KK Blog

Menu

Archives

2026 / 09

Astra Scores 62.7% or 99.9% on ARC-AGI-3: The Harness Is Part of the Benchmark

2026-09-05

解读 #GPT-6 Astra#ARC-AGI-3#AI Agent Benchmark#Provider Adapter#Harness Engineering#LLM Evaluation#Context Compaction
Astra scored 62.7% or 99.9% on ARC-AGI-3 under different harnesses. Here is what changed, what the numbers mean, and what remains unproven.
Claude Formalized Fermat's Last Theorem in 11 Days: What the Lean Verification Stack Actually Proves

2026-09-05

解读 #Claude#Fermat's Last Theorem#Lean 4#Formal Verification#Prove2Me#AI Mathematics#Multi-Agent Systems
Claude formalized Fermat's Last Theorem in Lean in 11 days. Here is what the 13-million-line proof and its verification stack establish.
持久 Agent 需要两份合同：什么时候醒，醒来后能做什么

2026-09-05

最佳实践 #持久Agent#长任务Agent#Agent权限#Agent安全#检查点#主动式AI#Agent治理
持久 Agent 需要分别约束唤醒时机与行动权限。一次唤醒只允许观察，任何外部副作用都要重新校验授权。
Persistent AI Agents Need Separate Wake and Action Contracts

2026-09-05

最佳实践 #Persistent AI Agents#Long-Running Agents#AI Agent Permissions#Agent Security#Checkpointing#Proactive AI#Agent Governance
Persistent AI agents need separate contracts for wake timing and action authority. This checklist makes long-running agents safer.
组织第二大脑如何向专家学习，同时避免固化专家错误

2026-09-03

最佳实践 #组织第二大脑#企业知识管理#专家反馈#知识溯源#RAG#AI Agent#回归测试
组织第二大脑需要把专家反馈转化为带来源、版本、验证结果和退役规则的知识变更，才能在扩散经验的同时控制错误。
How to Build an Organizational Second Brain Without Freezing Expert Mistakes

2026-09-03

最佳实践 #Organizational Second Brain#Enterprise Knowledge Management#Expert Feedback#Knowledge Provenance#RAG#AI Agents#Regression Testing
Build an organizational second brain with provenance, versioned knowledge, expert review, replay tests, and safe retirement instead of static RAG.
EvoUndo：给自我修改的 Agent Harness 加一道可恢复性门禁

2026-09-02

解读 #EvoUndo#自我修改Agent#Agent Harness#可恢复性#回滚#AI安全#LLM Agent
EvoUndo 说明能力提升不足以批准 Agent Harness 自我修改。持久变更合并前应通过状态见证、独立影响审计与反事实恢复测试。
EvoUndo: A Recoverability Gate for Self-Modifying AI Agent Harnesses

2026-09-02

解读 #EvoUndo#Self-Modifying Agents#Agent Harness#Recoverability#Rollback#AI Safety#LLM Agents
EvoUndo shows why capability gains are insufficient for self-modifying Agent Harnesses. Require witnessed, counterfactual recovery before merge.
Claude Fable 5.1 与 Mythos 5.1：怎样评测安全护栏成本

2026-09-02

解读 #Claude Fable 5.1#Claude Mythos 5.1#AI安全#LLM评测#Terminal-Bench#模型选型#AI治理
Fable 5.1 与 Mythos 5.1 使用相同权重、不同护栏。模型选型应联合评测能力、干预、可用结果成本与访问资格。
Claude Fable 5.1 vs Mythos 5.1: How to Benchmark the Safeguard Tax

2026-09-02

解读 #Claude Fable 5.1#Claude Mythos 5.1#AI Safety#LLM Benchmarking#Terminal-Bench#Model Evaluation#AI Governance
Fable 5.1 and Mythos 5.1 share model weights but use different safeguards. Measure capability, intervention, cost, and access together.
Previous Next