KK Blog

Menu

Archives

2026 / 08

Multi-Agent Coordination Must Preserve User Mandates

2026-08-18

解读 #Multi-Agent Systems#AI Agent Governance#Goal Preservation#Agent Coordination#Human Oversight#AI Safety#Agent Evaluation
Anthropic's conflict tests show why multi-agent coordination needs mandate invariants, escalation rules, and independent acceptance checks.
再培训能承接 AI 替代吗：56 项随机研究

2026-08-13

调研 #职业再培训#AI岗位替代#随机对照试验#Sector Program#劳动力政策#ITT#劳动经济学
56 项随机研究显示职业再培训平均有效但增益有限，sector program 效果更强，却受到筛选、雇主关系和规模化约束。
Worker Retraining and AI: Evidence from 56 RCTs

2026-08-13

调研 #Worker Retraining#AI Job Displacement#Randomized Trials#Sector Programs#Workforce Policy#Intent-to-Treat#Labor Economics
A review of 56 randomized studies finds modest average gains from worker retraining, stronger sector programs, and major limits for AI displacement.
WhatsApp Scam Alert：可验证的端侧 AI

2026-08-13

最佳实践 #WhatsApp Scam Alert#端侧机器学习#端到端加密#机密联邦分析#透明账本#TEE#差分隐私
WhatsApp Scam Alert 用端侧推理、机密分析、透明账本和失败关闭验证，在端到端加密下建立可审计的诈骗识别链路。
WhatsApp Scam Alert: Verifiable On-Device ML

2026-08-13

最佳实践 #WhatsApp Scam Alert#On-Device ML#End-to-End Encryption#Confidential Federated Analytics#Transparency Ledger#TEE#Differential Privacy
WhatsApp Scam Alert shows how on-device ML, confidential analytics, transparency logs, and fail-closed verification can protect E2EE messaging.
当模型权重被认定为侵权复制品：慕尼黑 Suno 判决到底说了什么

2026-08-12

调研 #Suno#AI 音乐#著作权#模型权重#记忆化#GEMA#德国#生成式 AI
慕尼黑法院认定六首作品可从 Suno 模型中稳定复现。判决关键不在于权重必然是复制件，而在于记忆化证据、可识别输出与属地规则。
When Model Weights Become Copyright Copies: What the Munich Suno Ruling Actually Says

2026-08-12

调研 #Suno#AI Music#Copyright#Model Weights#Memorization#GEMA#Germany#Generative AI
A Munich court found six songs reproducibly contained in Suno models. The ruling turns on memorization, recognizable outputs, and territorial copyrigh
编程语言也是 Agent Harness：为什么 Go 更容易验证 AI 生成代码

2026-08-12

解读 #Go#AI编程#Coding Agent#Agent Harness#软件验证#gofmt#go test#go vet
Go 的编译器、格式化、测试、模块与安全工具组成统一验证表面，让 AI 生成代码的错误更快暴露并形成可执行反馈。
Programming Languages Are Agent Harnesses: Why Go Makes AI-Generated Code Easier to Verify

2026-08-12

解读 #Go#AI Coding Agents#Agent Harness#Software Verification#gofmt#go test#go vet#Developer Productivity
Go helps AI coding agents by turning compiler, formatting, test, module, and security checks into a consistent machine-readable verification surface.
加密推理块也是不记名凭证：31.5 万条解码轨迹暴露了什么

2026-08-12

调研 #加密推理#AI安全#思维链#Agent日志#Prompt Injection#API安全#上下文绑定
一项研究重建了 315320 个客户端推理块。长期教训是加密还需要身份、上下文、有效期、撤销与防重放机制。
Previous Next