Most AI progress has rested on huge collections of examples written and labeled by people. Meta’s Dr. Zero skips that step. It trains AI systems that search the web to answer questions, known as search agents, without any human-written training data. Even so, on several standard question-answering tests, Meta reports, the agents beat systems trained the usual way, on human-written questions and answers.
The design is simpler than it sounds. It uses two copies of the same AI model, each with its own job. One copy writes questions that are difficult but solvable, because easy questions stall progress and impossible ones waste computing power. The other copy answers them and earns a reward, the score it learns from, only when its final answer is correct.
Both copies search the web as they work. An earlier study, ReAct, found that AI systems do better when they use tools while reasoning, instead of relying on memory alone. Checking outside information cuts down on made-up answers, and it makes failures easier to trace.
More accurate answers mean fewer manual checks and less rework, so in work that depends on search, problems get resolved faster and labor costs fall. The business measure is the cost of each correct answer, which a test score alone does not show.
Rewarding only correct answers also changes where the human effort goes. Training data is expensive, slow to produce and often a company’s biggest internal bottleneck, because it takes expert time, agreement across teams and quality control. With Dr. Zero, the human job is to define what “correct” means. Checking results, keeping logs and tracking errors become the points where people stay in control.
Meta also published the code, so teams that need repeatable results can inspect how the method works instead of trusting marketing claims. They can test it in a controlled setting before using it widely.
The approach works best when a task has a right or wrong answer, such as factual questions, math, logic and coding. Those results can be scored reliably, which is what the reward depends on. It does not apply to subjective work like brand voice or creative writing, where correctness cannot be scored cleanly.
