AI Models Are Starting to Learn by Asking Themselves Questions (www.wired.com)

A collaborative research project involving Tsinghua University, the Beijing Institute for General Artificial Intelligence (BIGAI), and Pennsylvania State University has introduced a groundbreaking system called the Absolute Zero Reasoner (AZR), which allows AI models to learn by posing their own questions. Unlike traditional AI that relies heavily on human-generated data or predefined tasks, AZR uses a large language model to generate and solve coding challenges autonomously, then refines its capabilities based on the outcomes of these attempts. This methodology not only enhances the coding and reasoning skills of AI models with billions of parameters but also enables them to outperform other models trained on human-curated datasets. The significance of this approach lies in its potential to push AI beyond mere imitation and rote learning towards a more independent and human-like way of acquiring knowledge. By adopting a technique known as "self-play," AZR could lead to AI that surpasses human instruction by continuously evolving its problem-posing and solving abilities. While the current implementation focuses on easily verifiable tasks like coding, researchers see potential applications in more complex areas such as agentic AI tasks. As the tech industry seeks innovative learning methods amid diminishing data resources, initiatives like Absolute Zero could pave the way for next-generation AI systems that exhibit more advanced intelligence akin to human reasoning.