AI News

Ban Yue Tan: Don't Let One-Sided Comparisons Derail AI Discussion

Shanghai ObserverUpdated Nov 16 18:28

AI summary

In response to the claim that "more than half of China's computing power is being used to make short dramas," a Doubao public relations head said on October 19 that the data is "somewhat unreliable," and that only about 10% of ByteDance's total computing power is used for training and inference of video generation models. Ban Yue Tan pointed out that token consumption does not equal computing power consumption, and that the two sets of data circulating online from China and the United States differ in statistical boundaries, model types, and samples, so simple comparisons distort the picture. The article also noted that structural imbalances in computing power resources objectively exist, and that a mechanism to make research computing power broadly accessible should be improved, along with a unified statistical standard that distinguishes training from inference and distinguishes application scenarios.

Body · AI translation

Recently, the claim that "more than half of China's computing power is used to make short dramas" has circulated online. Some have used this to compare China-US token consumption, concluding that "China feeds short dramas, the US forges the future," and even asserting that China's AI development has "gone off track." Tracing this back, the claim originated from a statement by a listed company during an investor exchange, whose original meaning was that 55% of domestic AI token consumption is used for "short dramas and video generation." On October 19, Doubao's PR head posted a response, saying this figure is "somewhat unreliable" and that, for ByteDance, only about 10% of total computing power is used for training and inference of video generation models. Clearly, making judgments based on one-sided numbers is itself not rigorous enough. Artificial intelligence is a strategic high ground in technological competition, and this discussion reflects the industry's concern about the allocation of computing power resources. Precisely for this reason, the discussion needs to be built on accurate concepts. Token consumption does not equal computing power consumption. A token is a unit of measurement for information processed by a model, reflecting invocation volume, while different models may require vastly different amounts of computation to process the same number of tokens. Video generation also does not equal short drama production. Training and inference distinguish technical stages, while scientific research and entertainment distinguish application purposes; the two divisions cannot be conflated, and "entertainment applications occupying inference computing power" does not lead to "scientific research lacking training computing power." The two sets of China-US data circulating online differ in statistical boundaries, model types, and samples, so simple comparisons will be distorted. Even if a certain type of application accounts for a relatively high share, that is not enough to judge the development direction of the entire industry. The binary oppositional narrative of "one good, one bad" is an important reason such views spread quickly. The complex computing power structure and differences in industrial development stages have a high threshold for understanding; simplified China-US comparisons are more likely to stir emotions and attract traffic. In the process of dissemination, key qualifying conditions are constantly lost, one company's statement is treated as national data, and the discussion easily drifts away from the issue itself. Some self-media then amplify partial phenomena into overall judgments, spawning claims such as "domestic AI only does entertainment and has no future" and "capital only seeks quick money and does not pursue hard technology," which can easily mislead public perception. It should be recognized that consumer scenarios themselves have industrial value, and the consumer-grade AIGC track, such as AI short dramas, cannot be wholly dismissed. While exploring commercialization, AI short dramas provide computing power clusters with opportunities to train in real scenarios and are also expected to help domestic inference chips achieve large-scale deployment. The laws of technological development show that many frontier technologies mature by starting from consumer scenarios close to the public, honing stability and optimizing scheduling capabilities in massive real-world business, and continuously accumulating engineering experience. Of course, the issues raised in this discussion, such as "emphasizing application deployment and neglecting underlying breakthroughs," also deserve the industry's attention. Although it is the mainstream for Chinese AI companies to focus on tackling underlying fundamental challenges and this must not be smeared, some market capital naturally favors short-term monetization projects, some basic research teams face high costs and long waiting periods for obtaining high-end computing power, and structural contradictions in computing power resources objectively exist. Solving the problem cannot rely on simply negating the application track; rather, market vitality and strategic orientation must be coordinated. On the one hand, a inclusive mechanism for scientific research computing power should be improved, special computing power quotas should be set up, and the threshold for using computing power in scientific research and the real economy should be lowered; on the other hand, policies and industrial funds should be used to guide capital toward tackling underlying technologies, the evaluation system for AI companies should be improved so that success is not judged solely by short-term traffic, and consumer AIGC should be guided toward quality upgrading and reduced low-level involution. In addition, the fact that the debate has featured people talking past each other also exposes the lack of authoritative statistics on the usage structure of computing power and tokens by scenario. Only by establishing unified standards that distinguish training from inference and distinguish application scenarios can coordinated scheduling have a basis. AI can both produce content products that enrich public cultural life and take root in industrial production and basic scientific research to tackle hard-core problems. When facing computing power topics, we should maintain self-reflection and shore up shortcomings in underlying innovation, while also refusing to be led by distorted data and traffic narratives, and view the diverse practices of China's AI industry objectively.

Source

Shanghai Observer