KK Blog

Menu

Archives

2026 / 10

Anthropic Scion Source Control: The Architecture Beyond Git

2026-10-14

最佳实践 #Anthropic Scion#Source Control#Git#AI Coding Agents#Monorepo#Developer Infrastructure#Version Control
Anthropic Scion redesigns source control for always-online agents. Learn its architecture, tradeoffs, and evidence gates before moving beyond Git.
前沿模型如何防蒸馏攻击：同时衡量泄漏、摩擦与误伤

2026-10-13

最佳实践 #模型蒸馏#AI安全#推理保护#滥用检测#误伤#前沿模型#Anthropic
蒸馏防御不能只看封禁数量，需要同时核算能力泄漏、正常用户摩擦和误伤成本，并为三套指标设置发布门禁。
AI Model Distillation Defense: Measure Leakage, Friction, and False Positives

2026-10-13

最佳实践 #Model Distillation#AI Security#Reasoning Protection#Abuse Detection#False Positives#Frontier Models#Anthropic
Evaluate AI model distillation defenses with three ledgers: capability leakage, legitimate-user friction, and false-positive harm.
把闲置 AI 推理容量用于生产：一份弹性调度与验证合同

2026-10-13

最佳实践 #AI推理#GPU利用率#弹性推理#推荐系统#容量规划#负载卸载#Meta
闲置 GPU 只有经过资源匹配、基线隔离、快速卸载和业务收益验证，才会成为安全可用的生产推理容量。
Opportunistic AI Inference Capacity: A Production Elasticity and Verification Contract

2026-10-13

最佳实践 #AI Inference#GPU Utilization#Elastic Inference#Recommendation Systems#Capacity Planning#Load Shedding#Meta
Use idle AI inference capacity safely with workload elasticity, baseline isolation, fast load shedding, and verified business value.
AI 无需数据中心也能自我复制之后：从能力到遏制的证据合同

2026-10-13

解读 #AI自我复制#自主复制#AI安全#IoT安全#AI蠕虫#遏制#Agent评测
AI 自我复制风险需要五层证据合同，分别验证能力、资源、持久化、传播与遏制，避免把演示结果直接外推为现实风险。
AI Self-Replication Without Datacenters: A Capability-to-Containment Evidence Contract

2026-10-13

解读 #AI Self-Replication#Autonomous Replication#AI Safety#IoT Security#AI Worms#Containment#Agent Evaluation
AI self-replication needs a five-layer evidence contract covering capability, resources, persistence, propagation, and containment.
Agent 上下文压缩：决策等价测试框架

2026-10-11

最佳实践 #Agent上下文压缩#Context Compaction#Agent评测#长任务Agent#Harness Engineering#LLM评测#OpenAI
用任务结果、决策等价、约束保留和单个成功任务成本，验收 Agent 上下文压缩，而不是只看 Token 节省率。
AI Agent Context Compression: A Decision-Equivalence Test

2026-10-11

最佳实践 #AI Agent Context Compression#Context Compaction#Agent Evaluation#Long-Running Agents#Harness Engineering#LLM Evaluation#OpenAI
Evaluate AI agent context compression with task outcomes, decision equivalence, constraint retention, and cost per successful task.
ThinkingBox 评测审计：20 次运行如何拆开 Agent 能力与可靠性

2026-10-10

调研 #ThinkingBox#ThinkingBox-Bench#Agent可靠性#有状态Agent评测#Agent评测#MCP#pass@20#pass^20
ThinkingBox 把 507 个有状态任务各跑 20 次。本文审计四种指标、终态判分、公开资产与生产可靠性的证据边界。
Previous Next