1. Home
2. World
3. When AI Learns Without Us: Absolute Zero Reasoner, the Machine That Thinks on Its Own

Developed in China, this artificial intelligence trains, improves and solves problems without any human data. A technological breakthrough that raises as many promises as questions.

What if an artificial intelligence no longer needed us to learn? No more relying on our data, our corrections, our examples? That’s the bold challenge taken on by a group of Chinese researchers with the development of Absolute Zero Reasoner (AZR).

This system does not rely on any human database. It learns in complete autonomy, generating its own tasks. It challenges itself, corrects itself and evaluates its own performance. This is what’s known as a self-play loop: a self-training cycle in which the AI simultaneously plays the roles of both student and teacher.

Proposing, Solving, Improving… With No Outside Help

AZR works on a simple yet powerful principle: alternating between generating problems and solving them. On one side, the task proposer creates an exercise — for example, coding a function or solving a complex equation. On the other side, the task solver attempts to answer it. If the answer is correct, a reward is given. If not, the algorithm adjusts, corrects and tries again.

This process is fully autonomous and enables the system to adapt without any human intervention. It doesn't need anyone to teach it how to reason.

Logical, Inductive and Abductive Reasoning

One of the most striking aspects of Absolute Zero Reasoner is its ability to use multiple forms of reasoning traditionally linked to human intelligence: deductive reasoning, to draw logical conclusions from a given set of facts; abductive reasoning, to generate hypotheses from observed outcomes (like in medical diagnostics); and inductive reasoning, to generalize a rule based on examples.

These are the cornerstones of complex thinking, decision-making and innovation. What AZR does, in essence, is not just execute — it thinks.

When the Machine Starts Surprising Us

Even more astonishing: AZR doesn’t just meet expectations — it innovates, sometimes in completely unexpected ways. Researchers have observed emergent behaviors: never seen before strategies, unfamiliar logic and approaches that are sometimes impossible to anticipate. Signs of internal planning have also been detected. Reasoning “outside the box.” A machine that surpasses the intentions of its creators. For experts, this is both fascinating… and unsettling.

Toward an Autonomous and Uncontrollable AI?

We are witnessing a new paradigm. Unlike models like AlphaZero, designed to excel within specific, rule-based environments (like chess), Absolute Zero Reasoner has no boundaries. It’s not limited to a specific game or a domain. It can code, solve, deduce, abstract… and all of this without human data, guidance or supervision.

This raises a critical question: who remains in control? And more importantly: how far can a self-made AI go?

AZR’s creators are clear-eyed: such an advance requires safeguards. Rules. An ethical framework. Because if a machine becomes capable of thinking, creating and reasoning on its own, it also becomes capable of stepping outside the limits we set for it.

We are still far from a science fiction scenario. But we are closer than ever to a truly autonomous artificial intelligence — with all that this implies for research, the economy and society.