A study of large language models (LLM) showed that the famous artificial intelligence systems from Openai and Google organize information on its own and in unexpected ways, although they were not trained in that, independent. Reports.
The newspaper’s authors, experts from the Chinese Academy of Science and the University of South China Technology, concluded that such models “have deep similarities with how humans are building concepts.” This challenges the common belief that artificial intelligence reproduces only answers by identifying patterns.
In the experiment, Openai’s Chatgpt-3.5 and Gemini Pro Vision from Google were asked to do a task to find the individual object. As a result, models created 66 conceptual dimensions to classify organisms.
“These dimensions can be understood and described, indicating that large language and multimedia models form human -like representations,” the researchers pointed out.
“This provides a strong argument that although these offers are not identical to humans, they participate in the common foundations and reflect the main features of the human knowledge of the world.”
The authors hope that these results will help create artificial intelligence with more cognitive capabilities “human” – those that can cooperate better with people.
At the same time, systems are already developed closer to human nature around the world. For example, an emerging company from Australia recently provided the world’s first commercial biological computer based on live brain cells.
The cortical laboratory system is a “body in a square”: depends on the neurons planted in the laboratory on the silicone chip that can receive and transport electrical signals.
According to developers, this combination of biology and electronics allows them to learn and adapt faster than traditional computers. Even at an early stage, 800,000 nerve cells in humans and mice managed to mastery of the Pong game independently.
As stated in a sheet in the magazine cell, neurons showed signs of awareness when interacting with the game environment.
“We combine biology with classic computing to create a global educational engine,” the company website says. “Nerve cells are self -refinement, and it has almost flexibility, and it is a product of development for four billion years.”
the Ticket The research, titled “The Concept of Human Objects”, appears naturally in the large multimedia models, “in the journal Nature Machine Intelligence.
