Archives

DeepInfer Deterministic Inference: An Output Contract

2026-10-21

解读 #DeepInfer #Deterministic Inference #Batch Invariance #DeepSeek #LLM Inference #Reproducibility #AI Infrastructure #Inference Evaluation

DeepInfer makes determinism a testable inference feature. Audit batch invariance, seed conditions, hardware scope, and serving gaps before deployment.

WhatsApp 隐藏 LLM 指令：检索血缘审计

2026-10-21

调研 #WhatsApp #检索血缘 #间接提示注入 #RAG安全 #AI供应链 #模型可见内容 #AI治理

WhatsApp 隐藏指令事件说明，官方来源的事实可信度与指令权限必须分开，RAG 系统需要审计模型实际看到的内容。

WhatsApp Hidden LLM Instructions: A Provenance Audit

2026-10-21

调研 #WhatsApp #Retrieval Provenance #Indirect Prompt Injection #RAG Security #AI Supply Chain #Model-Visible Content #AI Governance

WhatsApp's hidden chatbot instructions show why RAG systems must separate source trust from instruction authority and audit model-visible content.

低资源编程语言需要迁移评测，而不只是代码翻译演示

2026-10-20

解读 #遗留系统迁移 #低资源编程语言 #COBOL现代化 #AI编程 #软件评测 #行为等价 #MultiLangBench

OpenAI 的低资源编程语言评测把问题从代码翻译推进到行为保留。生产迁移还需覆盖依赖、故障、影子运行、切流和回滚。

Legacy Code Migration Evals: Beyond Low-Resource Language Translation

2026-10-20

解读 #Legacy Code Migration #Low-Resource Programming Languages #COBOL Modernization #AI Coding #Software Evaluation #Behavioral Equivalence #MultiLangBench

Legacy code migration evals must test behavioral equivalence, dependencies, failure modes, and rollback. OpenAI's new benchmarks are a starting point.

薄编排器模式：让 AI Agent 路由可观测、可恢复

2026-10-20

最佳实践 #薄编排器 #AI Agent架构 #多Agent系统 #Agent路由 #可靠性 #上下文隔离 #Agent可观测性

薄编排器把路由、状态与门禁留在确定性控制平面，把领域推理交给专门 Agent。本文给出职责边界、状态机、恢复矩阵和迁移清单。

Thin Orchestrator Pattern: Reliable AI Agent Routing, Recovery, and Observability

2026-10-20

最佳实践 #Thin Orchestrator #AI Agent Architecture #Multi-Agent Systems #Agent Routing #Reliability #Context Isolation #Agent Observability

The thin orchestrator pattern keeps AI agent routing deterministic while specialists reason. Use this contract for state, retries, and recovery.

Muse Spark Open 硬件需求：42B 激活参数为何仍需 816B 部署方案

2026-10-19

最佳实践 #Muse Spark Open #Meta AI #MoE #本地大模型 #Mac Studio #MLX #M5 Ultra #AI基础设施

Muse Spark Open 每个 Token 激活 42B 参数，但部署仍要容纳 816B 模型。本文拆解权重、KV Cache、内存和双机拓扑。