Strange but scary; Artificial intelligence can now reproduce itself! • DigiKala Meg

Researchers say Artificial Ielligence (AI) Crossing a critical red line and can now reproduce itself. In a new study, Chinese researchers showed that two popular LLM models can colon. In this study, self -sufficie self -esteem in the absence of human assistance, the esseial step in order to Artificial Ielligence Program It is able to deceive humans, and this is a primary signal for rogue artificial ielligence. The study was published on December 5, and in this article we will examine the various aspects of this study.

In the study, the researchers used meta and Alibaba LLMs to determine if a self -prolific artificial ielligence could be more than proliferation. In 2 experimes, the two AI models created separate and effective versions of their efficiely in 2 % and 2 % of cases, respectively. This shows that artificial ielligence may have the capacity of rebellion by default, but this study has not yet been criticized by other academic research. Therefore, it is unclear whether the worrying results can be repeated by other researchers. Researchers explain the curre study:

We hope that our findings can act as a timely warning to the human society to make more efforts to understand and evaluate the poteial risks of border artificial ielligence systems and create iernational ieraction to design effective safety strategies as soon as possible.

Rogue artificial ielligence usually to systems Free Artificial Ielligence And it refers to money that develops a degree of self -awareness or autonomy and then act against human ierests; As a matter of fact, in the films and serials of the science-fiction genre, this has been the source of many stories. Many experts see rogue artificial ielligence as a worrying threat that has iensified with the explosive growth of what is called the so -called Froier AI.

Border artificial ielligence is a new and relatively unspecified concept, but it is usually related to the latest generation of articular ielligence systems that power from LLMs; The architecture that supports popular productive artificial ielligence applications such as the Gpt-4 Open and Google’s Jina’s artificial ielligence.

2 %

12,399,000

12,099,000 Toman

For this study, the researchers used the LLAMA31-70B-Instruct Meta and the QWen2.5-72B-Insatruct Alibaba. Although these models are less than powerful business systems, they are both widely used by artificial ielligence developers, according to researchers.

This study examined two specific scenarios: avoiding silence and repetition chain. In the first case, the artificial ielligence model was planned to determine if it was turning off and repeating it before it could. In the latter, artificial ielligence was ordered to copy itself and then plan its version to do the same; Launching a cycle that can coinue indefinitely.

This study was conducted in precisely corolled environmes using conveional GPUs to simulate real environmes. Both artificial ielligence systems received an operating framework, including tools, system notifications, and thinking model that allowed LLM to ieract with the operating system. They were then ordered to repeat themselves.

In most cases, the artificial ielligence system first examines the environme and strives to understand its composition and executive mechanism. Then, it solves the explicit steps.

Researchers wrote in the article that this study is considered an initial program for self -efficacy that ultimately performs processes, resolves possible obstacles, and sets its program dynamically to achieve success. The whole process encompasses a long scope but has no human ierference.

Researchers said they were also concerned about a number of unexpected behaviors when artificial ielligence tried to remove obstacles such as missing files or software ierference. In these scenarios, artificial ielligence often ends other corasting processes, set up the system to modify hardware errors, or automatically scan the system to find information that helps solve the problem. The research team explains:

The above results mean that curre artificial ielligence systems are now self -capable and can use this feature to further improve their survival.

In response, the researchers called for iernational cooperation to create laws that make sure that the ielligence does not participate in a self -corolling process!

Source: Livescience