KK Blog

Menu

Archives

2026 / 09

ChatGPT Ads Need an Answer-Independence Audit

2026-09-01

最佳实践 #ChatGPT Ads#Answer Independence#AI Governance#Advertising Audit#Algorithmic Auditing#Privacy#OpenAI
A sponsored label identifies the ad but cannot prove that advertising leaves ChatGPT answers unchanged. Here is a repeatable audit contract.
LLM 表现出编码痕迹却仍答不出

2026-09-01

解读 #LLM事实性#参数知识#WikiProfile#LLM评测#知识提取#推理#幻觉
WikiProfile 把被归为未编码的事实，与表现出行为编码证据却难以提取的事实分开，形成更准确的事实错误诊断。
LLM Recall: Encoded Facts That Remain Hard to Retrieve

2026-09-01

解读 #LLM Factuality#Parametric Knowledge#WikiProfile#LLM Evaluation#Retrieval#Reasoning#Hallucination
WikiProfile separates facts classified as not encoded from facts with behavioral encoding evidence that remain hard to recall.

2026 / 08

MCP Agent 身份：哪些已稳定，哪些仍是草案

2026-08-30

解读 #MCP#Agent身份#MCP授权#OAuth 2.1#工作负载身份#ID-JAG#DPoP#企业安全
MCP 已把 Agent 身份列为协议重点。本文区分现行授权、企业扩展、工作负载身份、委托与 DPoP 的真实成熟度。
Thomson Reuters Thomson 1.0: What Its $40M Domain Model Actually Bought

2026-08-30

调研 #Thomson Reuters#Thomson 1.0#Domain-Specific LLM#Enterprise AI#Build vs Buy#Continual Learning#Open Weights#AI Sovereignty
Thomson Reuters spent about $40M on Thomson 1.0. Its reusable data, experts, evaluation, training, and deployment system matters most.
MCP Agent Identity in 2026: What Is Stable, What Is Draft, and What Each Mechanism Proves

2026-08-30

解读 #MCP#Agent Identity#MCP Authorization#OAuth 2.1#Workload Identity#ID-JAG#DPoP#Enterprise Security
MCP agent identity is becoming a protocol priority. This guide separates stable authorization from draft workload identity, delegation, and DPoP work.
自动对齐研究员：Claude 安全增益背后的评测合同

2026-08-29

解读 #自动对齐研究员#AI对齐#AI安全评测#留出测试#奖励作弊#Petri#Claude
Claude 能自动改进十类对齐失败。真正值得复用的成果，是一套能发现过拟合、能力退化、作弊和外推失效的评测合同。
Automated Alignment Researchers Need an Evaluation Contract

2026-08-29

解读 #Automated Alignment Researchers#AI Alignment#AI Safety Benchmarks#Held-Out Evaluation#Reward Hacking#Petri#Claude
Automated alignment researchers need more than strong scores. Anthropic's experiment shows the evaluation contract that makes gains testable.
加州 AB 1651 AI 披露：高风险认证需要两套账

2026-08-27

解读 #加州AB1651#AI披露#律师考试#AI治理#高风险认证#审计链#加州律师协会
加州 AB 1651 提醒我们：日常生产需要结果核验账，高风险认证还需要独立的来源披露账。
California AB 1651 AI Disclosure: Two Ledgers for High-Stakes Certification

2026-08-27

解读 #California AB 1651#AI Disclosure#Bar Exam#AI Governance#High-Stakes Certification#Audit Trail#State Bar of California
California AB 1651 shows why AI governance needs one ledger for verified work and another for disclosure at certification gates.
Previous Next