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

Unverified50% confidenceTradeoffExact time

ReAct的Thought→Action→Observation循环以append-only方式增长，导致上下文线性膨胀

1

Sources

50%

Confidence

Long-term

Relevance

7/8/2026

First Seen

Sources

Context Warp Drive：AI智能体确定性上下文折叠详解

hackernewshackernews7/7/2026

Related Claims

Unverified扩散模型的迭代步数与输出长度不呈线性正相关，耗时增长比自回归模型平缓71% similar Unverified当链式法则连乘的因子持续小于1时梯度指数衰减（消失），持续大于1时指数增长（爆炸）67% similar Unverified新模型在实测中偶尔会陷入复读循环，同一句话连续重复多次，稳定性仍有提升空间64% similar Unverified理论分析表明，在时间步长为T的轨迹上，行为克隆的累积误差以O(T²)的速度增长64% similar Unverified随着对话轮次增加，每轮需携带完整历史上下文，输入token数量持续膨胀，导致推理时间呈现非线性增长63% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/303754

API

curl https://kongchang.com/api/v1/knowledge/claims/303754

MCP

get_claim(id=303754)