AI models are starting to learn by asking themselves questions

Even the smartest artificial intelligence Models are basically imitations. They learn either by consuming examples of human action or by trying to solve problems assigned to them by human trainers.

But perhaps AI can, in fact, learn in a more human way, by discovering interesting questions to ask itself and trying to find the right answer. Project from Tsinghua Universitythe Beijing Institute of General Artificial Intelligence (BIGAI) and Pennsylvania State University have shown that artificial intelligence can learn to think this way by playing with computer codes.

The researchers created a system called Absolute zero mind (AZR) which first uses a large language model to generate difficult but solvable Python coding problems. It then uses the same model to solve those problems before verifying its operation by trying to run the code. Finally, the AZR system uses successes and failures as a signal to improve the original model, increasing its ability to pose and solve better problems.

The team found that their approach significantly improved coding and reasoning skills for both 7 billion and 14 billion parameter versions of the program. Open source language model Coin. Impressively, the model outperformed some models that received human-curated data.

I talked to Andrew Chowa doctoral student at Tsinghua University who came up with the original idea of ​​absolute zero, as well Zilong Chengwho is a researcher at BIGAI and worked with him on the project via Zoom.

This approach is similar to the way human learning goes beyond rote memorization or imitation, Chow told me. “At first you imitate your parents and act like your teachers, but then you have to ask your own questions,” he said. “And in the end you can outperform those who taught you in school.”

Zhao and Zheng noted that the idea of ​​AI learning in this way, sometimes called “self-playing,” goes back years, and has been previously explored by the likes of Jürgen Schmidhuberone of the well-known pioneers of artificial intelligence, and Pierre-Yves Audiera computer scientist at Inria in France.

One of the most exciting elements of the project, according to Zheng, is the way the model’s problem-posing and problem-solving skills are measured. “The difficulty level grows as the model becomes more robust,” he says.

The main challenge is that the system currently only works on problems that can be easily verified, such as those involving mathematics or programming. As the project progresses, it may be possible to use it for agent AI tasks such as browsing the web or doing office work. This might involve the AI ​​model trying to judge whether the agent’s actions are correct.

One fascinating possibility for an approach like absolute zero is that it is possible, in theory, to allow models to go further than human teaching. “Once we have that, this becomes a way to reach superintelligence,” Li Cheng said.

There are early signs that the absolute zero approach is starting to catch on in some large AI labs.

Project called Agent0from Salesforce, Stanford, and the University of North Carolina at Chapel Hill, involves an agent using a software tool that improves itself through self-play. As with absolute zero, the model improves in general reasoning through experimental problem solving. A Recent paper Written by researchers from Meta, the University of Illinois, and Carnegie Mellon University, it presents a system that uses a similar type of self-play in software engineering. The authors of this work suggest that it represents “a first step toward training models for superintelligent software agents.”

Finding new ways to learn AI is likely to be a big topic in the tech industry this year. As traditional sources of data become scarcer and more expensive, and as laboratories look for new ways to make models more capable, a project like Absolute Zero could lead to AI systems that are less like mimics and more like humans.