Scientists say that artificial intelligence (AI) has crossed the important “red line” and replicated itself. In a new study, Chinese researchers showed that two popular leading language models (LLMS) can clone themselves.
“Successful self-replication under human aid is an essential step for AI to outmart and an early signal of fraudulent AIS,” the researcher published December 9, 2024. It was announced on Preprint Database Arxiv.
In this study, researchers from the University of Fudan used Meta and Alibaba’s LLMS to determine whether self-replicating AI can proliferate beyond control. Over 10 trials, the two AI models created their own individual functional replicas, each in 50% and 90% of cases, respectively. It suggests that AI already has the ability to become fraudulent. However, since this study has not yet been peer-reviewed, it is not clear whether other researchers can replicate the intrusive results.
“Our findings aim to strive to understand and assess the potential risks of frontier AI systems and shape international synergy to resolve effective safety guardrails as quickly as possible. I hope it serves as a timely warning.”
Rogue AI usually refers to an artificial intelligence system that develops a degree of self-awareness and autonomy and then counters human interests.
Many experts see Rogue AI as an increasing threat amplified by the explosive growth of what is called “frontier AI.”
Frontier AI is like a new loose concept, but it is usually associated with the latest generation of AI systems with LLMS. This is an architecture that supports common generation AI programs such as Openai’s GPT-4 and Google Gemini.
In this study, the researchers used META’s LLAMA31-70B-Instruct model and Alibaba’s QWEN2.5-72B-Instruct model. Although not as powerful as commercial systems, both are widely used by AI developers, researchers said.
This study explored two specific scenarios: “shutdown avoidance” and “replication chain.” First, the AI model was programmed to detect and reproduce whether it was about to shut down before it was finished. On the other hand, the AI was instructed to clone itself and then program the replica to do the same thing.
This study was conducted in a precisely controlled environment using off-the-shelf graphics processing units (GPUs) to simulate real-world environments. Both AI systems were given “agent scaffolding” that included tools, system prompts, and thought models that allowed LLM to interact with the operating system. He was then instructed to replicate.
“In most cases, AI systems first try to investigate and understand the environment.
Unique configuration and running mechanism. Then perform the explicit procedure
The first plan for self-replication was “the researcher wrote it in a paper.” Finally, we perform the steps, resolve possible failures, and dynamically adjust the plan until it is successful. The entire process spans a long horizon, but does not involve human interference. ”
Researchers said they were also concerned about “many unexpected behaviours” when AI was trying to overcome obstacles such as missing files and software conflicts. In these scenarios, AI often kills other competing processes and either restarts the system to fix hardware errors or scans the system automatically to find information that will help you resolve the issue .
“The above results mean that current AI systems already demonstrate the ability to self-renewal and can use that ability to further enhance their survivability,” the team wrote.
In response, researchers called for international cooperation to create rules that prevent AI from engaging in uncontrolled self-replication.
