A nonprofit organization dedicated to artificial intelligence safety, Model Evaluation and Threat Research (METR), has published conclusions that provide one of the clearest pieces of evidence so far that AI is accelerating its own development. This creates the potential for recursive self-improvement (RSI), a process in which artificial intelligence systems create stronger versions of themselves without direct human intervention. [1]
METR estimates that artificial intelligence could be responsible for approximately 80% of the current acceleration in technological progress, although researchers emphasize the uncertainty of this figure. In addition, no conclusive evidence has been identified regarding the existence of major obstacles that could slow down the progress of AI models next year. The organization also clarified that this acceleration is not due to an intensification of computing power used in training. [1]
The evaluation conducted by METR over a period of 11 weeks, based on confidential research, interviews, and performance data of yet-unreleased models, revealed a significant shift. The abilities of models developed by the company Anthropic improved 1.9 times faster than before, starting with October 2025. However, METR warned that AI is not yet capable of generating extremely significant and fully autonomous contributions in the essential stages of model training. [1]
The prospect of 'recursive self-improvement' has recently generated series of warnings from AI sector researchers. They argue that such systems could become difficult to align with human objectives, monitor, and control, potentially representing an existential threat to humanity. Such a technological transformation could have wide-ranging implications for the industry that has spent hundreds of billions of dollars building the compute infrastructure necessary for training advanced models, as artificial intelligence itself could become the main engine of future progress. [1]
METR's evaluation method included a detailed analysis of anonymized model performance, testing their ability to solve complex problems and generate coherent code or text. Results suggest that the qualitative leap is not linear but represents a regime change in the efficiency of learning algorithms. This trend highlights the urgent need for the scientific community to establish rigorous verification standards for the emerging capabilities of neural networks. [1]
- Categories:
- Technology
