Foreign media commentators believe that when deploying AI, companies are using an easily quantifiable number as an efficiency indicator: token usage. This approach is similar to the early days of real-time strategy game players' obsession with APM (actions per minute), which may seem busier but doesn't necessarily lead to better results.
Token usage may not be more effective
The article mentions that as AI agents become a hot application in enterprises, the number of model calls has increased significantly. Anthropic previously found that the tokens consumed by agent tasks are typically about four times that of ordinary chat, and up to about 15 times that of multi-agent systems.
The author argues that such numbers are easily mistaken for proof of ability, but high consumption does not necessarily indicate more valuable output. On the contrary, the more content generated, the more work is required for subsequent review, verification, and correction.
APM analogy in games
The article uses games like StarCraft as examples, stating that APM (Actions Per Minute) was long considered an indicator of skill level, but later research has not found that high-level players necessarily have higher APMs. The more crucial differences often lie in judgment, attention allocation, and more efficient execution.
The author also mentions that Google DeepMind found during the training of AlphaStar that allowing the system to perform more clicks did not improve performance; instead, it might waste computing power on minor details. In demo matches against professional players, the system had a lower average APM than human players, yet still won.
The ability to review is the constraint.
The article argues that companies should focus less on the quantity of results generated and more on their ability to reliably validate those results. It cites an economics paper by researchers from MIT, the University of Washington, and UCLA, stating that the growth constraint imposed by AI is no longer just intelligence itself, but rather the human validation bandwidth.
Based on this, the author suggests that when deploying AI, enterprises should focus on observability, review processes, and verification systems, rather than simply pursuing higher token consumption. The article also cites a 2025 analysis of the impact of AI on software development, stating that engineers are spending more time reviewing and modifying AI-generated code than waiting for the code to be generated itself.
In the author's view, if companies only optimize easily countable proxy metrics, they may stray from their true goals. For AI systems, the problem is not continuing to increase the amount of data generated, but rather establishing sufficiently robust verification capabilities.
