Unverified50% confidenceTradeoffExact time
推理成本会随用户量线性乃至超线性增长，而训练是一次性或周期性成本
1
Sources
50%
Confidence
Long-term
Relevance
7/12/2026
First Seen
Sources
AI全栈方法解读：谷歌从芯片到应用的垂直整合战略
rss6/29/2026
Related Claims
UnverifiedIn the training of hundred-billion-parameter large models, inter-node communication overhead can account for 30%–50% of total training time.70% similarUnverified参数化知识固化在模型神经网络权重中，更新代价极高，通常需要数周乃至数月的重新训练；非参数化知识存放在独立外部索引中，可随时增删改查，更新成本接近于零70% similarUnverifiedFine-tuning costs are typically only one-thousandth to one-ten-thousandth of pre-training costs.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/494203API
curl https://kongchang.com/api/v1/knowledge/claims/494203MCP
get_claim(id=494203)