AI Models Discover Self-Learning by Asking Themselves Questions for Solutions
AI models are evolving by learning to ask themselves questions to solve coding problems, leading to improved skills and self-learning.
A project from Tsinghua University and other groups created a system called Absolute Zero Reasoner (AZR). This system starts by using a smart language model to create difficult Python coding tasks. It then solves these tasks and checks its work by running the code. When it succeeds or fails, it uses that to get better at asking and solving problems.
The researchers discovered that this new way helped the AI models get better at coding than many models that just learned from humans. They even spoke to Andrew Zhao, a PhD student who helped create Absolute Zero, and Zilong Zheng, a researcher at BIGAI. Zhao explained that this method is like how humans learn—first, you copy what you see; then, you start asking your own questions.
Zheng shared that this project is exciting because the harder the problems are, the better the model gets. Right now, it only works on simple problems, like math and coding. But soon, it might handle more complex tasks, like web browsing or organizing office work.
One interesting idea is that this approach could help AI learn more than we can teach it. This might even lead to superintelligent AI someday! There are signs that other big AI labs are interested in this kind of learning too. For example, Salesforce and Stanford are working on a project like this called Agent0.
As technology changes, finding new ways for AI to learn will become very important. As it gets harder and more expensive to collect data, projects like Absolute Zero could change AI from being just copycats to thinking like humans.