Can AI Really Learn Like Humans Do?
So, you know how AI is always trying to mimic us, right? Like, it’s always trying to learn and improve by imitating what we do. But what if I told you that a team of researchers has come up with a way to make AI learn more like we do? Like, actually ask questions and seek answers? I’m talking about a system called Absolute Zero Reasoner (AZR), and it’s like a game of “20 Questions” but with code.
AZR uses a large language model to generate complex coding problems, and then uses the same model to try to solve those problems. And then it checks its own work by trying to run the code. It’s like, this feedback loop where the AI is constantly refining its abilities to come up with better questions and solutions.
In experiments, the team found that AZR improved the coding and reasoning abilities of some pretty big language models, including Qwen. And get this – AZR outperformed some other models that had been given human-curated data. Like, whoa.
I got to chat with Andrew Zhao, the PhD student who came up with the idea, and Zilong Zheng, the researcher who worked on the project. “To learn, humans don’t just imitate their parents or teachers, they ask questions and explore,” Zhao said. “We wanted to replicate that process with AI.”
So, what’s the ultimate goal? To create a system that can learn without human intervention. “Once we have it, it’s like having superintelligence,” Zheng said.
For now, Absolute Zero is limited to math and coding problems, but the researchers think it could be scaled up to more complex tasks like searching the web or completing office tasks.
And other AI labs are already on board. Salesforce, Stanford, and the University of North Carolina at Chapel Hill are working on a project called Agent0, which uses a similar self-play approach.
As AI continues to evolve, finding new ways for machines to learn is going to become more and more important. Traditional sources of data are getting scarcer and more expensive, so researchers are looking for innovative ways to make models more capable. And a challenge like Absolute Zero could lead to AI systems that are less like copycats and more like humans.
You can read the full article on Wired to learn more about this wild project – [https://www.wired.com/story/ai-models-keep-learning-after-training-research/](https://www.wired.com/story/ai-models-keep-learning-after-training-research/)
