Feed Entities Podcasts Deep Dives Developers About

[KongchangAI]

Real-time AI knowledge base. Every claim is traceable to its source, with data freshness under 24 hours.

Powered by 240+ first-hand AI sources

Knowledge

Feed
Hot Entities
Podcasts

Developers

CLI
MCP Server
REST API
About

© 2026 KongchangAI. The freshest, most accurate AI knowledge

Verified70% confidenceFactExact time

上下文长度与计算量呈近似平方级增长关系，受注意力机制影响

4

Sources

70%

Confidence

Long-term

Relevance

7/12/2026

First Seen

Sources

Claude用量限制收紧：AI无限畅享时代终结了吗？

redditr/AnthropicAi7/9/2026

Related Claims

Unverified标准Attention的显存消耗随序列长度平方增长，Flash Attention将其压缩到线性增长85% similar Verified自注意力机制的计算复杂度随序列长度呈近似二次方增长，序列长度为n时注意力矩阵计算量正比于n²81% similar Unverified自注意力使任意两个位置之间的信息交互只需一步计算，但显存占用随序列长度平方增长79% similar Unverified原始 Self-Attention 的计算量随序列长度呈平方级增长，Anthropic 通过优化位置编码（如 ALiBi）等技术手段拓展了可用上下文78% similar UnverifiedTransformer架构的注意力机制中，序列长度增加时早期token的注意力得分在softmax归一化后被后续token稀释，形成注意力沉没效应76% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/490971

API

curl https://kongchang.com/api/v1/knowledge/claims/490971

MCP

get_claim(id=490971)