jietang (@jietang) on X

The article discusses that scaling laws are not only about parameter count but also require balancing data volume, compute allocation, and inference cost. From Kaplan to Chinchilla and then to MoE, the optimal ratio varies with the task and deployment conditions.