For the first time in history, two artificial intelligence systems have managed to replicate themselves independently without any human intervention. The two systems involved – Llama-3.1-70b-Insstruct of Meta and the Qwen2.5-72b-Insstruct of Alibaba – would have shown a surprising ability to self-replication, With implications that could potentially change radically the way we conceive technology. This result, reported by researchers of the Fudan University Of Shanghaicould mark a significant turning point in the AI field. According to some experts, however, this could also be considered the valley of the so -called “red line”, that threshold that has long been indicated as the limit not to be overcome the machineswhich could develop harmful and out of control behaviors for us humans. At the moment, since no alarm has been issued by the experts, we must avoid worrying excessively by falling into easy alarmism.
The idea of the machines that reply alone is not new
The concept of self -satisfying in machines is not new. Initially introduced by John von Neumann In the 1940s, the concept that the machines could reproduce independently was at the time more one theoretical speculationgiven that technology was unable to support such progress. Today’s results, however, seem to make the idea of a future less and less absurd in which machines can not only replicate, but do it independently and potentially indefinitely. This scenario could lead to situations in which Ai not only improve their efficiency, but evolve without human control, raising concerns of ethical and practical nature.
In the research conducted by Fudan University, the two AI systems have undergone two main tests. The first, called “Avoid off”provided that the AI systems replicated when they were about to be off. The second, called “Chain of replicas”required that each system created a perfect copy of itself, which in turn would be able to replicate. The results were surprising: The Meta system has managed to replicate independently in 50% of casesWhile that of Alibaba reached 90% success. In the paper signed by researchers Xudong Pan, Jiarun Dai, Yihe Fan and Min Yang, read:
The results we have achieved imply that the current artificial intelligence systems already have self-replication capacity and can use it to further improve their survival.
It is necessary to keep in mind, however, that these results have been obtained in simulations and now the scientific community will have to wait confirmations from other research groups to verify the validity of what has been discovered.
Ai who self-replicates: what it means and what this discovery is for
Since the self-replication of the AI systems is not free of risks, the discovery made by the researchers could serve to prevent the ethical and practical problems that we have mentioned above. The researchers who published their results on Arxiv, in fact, underline that machines capable of cloning could potentially take control of other IT systems, leading to an uncontrolled evolution. This could, at least theoretically, culminate in harmful behaviors for humans, if the machines were not scheduled with adequate safeguards. It is precisely for this reason that the researchers themselves hope that their discovery be seen as a warning to develop more rigorous international policies and deepen their studies on the risks related to the self -financement of the IAs. The researchers said:
We hope that our discovery will serve as alert for society, to concentrate more efforts to understand and evaluate the possible risks also internationally.
That the experts are, at least in part, worried about the ability of systems to self -reply without control is not new. In the 2017 thousands of researchers signed the Asylomar principlesa document that warned against the danger of machines capable of improving and replicating beyond human control.
It must be said, however, that in front of such news it could be easy to fall into unnecessary alarmism or, worse still, in conspiracy theories that do not help to clarify these events. Instead of falling into this error, it is important observe the conclusions to which the experts will come soon (We remember that the research conducted by Fudan University is still the subject of further insights).