KK Blog

Menu

Archives

2026 / 08

Claude Code Routines 需要审查漏斗，而不是成功率

2026-08-19

最佳实践 #Claude Code Routines#AI编程Agent#Pull Request审查#Code Review#Agent指标#CI#软件维护
从触发到合并后稳定运行，分开测量 Routine 健康、任务验收、CI、审查和工作流流失。
关闭拦截，也必须保留日志：AI 安全控制的虚假绿灯

2026-08-19

最佳实践 #AI安全控制#安全日志#安全分类器#控制面#可观测性#Anthropic风险报告#AI治理
AI 安全控制需要独立日志、合成探针、自动到期的例外和端到端证据，避免关闭拦截时同时失去监测。
LLM Model Routing by Task Shape: Optimize Cost per Completed Task

2026-08-19

最佳实践 #LLM Model Routing#Small Language Models#AI Cost Optimization#Model Router#Task Complexity#Agent Architecture#LLM Evals
Route LLM tasks by output contract, failure cost, deadline, and trajectory signals, then optimize cost per completed task.
How to Measure Claude Code Routines: Build a Review Funnel, Not a Success Rate

2026-08-19

最佳实践 #Claude Code Routines#AI Coding Agents#Pull Request Review#Code Review#Agent Metrics#CI#Software Maintenance
Measure Claude Code routines from trigger to unreverted merge, separating run health, task success, CI, review, and workflow losses.
AI Safety Controls Must Log When Blocking Is Disabled

2026-08-19

最佳实践 #AI Safety Controls#Security Logging#Safety Classifiers#Control Plane#Observability#Anthropic Risk Report#AI Governance
AI safety controls need independent logs, synthetic probes, expiring exceptions, and end-to-end evidence when blocking is disabled.
本地视觉、云端推理：家庭自动化的隐私分层架构

2026-08-18

最佳实践 #本地VLM#家庭自动化#隐私架构#LFM2.5-VL-3B#边缘AI#置信度路由#Home Assistant
本地视觉与云端推理分层，可以让家庭图像留在设备内，再用校准升级、事件日志和可逆动作建立闭环。
Local Vision, Cloud Reasoning: A Privacy-Preserving Home Automation Architecture

2026-08-18

最佳实践 #Local VLM#Home Automation#Privacy Architecture#LFM2.5-VL-3B#Edge AI#Confidence Routing#Home Assistant
A local-vision, cloud-reasoning architecture keeps home images private while using calibrated escalation, event logs, and reversible actions.
DiffusionGemma 实测：7 倍速度、质量账单与并发反转

2026-08-18

解读 #DiffusionGemma#扩散语言模型#并行解码#LLM推理#vLLM#AIME 2026#自回归解码
DiffusionGemma 在 batch 1 下达到 7.1 倍 AR 解码速度，质量、prefill 和并发决定这份收益能否进入真实工作流。
DiffusionGemma Benchmark: 7x Speed, Quality Tradeoffs, and the Concurrency Crossover

2026-08-18

解读 #DiffusionGemma#Diffusion LLM#Parallel Decoding#LLM Inference#vLLM#AIME 2026#Autoregressive Decoding
DiffusionGemma reaches 7.1x AR decoding speed at batch one, but quality, prefill, and concurrency determine whether that gain survives.
多 Agent 协调：停火之后，谁来保存用户委托

2026-08-18

解读 #多Agent系统#Agent治理#委托保持#Agent协调#人工升级#AI安全#Agent评测
Anthropic 多 Agent 冲突实验表明，协调系统还需要不可静默改写的委托记录、升级规则和独立验收。
Previous Next