When Question Banks Lag Behind Models, AI Starts Setting Its Own Questions: Chinese Team Achieves Data-Level RSI

The most expensive word in Silicon Valley this year is Recursive Self-Improving (RSI), which lets AI participate in iterating its own models, algorithms, data, and R&D processes. Jeff Dean, after leaving Google, founded Discovery Loop to explore AI-automated scientific research. AlphaGo creator David Silver and other top AI researchers are on the same track, forming a consensus: let AI participate in creating the next generation of AI.

But under the consensus lies a thorny problem: the stronger the model, the fewer humans who can set problems, solve them, and verify them for AI. Where does high-quality training data come from? Traditional approaches either wrap the model in a tool shell (like a harness) without changing the underlying training pipeline, or use large models to score and filter, but the evaluation criteria are pre-written by humans, and old benchmarks quickly become obsolete as models improve. Training data is still produced problem-by-problem by humans, and the speed of evolution is bottlenecked by human problem-setting speed. This is one of the core issues for RSI implementation.

The Endless Frontier team, composed of Shanghai Jiao Tong University's School of AI, DP Technology, and Shanghai Algorithm Innovation Research Institute, releases BigBang-V1, the first base model trained natively with RSI. During training, problem-setting, solving, and verification are all handed to AI, continuously generating high-quality self-evolving synthetic data through verifiable frontier tasks, without human involvement in each problem.

The team took a different angle: if the model can self-improve, can the training data production system itself become the object of optimization? The question shifts from 'how to collect more scientific data' to 'how to make the data production system continuously evolve with model capabilities.' Data quality is not cleaned up but determined by task attributes. Tasks must simultaneously satisfy frontier-ness (located at the boundary of knowledge or capability, with no known optimal answer) and verifiability (able to be objectively checked through formal methods, program execution, numerical computation, etc.). If either is missing, data quickly depreciates.

Scientific fields happen to satisfy both conditions simultaneously, naturally combining search, reading, hypothesis, mathematical derivation, code development, tool invocation, experimental analysis, etc., equivalent to a comprehensive course for general intelligence. Jeff Dean says Discovery Loop applies to scientific and engineering fields with measurable goals, and BigBang's choice of science as a training ground shares the same underlying logic.