KK Blog

Menu

Archives

2026 / 08

OpenAI Astra Critical 网络门槛解析

2026-08-08

解读 #OpenAI Astra#Critical网络能力#Preparedness Framework#AI网络安全#Agentic Coding#AI安全#网络安全评测
OpenAI 正按 Critical 网络风险对待 Astra。本文区分初步评测、保守处置和正式定级，并拆解研发期必须验证的控制。
OpenAI Astra Critical Cyber Threshold Explained

2026-08-08

解读 #OpenAI Astra#Critical Cyber Capabilities#Preparedness Framework#AI Cybersecurity#Agentic Coding#AI Safety#Cyber Evaluations
OpenAI is treating Astra as Critical for cyber safety. Here is what the threshold proves, what remains preliminary, and which controls must follow.
Meta 多阶段广告排序：在延迟约束下扩展用户序列

2026-08-07

最佳实践 #Meta广告排序#多阶段推荐系统#LLaTTE#Scaling Law#用户序列建模#推荐系统#在线推理
Meta 多阶段广告排序把长序列计算移出请求链路，并用迁移效率、延迟和业务结果共同约束扩容。
Meta Multi-Stage Ads Ranking: Scaling User Sequences Under Latency Constraints

2026-08-07

最佳实践 #Meta Ads Ranking#Multi-Stage Recommender Systems#LLaTTE#Scaling Laws#User Sequence Modeling#Recommendation Systems#Online Inference
Meta multi-stage ads ranking separates long user modeling from online scoring. Learn its scaling laws, transfer ratio, and release gates.
Agent 安全钩子需要能力矩阵：先统一证据，再谈策略阻断

2026-08-06

洞察 #AI Agent安全#安全钩子#运行时阻断#Agent可观测性#CEL#端点安全#Numbat
用能力矩阵区分 Agent 安全钩子的可见性、证据质量、激活状态、失败语义与阻断能力。
国家实验室里的 AI：科学 Agent 需要怎样的验证栈

2026-08-06

最佳实践 #科学Agent#国家实验室#科学计算#AI验证#模拟校准#可复现研究#Codex
从参考实现、参数容差、模拟真值、数值诊断到真实数据，建立科学 Agent 的五层验证栈。
"Agent Security Hooks Need a Capability Matrix: Normalize Evidence Before Enforcing Policy"

2026-08-06

洞察 #AI Agent Security#Security Hooks#Runtime Enforcement#Agent Observability#CEL#Endpoint Security#Numbat
A capability matrix for agent security hooks that separates visibility, evidence quality, activation, failure behavior, and enforcement.
"AI for National Laboratories: The Verification Stack Behind Scientific Agents"

2026-08-06

最佳实践 #Scientific AI Agents#National Laboratories#Scientific Computing#AI Verification#Simulation-Based Calibration#Reproducibility#Codex
A five-layer verification stack for scientific AI agents, from reference parity and parameter tolerances to simulation truth and stewardship.
Meta GEM 训练效率：LLM 规模推荐模型需要一套专用优化栈

2026-08-05

最佳实践 #Meta GEM#推荐系统#LLM训练#GPU效率#MFU#Jagged Tensor#分布式训练#AI基础设施
Meta GEM 将端到端 MFU 提升至 20% 至 25%。真正可复用的是围绕真实负载持续识别并转移瓶颈的方法。
Meta GEM Training Efficiency: Why LLM-Scale Recommenders Need a Different Stack

2026-08-05

最佳实践 #Meta GEM#Recommendation Systems#LLM Training#GPU Efficiency#MFU#Jagged Tensors#Distributed Training#AI Infrastructure
Meta GEM training efficiency shows why LLM-scale recommenders need workload-specific kernels, precision, parallelism, memory, and profiling.
Previous Next