Researchers have developed a new system that allows AI models to train independently by asking themselves questions. The approach could pave the way to artificial superintelligence.
Current AI models mostly learn by being fed data and mimicking human intelligence by recognizing patterns and probabilities. A new research project from Tsinghua University could fundamentally change this approach by teaching AI to ask itself questions in order to learn.
The system, called Absolute Zero Reasoner (AZR), uses a language model to independently generate challenging programming problems in Python and then solve them itself. AZR executes the code independently in order to learn directly from successes or its own failures.
Through this technical cycle, the model should be able to refine its skills both in setting tasks and solving them. The researchers found that models with seven billion and 14 billion parameters massively improved their performance using this method.
Artificial curiosity: AI trains itself
The idea of artificial curiosity is not new, but rather takes up concepts from pioneers such as Jürgen Schmidhuber and Pierre-Yves Oudeyer, who explored the potential of “self-play” early on. With its approach, the Absolute Zero Reasoner even outperforms systems that were trained with data sets that were carefully curated by humans.
This progress demonstrates the effectiveness of autonomous learning methods for the future development of intelligent systems. The difficulty of the tasks increases in parallel with the performance of the model.
According to the researchers involved, this scaling effect could pave the way to a future superintelligence that goes beyond the knowledge of human teachers. In daily practice, AI could, for example, take on complex office work or independently carry out more in-depth research on the Internet.
Digital agents of the future
The new process is becoming increasingly important in the tech industry as traditional data sources for training new models are becoming increasingly scarce and expensive. A project called Agento from Salesforce is already using similar principles to strengthen the general reasoning skills of its digital agents through experimental problem solving.
Researchers at Meta are also developing systems that use self-play for software engineering, thereby creating the basis for highly talented software agents. The move away from simply copying human templates represents a turning point in AI development.
Instead of simply reproducing existing knowledge, the system discovers new solutions by experimenting with code.
Also interesting: