major foreign media, "LLMs have a very limited understanding of logic."

[Image source=AP Yonhap News]

LeCun added, "(LLMs) do not understand the physical world, lack persistent memory, cannot infer reasonable definitions of terms, and cannot plan hierarchically."

He also argued against relying solely on LLMs to create human-level AI. He explained, "LLMs can only respond accurately if provided with the right training data" and "they are inherently unsafe."

Recently, big tech companies have been advancing LLM technology as a way to achieve artificial general intelligence (AGI) that thinks like humans and surpasses human capabilities. OpenAI recently released the GPT-4o model, Google unveiled the AI agent 'Astra,' and Meta also announced LLaMA 3.

LeCun evaluated, "To most people, it looks like reasoning, but the majority is leveraging knowledge accumulated from large amounts of training data" and "(LLMs) have limitations but are useful."

Instead, he said they are developing a completely new AI system that thinks and reasons like humans, which could take 10 years to realize. LeCun is developing 'world modelling' at Meta's AI research lab (FAIR), which learns in a way similar to humans.

Foreign media assessed that in a situation where investors want to quickly secure profitability from AI investments, such experimental challenges are risky and costly gambles. In fact, last month, when Meta CEO Mark Zuckerberg said he would increase investments to make Meta the world’s top AI company, Meta’s market capitalization dropped by about $200 billion (approximately 273 trillion KRW).

However, LeCun emphasized that Meta is already ahead in the race toward next-generation AI. He said, "We are probably at the point where we think we are at the pinnacle of next-generation AI systems."

At FAIR, various experiments are underway to create AI that thinks like humans. They mimic how children learn by omitting parts of videos and then having AI predict what happens next, or attempt to build a universal text encoding system that applies abstract expressions in text to video and audio.

LeCun said, "In the long term, this technology will enhance AI agents that interact with users through wearable technologies such as extended reality (XR), smart glasses, and electromyography bracelets," adding, "For AI agents to be truly useful, a level of intelligence similar to human-level intelligence is necessary."