企业 AI 的第三条路:Thomson Reuters 的 4000 万美元买了什么

Thomson Reuters 为 Thomson 1.0 投入约 4000 万美元。真正可复用的资产是数据、专家、评测、训练与部署系统。

Administrator Administrator Published on 2026-11-07

AI Memory 的写入精度问题:为什么更强的 Recall 反而会降低 Agent 可靠性

可靠的 Agent 记忆需要写入准入、来源绑定、纠错传播和下游评测。只提高 Recall,可能让错误被更稳定地复用。

Administrator Administrator Published on 2026-11-07

AI Memory Has a Write-Precision Problem: Why Better Recall Can Make Agents Less Reliable

Reliable AI agent memory needs write-time admission, source binding, correction propagation, and downstream evaluation, not recall alone.

Administrator Administrator Published on 2026-11-07

政府 AI Agent 越权后,先冻结两条证据链

面向政府与关键基础设施 Agent 的事故处置方法:分开权限证据与效果证据,用双时钟完成遏制、初报和可验证关闭。

Administrator Administrator Published on 2026-11-04

第三方 AI 评测需要三份合同

第三方 AI 评测的可信度取决于独立性、访问权与披露权三份可执行合同,而不是第三方这个标签。

Administrator Administrator Published on 2026-10-30

AI 搜索引用供应链攻击面

复旦大学对 17211 次引用的研究证明了公开发帖进入 AI 搜索引用的路径,也给出了清晰的因果边界。

Administrator Administrator Published on 2026-10-27

LLM 形式化验证上线评审:三张表拆清证明覆盖范围

Anthropic 用 Claude 为 123 个开源库的 239 个入口点生成证明。上线评审仍要核对证明命题、部署工件和残余风险,避免把有边界的证明误读为整库无漏洞。

Administrator Administrator Published on 2026-10-23

LLM Formal Verification Needs a Proof-Coverage Contract

Anthropic's 123-library experiment shows that LLMs can scale proof writing. A deployment assurance manifest shows what was proved, what ships, and wha

Administrator Administrator Published on 2026-10-23

WhatsApp 隐藏 LLM 指令:检索血缘审计

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。

Administrator Administrator Published on 2026-10-21

WhatsApp Hidden LLM Instructions: A Provenance Audit

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

Administrator Administrator Published on 2026-10-21
Previous Next