Tsinghua Professor Tang Jie: Replace Scaling Laws with "Scaling Correctness" to Pave the Way to AGI

Anyone with some understanding of large models knows the "Scaling Law." At the main forum of the 2026 World Laureates Forum held today, Tang Jie, a professor at Tsinghua University and co-founder of Zhipu AI, said in his speech that "scaling correctness" should replace the "Scaling Law" and that a science of model scaling should be explored. In his view, pursuing the path of "scaling correctness" can enable large models to autonomously complete increasingly complex tasks, ultimately leading to AGI (artificial general intelligence). Tang Jie appeared on the 2026 Hurun Rich List released in October this year, with a fortune of about 27.5 billion yuan, and has been called China's "richest professor" by some media. He not only founded the listed company Zhipu AI, but is also the mentor of Zhipu AI CEO Zhang Peng and Moonshot AI founder Yang Zhilin, having trained a group of young talents in the field of artificial intelligence. Professor Tang Jie speaks at the 2026 World Laureates Forum. What insights does this Tsinghua professor have on the future direction of AI research and development? He proposed the concept of "scaling correctness," challenging the "Scaling Law." Since ChatGPT launched in November 2022, large model research and development following the "Scaling Law" has progressed rapidly. This law is an empirical rule meaning that, when other conditions do not constitute a bottleneck, the test loss of model pretraining and the three factors of parameter count, training data volume, and training compute each satisfy a power-law relationship; as parameters, compute, and data are scaled up in tandem, the predicted loss continues to decline smoothly, and model capability improves accordingly. This means that the larger the model's parameter count, the greater the training compute, and the more training data, the higher the model's level of intelligence. However, in Tang Jie's view, the "Scaling Law" only fits training loss and cannot guarantee correctness on real tasks; a decline in training loss does not equal reliable reasoning, runnable code, or long tasks free of hallucinations. Blindly following this law is in fact a lazy path; a systematic methodology should be established to answer "whether conclusions still hold after scaling up." He calls this "scaling correctness" and summarizes this research route as: moving from the "Scaling Law" to a science of scaling. The proposal of this research route is timely: after dialogue large models were upgraded to agents, the cost of trial and error in training increased substantially. A single training run of a frontier large model requires enormous investment; if an algorithm works at small scale but fails when applied to a large model, all the compute and time will be wasted. On the other hand, the real tasks handled by AI are becoming increasingly complex, far beyond the scope that any single benchmark can cover. Against this backdrop, conclusions that hold only under small-scale conditions often cease to be valid after scaling up. Scaling failure is not a single problem but is distributed across six dimensions: optimization and hyperparameters, model architecture, numerical precision, parallelism and systems engineering, data, and post-training. Failure in any one dimension can cause an entire project to fail. Therefore, "scaling correctness" is a systemic issue spanning the entire chain of algorithms, chip clusters, data, and post-training, and does not depend on a single metric. The 2026 World Laureates Forum is held at the Shanghai Lingang Center. Tang Jie pointed out that the core methodological innovation in pursuing "scaling correctness" lies in changing the unit of evaluation. The old paradigm of large model research and development and evaluation was "task-driven, single-point comparison," that is, training models separately for different tasks and looking only at their respective single-point SOTA (state of the art in a given field); such conclusions may well be detached from scale and cannot be extrapolated. The new paradigm is "scaling paradigm, curve comparison," that is, using training loss as the vertical axis and compute, parameter count, and data volume as the horizontal axes to construct a scaling curve, with the goal of accurately predicting the performance of large-scale (trillion-parameter-level) models through small-scale fitting. This new paradigm is of great significance to the emerging "recursive self-improvement" (RSI). "Recursive self-improvement" refers to the process in which an AI system uses the results of its own improvements as input, continuously optimizing its own capabilities and modifying its own architecture to form a self-improvement loop; it is widely regarded in the industry as a key path to achieving artificial general intelligence. Tang Jie believes that whether recursive self-improvement can work does not depend on whether the model can generate the next step, but on whether it can accurately judge whether this step is "progress." He calls this internalized evaluation capability "self-purification." Only when large models possess "self-purification" capability can they iterate autonomously, improve themselves, and ultimately achieve new breakthroughs in intelligence.