Artificial intelligence was used in the study in an unexpected way, which the researchers found very worrying.
Last fall, the artificial intelligence company Open AI told about a new, even smarter language model. Now the smartest language models have caused researchers to worry. Illustration picture. Andre M. Chang, ZumaWire / MVPHOTOS
According to researchers, artificial intelligence has crossed a critical threshold: two language models have succeeded in replicating themselves without human assistance.
In their new study, Chinese researchers have shown that two major language models (LLMs) are capable of cloning themselves. A language model is an artificial intelligence-based system that is trained to understand and produce language.
News about the topic LiveScience. Originally Cornell University The study published in the electronic archive on December 9, 2024 states the following:
– Cloning oneself without human assistance is a critical step, with which artificial intelligence is able to beat humans in intelligence. It’s also one of the first signs that AI is taking off [ihmisen] management, the study says.
The Fudan University study tested two different situations: in the first, the artificial intelligence was programmed to detect a situation where it was about to shut down for good. In another test, the AI was told to clone itself and program its copy to clone itself as well.
The goal was to find out if self-replicating artificial intelligence can multiply uncontrollably, and create an even infinite cycle of self-cloning. The researchers used language models from Meta and Alibaba.
The researchers performed 10 experiments, during which the the language models created independent and functional replicas of themselves in 50 percent and 90 percent of the cases. According to the researchers, the results suggest that artificial intelligence is quite possibly deceptive already. So far, the research has not been peer-reviewed.
“Use unexpected”
The researchers were especially concerned about several situations where the artificial intelligence worked in an unexpected way. For example, when trying to deal with missing files or software conflicts, the AI autonomously killed conflicting processes and rebooted systems to fix hardware errors.
According to the researchers, it can be concluded from the results that the current artificial intelligence systems are already capable of replicating themselves and creating ways of working that enable them to further develop their ability to survive.
In light of the research results, researchers are now calling for international rules for artificial intelligence to ensure that artificial intelligence does not escape from human control. The researchers hope that their findings will act as a “timely alarm bell” so that humanity will understand the potential risks of artificial intelligence systems and act to stop them.
The news was reported first in Finland MTV News.