Even the most advanced AI models learn by copying humans and solving problems set by their creators. However, recent research suggests that AI can learn in a more human-like way by asking itself insightful questions. A collaboration between Tsinghua University, the Beijing Institute for General Artificial Intelligence (BIGAI), and Pennsylvania State University explored this idea.
The researchers developed a system known as the Absolute Zero Reasoner (AZR). This system generates coding challenges and then solves them, using its successes and failures to improve over time. They tested this approach on a language model called Qwen, with versions of varying sizes (7 billion and 14 billion parameters). Remarkably, the AZR system outperformed some other models trained on curated data.
Andrew Zhao, a PhD student at Tsinghua, explained that their method mirrors human learning. “Initially, you imitate your parents or your teachers, but eventually, you ask your own questions,” he noted. This “self-play” concept isn’t brand new. It has roots in the work of AI pioneers like Jürgen Schmidhuber and Pierre-Yves Oudeyer.
Looking ahead, the researchers are excited about the scalability of their approach. The model can tackle increasingly difficult problems as its abilities grow. However, the current limitation is that it mainly handles problems that can be easily verified, like math or coding tasks.
One intriguing possibility is that this method could allow AI to learn beyond what humans teach. Zheng mentioned that achieving this could potentially lead to superintelligence. Signs of this approach gaining traction can already be seen in projects like Agent0 from Salesforce, Stanford, and the University of North Carolina. It emphasizes self-play to enhance reasoning abilities.
As the tech industry moves forward, finding innovative ways for AI to learn will be crucial. Traditional data sources are becoming limited and costly. With projects like Absolute Zero, we might see AI that’s not just mimicking humans, but genuinely evolving its understanding and capabilities.
For further insights on self-learning AI, you can explore more at Tsinghua University and BIGAI.
Source link
ai lab,artificial intelligence,china,models,research
