French scientist: Artificial intelligence will not reach the level of human intelligence in large linguistic models

-
AI systems built on a large linguistic model will never be able to reach the level of human intelligence.

This opinion was expressed by the French mathematician Yann Licon, vice president of the company "META" (banned in Russia as an extremist organization), where he is responsible for studies in the field of artificial intelligence.

The rapid progress in developing artificial intelligence-based products in recent years has raised concerns and speculation that such systems will soon surpass human intelligence. For example, last March, Elon Musk expressed the opinion that “artificial intelligence could become smarter than anyone by 2025.”

Lacon commented on Musk's statements, expressing his doubts about the speed of progress of artificial intelligence. In an interview with the Financial Times, he said that systems based on large language models (LLM), such as ChatGPT and many of its counterparts, will not be able to reach the human level. According to the researcher, she "has a very limited understanding of logic, does not understand the real world, does not have a permanent memory, and cannot plan hierarchically, that is, achieve a goal by solving a series of small problems."

The scientist pointed out that LLM systems have limited capabilities due to the restrictions imposed on them, and their capabilities depend on the information taught by their developer, and they do not learn independently, as people do. “Certainly many people think that AI does logical reasoning, but most of the time it uses knowledge accumulated from a huge amount of data,” LeCun said.

Instead, the AI ​​Lab at Meta studies methods that rely on the principle of “simulating the real world.” In particular, the tester shows the system video recordings that expire before they are completed, and the system must guess what happened in the end. The mission Lacon mentioned is to develop an encryption system that allows artificial intelligence to understand information presented that is formulated in the form of abstract text.

According to the scientist, who is considered one of the pioneers in developing artificial intelligence, the implementation of this project may take 10 years.