Unverified50% confidenceFactExact time
The Chinchilla scaling law states that the optimal number of training tokens should be proportional to the number of model parameters, meaning that when parameters are doubled, training data and computational cost should at least double as well.
1
Sources
50%
Confidence
Long-term
Relevance
8/17/2026
First Seen
Sources
伽利略平方立方定律：为什么巨人在物理上不可能存在
Related Claims
UnverifiedThe Chinchilla law states that under a fixed compute budget, appropriately reducing model size and increasing training data volume often yields better performance.78% similarVerifiedDeepMind's Chinchilla research (2022) pointed out that optimal training should keep data volume and parameter count growing at roughly a 1:1 ratio75% similarVerifiedIn 2022, DeepMind's Chinchilla paper pointed out that under the same compute budget, model size and training data volume should be scaled up proportionally.73% similarUnverified批次归一化在训练模式使用当前批次统计量，在推理模式使用训练过程中累积的移动平均值71% similarUnverified根据Chinchilla定律及其后续研究，模型性能与训练计算量之间存在幂律关系71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/763864API
curl https://kongchang.com/api/v1/knowledge/claims/763864MCP
get_claim(id=763864)