26 December, according to Tsing Tsing, Associate Professor, Department of Computers, Qinghua UniversityLiu ZhiyuanOn the cover of Nature Machine Intelligence, a study by its team officially proposes Densing Law, which is used to quantify the power density of a large model。
Based on a reflection of the 51 major mainstream models, the study noted that between 2023 and 2025, the intellectual density of the large models accelerated at a rate of double per 3.5 months, meaning that half of the amount of parameters available per 100 days reached the equivalent performance of the current best model, with the associated cost being halved。
Liu Zhiqiang put it bluntly that if a model company “three to six months without cost recovery” were to publish new products, the business model would be difficult to sustain, as the latter would soon be able to achieve equivalent capacity with a quarter of its resources。
"CONSTRUCTION OF AI WITH AI" IS SEEN AS AN INDICATOR OF THE PRODUCTIVITY OF THE AI ERA AND THE DIRECTION OF THE INDUSTRY。
Liu Zhiyuan considers the "density rule" and the "Scaling Law" as "two sides of the coin":
The former emphasizes greater capacity with smaller parameters through continuous innovation in architecture, data governance and learning
The latter depicts the capacity of the expansion in the size of the parameters as increasing。
He noted that after ChatGPT had triggered global inputs, the density doubling cycle had contracted from about five months to about 3.5 months, much faster than the 18-month rhythm of Moore ' s Law. This trend has made cloud-based API services extremely competitive and may end up being left to the head of a large-volume user with a strong technological iterative capacity
At the same time, "end-of-side intelligence" with clear constraints and sensitivity to the time of effort and response will be a more certain window of opportunity for start-ups。
With regard to multi-modular progress, Liu Zhiyuan views Google’s latest release of Gemini 3 as a milestone: the high consistency and control of text in image generation reflects the model’s “several refinement” of the world’s understanding and generation process。
He speculates that this capability is not only dependent on Diffusion, but is also likely to be integrated into a new paradigm of self-regression, thus generating coherence; it also confirms the extension of the density rule - as long as some intellectual ability can be realized, it will be able to operate on smaller terminals, such as mobile phones, PCs or car-mounted chips。
HE'S OPTIMISTIC ABOUT THE LONG-TERM IMPACT OF AI, AND BELIEVES THAT BY 2030-2035 GLOBAL INCLUSION WILL BE ACHIEVED AGITHE SUBJECT OF THE INTERNET WILL NO LONGER BE A HUMAN BEING, BUT WILL ALSO HAVE COUNTLESS INTELLIGENT BODIES; ALTHOUGH THE TRAINING WORKSHOP'S CHAMBERS OF COMMERCE ARE CONSTRICTED, “AGI DEVELOPMENT IS NOT CONDENSED”, THE REASONED NEED WILL GROW EXPLOSIVELY AND HUMAN SYNERGIES WILL BECOME NORMAL。