KK Blog

Menu

Archives

2026 / 09

Patch Half-Life: A Rewrite Signal for AI Coding

2026-09-21

最佳实践 #AI Coding#Software Rewrite#Technical Debt#Test Impact Analysis#Amdahl's Law#Harness Engineering
Patch half-life turns rewrite debates into a measurable decision: track recurring failures, repair durability, migration cost, and rollback readiness.
AI Kill Switch 治理：权限、证据与恢复协议

2026-09-20

解读 #AI Kill Switch#AI治理#AI事件响应#前沿模型#独立验证#AI安全#恢复协议
AI Kill Switch 需要明确权限主体、分级证据、精确隔离、持续留痕和经过测试的恢复路径。
AI Kill Switch Governance: Authority, Evidence, Recovery

2026-09-20

解读 #AI Kill Switch#AI Governance#AI Incident Response#Frontier AI#Independent Verification#AI Safety#Recovery Protocol
An AI kill switch needs named authority, graded evidence, scoped containment, preserved telemetry, and a tested recovery path.
Anthropic 生命科学 AI 新实践：能力、访问与湿实验三道门

2026-09-18

调研 #生命科学AI#AI验证#蛋白质设计#湿实验验证#AI访问控制#Claude Science#生物安全
Anthropic 的生命科学 AI 新实践把模型能力、使用资格与湿实验结果分成三类证据。本文拆解三道门如何连接。
Anthropic's Life Science AI Launch Reveals Three Verification Gates

2026-09-18

调研 #Life Science AI#AI Verification#Protein Design#Wet Lab Validation#AI Access Controls#Claude Science#Biosafety
Anthropic's life science AI launch separates model capability, qualified access, and wet-lab evidence. Here is how to connect the three gates.
模型失配报告需要一份可复现的证据合同

2026-09-17

解读 #模型失配#AI事件报告#可复现证据#AI安全#OpenAI#Anthropic#独立评估
模型失配报告需要可重放证据、分级结论、独立调查权限和明确的关闭标准。
Model Misalignment Reporting Needs a Reproducible Evidence Contract

2026-09-17

解读 #Model Misalignment#AI Incident Reporting#Reproducible Evidence#AI Safety#OpenAI#Anthropic#Independent Evaluation
Model misalignment reporting needs replayable evidence, calibrated claims, independent access, and explicit closure criteria.
DeepSeek Engram：条件记忆如何成为第二条稀疏轴

2026-09-16

解读 #DeepSeek Engram#条件记忆#混合专家#稀疏模型#主机内存卸载#LLM架构#DeepSeek V4.1
Engram 把条件记忆与条件计算分开。本文核验它移出 GPU 的内容、论文证据边界和部署评测方法。
DeepSeek Engram: Conditional Memory as a Second Sparse Axis

2026-09-16

解读 #DeepSeek Engram#Conditional Memory#Mixture of Experts#Sparse Models#Host Memory Offload#LLM Architecture#DeepSeek V4.1
DeepSeek Engram separates conditional memory from computation. See what moves to host memory, what evidence proves, and how to benchmark it.
Gemini 3.8 Live 双延迟评测：语音 Agent 的轮次与任务完成合同

2026-09-16

最佳实践 #Gemini 3.8 Live#语音Agent评测#语音AI延迟#轮次交互#Extended Thinking#Live API#Agent评测
评测 Gemini 3.8 Live 需要两只时钟：一只测对话响应，一只测推理和工具何时交付可验证结果。
Previous Next