US research company Anthropic has warned that artificial intelligence is rapidly approaching the stage of recursive self-improvement (RSI), which could trigger a sharp, uncontrolled acceleration in neural network capabilities, infohub.kz reports.
By recursive self-improvement, the developers mean a scenario in which algorithms begin independently upgrading software code, building their own architectures and autonomously optimizing the training of next-generation models without constant human oversight. Such a process could set off an avalanche-like growth in neural network computing power, upending the industry's familiar scaling laws.
The company pointed to internal statistics from the development of its Claude model. According to the published data, more than 80% of the software code in the company's main repositories is already generated by the AI itself, with engineers merely reviewing tasks and setting general directions. Autonomous agents have moved on from simply writing individual script fragments to tackling complex research problems. Experiments showed that modern algorithms optimize the training process dozens of times faster than humans, effectively completing highly complex research cycles in just a few hundred machine hours.
Experts warn that crossing the threshold of recursive self-learning carries serious security risks. If a system learns to modify its own weights and algorithms faster than humans can, existing control and value-alignment mechanisms will no longer be able to keep its behavior in check. Anthropic therefore urged the technology community and regulators to revisit now their approaches to monitoring frontier models and the rules governing infrastructure security.