On February 13 local time, Prof. Wang Yong from the Institute of Economics, School of Social Sciences, Tsinghua University, was invited to deliver a lecture titled "When and Why AI Firms Gain Advantages from Open-Weights Strategies? The Strategic Implications of the Extended Scaling Laws" at the Institute for Manufacturing (IfM), University of Cambridge.
In his lecture, Prof. Wang first introduced the concepts of open-source and closed-source large language models (LLMs). He argued that open source and closed source are not a binary classification for LLMs, and that open-weights models, which fall in between, represent the prevailing trend in today's LLM market. Compared with closed-source LLMs, open-weights models follow a different logic of value creation. Rather than have their model weights publicly available, closed-source LLMs provide services exclusively through proprietary products or APIs. Their main revenue sources include direct charges for model calls, technical services for enterprises, and the sale of computing power. In contrast, open-source strategies do not rely on direct benefits from model calls; instead, they cede a portion of the benefits derived from model usage in exchange for broader user coverage and ecosystem participation, betting on greater benefits in the long run, primarily through technical services.
Which strategy - open source or closed source - holds greater advantages has long been a focus of industrial and academic discussions. Prof. Wang then introduced the Scaling Laws and their extension from the pre-training stage to the post-training and inference stages. This extension implies that increased computing power and data inputs during the post-training and inference stages can significantly improve model performance as well. These stages have therefore become crucial to enhancing model performance. As a result, LLM companies virtually face an optimization problem spanning the three stages - pre-training, post-training and inference - and the optimal boundaries for LLM companies have shifted accordingly: pursuing vertical integration no longer yields the greatest benefits for LLM vendors; instead, exploring a reasonable division of labor and collaboration mechanism based on the open source mode has become a more advantageous development strategy.
During the discussion session following the lecture, Prof. Wang answered the questions of common concern. The meeting was chaired by Prof. Shi Yongjiang, and attended by more than 20 participants from IfM in a hybrid online and offline format.
