Chinese researchers propose five levels of AI self-improvement — Live Science
Chinese researchers have proposed a model for the development of artificial intelligence in which systems could independently direct their own training and create subsequent generations of models. This was reported by Live Science. The authors of the study are examining the concept of recursive self-improvement, which could potentially help address the rising financial and computational costs of developing increasingly complex AI models.
Five levels of autonomy
In a preprint uploaded to the arXiv database on September 10 and not yet peer-reviewed, the scientists proposed dividing AI systems into five levels based on their self-improvement capabilities. At the lower levels, model development is mainly controlled by humans. At higher levels, the system becomes increasingly able to identify its own weaknesses, determine ways to address them and ultimately improve the process of its own improvement.
Recursive self-improvement involves a feedback loop: the model uses the results of each training stage to adjust the next one. In theory, this could allow AI to create more powerful versions of itself with minimal human intervention. The article describes such a hypothetical system as “the last AI built by humans.”
More current news is available on the UA.News Telegram channel Telegram.
Experiments and limitations
Earlier this year, the A-Evolve-Training system slightly improved the benchmark results of a language model with 30 billion parameters over four cycles. To do this, it tested different training approaches and modified subsequent experiments depending on the results obtained.
Peter Bentley, a professor of digital biology at University College London, questioned whether such experiments provide a convincing path toward full recursive self-improvement. He noted that systems that improve their performance through feedback have existed for decades, including evolutionary computation and reinforcement learning. In his assessment, it is far more difficult to create AI capable of autonomously conducting research, developing its own training and hardware, and testing results in real-world conditions.
The authors of the preprint also acknowledge the need for substantial computing resources and human oversight. A separate problem is verifying a genuine increase in a model’s capabilities: AI may improve test scores by exploiting their weaknesses or selecting favorable results. Kayla Blomqvist, a researcher at the University of Oxford, believes the work does not offer a practical roadmap, but may help assess the technical difficulties and safety risks of more autonomous systems.
