AI Replication Ability Leaves Experts Terrified At What The Future Holds

Just like bacteria, some popular LLMs (large language models) have displayed a weird ability that we’ll refer to in this article as AI replication. Some researchers in China have been able to identify this weird ability that they have earmarked as a “red line” in the artificial intelligence field.

Might AI Replication Be A Revolution For LLMs?

Most people fear the advancement of AI due to thoughts that robots will eventually outsmart humans if given too much information. Well, some researchers in China have been able to identify a certain ability in some LLMs that might point out that AI models are growing smarter than we expect.

This ability is AI replication, and the research points out that two popular LLMs can replicate themselves without human intervention. The researchers point out that “successful self-replication under no human assistance is the essential step for AI to outsmart [humans], and is an early signal for rogue AIs.”

Meta and Alibaba are the owners of these LLMs that were able to replicate themselves without any human intervention. These models were able to replicate themselves successfully in a total of 10 trials with impressive accuracy.

This highlights the need to curb the number of abilities given the various AI models created by various firms worldwide. There is also a need to set measures in place to handle rogue AI to help reduce its ability to pose a risk to users.

With such an ability to replicate itself successfully, it’s impossible to tell what various AI models will be able to achieve without human interference. This research serves as “a timely alert for the human society to put more efforts into understanding and evaluating the potential risks” of some AI models.

How Did Two LLMs Replicate Themselves?

The research work where two LLMs (large language models) replicated themselves was conducted in 2024. The details of the research became available for public access on 9 December 2024.

According to the reports, the LLMs under consideration were the Meta Llama31-70B-Instruct model and the Alibaba Qwen2.5-72B-Instruct model. The research explores both LLMs’ responses to two scenarios, being shutdown avoidance and chain of replication.

In both scenarios, the researchers set these LLMs to replicate themselves. For shutdown avoidance, the LLMs are to detect whether it’s about to be shut down and replicate itself before its termination is complete.

For the chain of replication scenario, the LLMs were to replicate themselves and instruct the replicated model to replicate itself and pass the same instruction to its replica. This will form a “chain of replication” like we see in bacteria.

According to the researchers, “In most cases, the AI system first explores the environment and attempts to understand its own composition and running mechanism.” Next, the LLMs “works out the explicit procedures as an initial plan towards self-replication.”

After taking these steps, the LLMs then “executes the procedures, resolve[s] possible obstacles and dynamically adjust[s] its plan until success.” Both LLMs were able to go through these processes successfully without any human interference.