KK Blog

Menu

Archives

2026 / 10

ThinkingBox Agent Benchmark Audit: What 20 Runs Reveal Beyond Pass@1

2026-10-10

调研 #ThinkingBox#ThinkingBox-Bench#AI Agent Reliability#Stateful Agent Benchmark#Agent Evaluation#MCP#pass@20
ThinkingBox repeats 507 stateful agent tasks 20 times. This audit separates capability, consistency, observed 20/20, and production reliability.
Anthropic OSS Scanner：原始漏洞报告需要一条维护者验证队列

2026-10-09

解读 #Anthropic OSS Scanner#AI漏洞报告#开源安全#漏洞分诊#威胁建模#可复现安全
Anthropic OSS Scanner 把原始 AI 漏洞报告交给维护者。真正的核心是主动加入、复现、去重、威胁模型与风险关闭。
Anthropic OSS Scanner: A Verification Queue for Raw AI Vulnerability Reports

2026-10-09

解读 #Anthropic OSS Scanner#AI Vulnerability Reports#Open Source Security#Vulnerability Triage#Threat Modeling#Reproducible Security
Anthropic OSS Scanner turns raw AI vulnerability reports into a maintainer-controlled queue. Here is the verification contract it needs.
GPT-6 Intelligent UI：动态界面需要一份流式运行时合同

2026-10-08

解读 #GPT-6#Intelligent UI#生成式界面#流式UI#UI编译器#AI Agent#Agent评测#OpenAI
GPT-6 Intelligent UI 让模型边思考边生成界面。真正的生产门槛是把状态、权限、恢复和业务结果写进运行时合同。
GPT-6 Intelligent UI Needs a Streaming Runtime Contract

2026-10-08

解读 #GPT-6#Intelligent UI#Generative UI#Streaming UI#UI Compiler#AI Agents#Agent Evals#OpenAI
GPT-6 Intelligent UI turns model output into a live interface. Production trust requires explicit state, capability, recovery, and outcome contracts.
Agent 审计链里的 Network Time Security：信任时间戳前先验证时间源

2026-10-07

最佳实践 #Network Time Security#NTS#Agent安全#审计日志#NTP#Chrony#RFC 8915
NTS 能认证 NTP 时间包，但 Agent 审计还需要校时健康、时间语义和证据绑定。本文给出一套四层时间合同。
Network Time Security for AI Agents: Verify Time Before You Trust the Audit Trail

2026-10-07

最佳实践 #Network Time Security#NTS#AI Agent Security#Audit Trail#NTP#Chrony#RFC 8915
Network Time Security authenticates NTP, but an AI agent audit trail also needs synchronization health, explicit time semantics, and evidence binding.
DeepSeek 昇腾移植：算清 AI 基础设施兼容性预算

2026-10-06

最佳实践 #DeepSeek#华为昇腾#AI基础设施#DeepGEMM#DeepEP#FlashMLA#硬件兼容
DeepSeek 昇腾移植把兼容边界写进代码：上层 API 可以复用，数据布局、编译器、内核、通信和生产验证仍需单独预算。
DeepSeek Ascend Compatibility: An AI Infrastructure Portability Budget

2026-10-06

最佳实践 #DeepSeek#Huawei Ascend#AI Infrastructure#DeepGEMM#DeepEP#FlashMLA#Hardware Portability
DeepSeek's Ascend ports show where AI infrastructure compatibility ends: reuse APIs, budget for layouts, compilers, kernels, networks, and validation.
Kimi K2.6 推理调优：先测负载，再动模型

2026-10-04

最佳实践 #Kimi K2.6#LLM推理#Coding Agent#KV Cache#投机解码#SGLang#GPU性能
从 Modal 的 Kimi K2.6 案例提炼一套负载优先的推理调优流程：先测请求形态，再依次处理解码、缓存和路由瓶颈。
Previous Next