KK Blog

Menu

Archives

2026 / 07

数学 Agent 的双循环协议：把开放探索与可信知识分开

2026-07-24

最佳实践 #AI Agent#数学证明#证明验证#多 Agent#AI4Math#Codex#形式化验证
一套可落地的数学 Agent 工作流：用持久状态、敌意审计、盲重构和证据晋升，把开放探索转化为可信知识。
GigaToken 基准审计：1000 倍加速何时真的有效

2026-07-24

最佳实践 #LLM#Tokenization#GigaToken#CPU性能#基准测试#Hugging Face#tiktoken#推理优化
审计 GigaToken 约 1000 倍加速的测试条件、独立复现和兼容风险，并判断 CPU Tokenization 何时真正影响 LLM 吞吐。
AI Agent Sandbox 状态模型：什么会留下

2026-07-23

最佳实践 #AI Agent#Sandbox#沙箱#状态管理#持久化#快照#Agent基础设施#安全
AI Agent Sandbox 的持久化语义取决于生命周期动作。本文用五层状态模型核验文件、内存、快照、外部存储、连接与副作用。
OpenAI Presence: What an Enterprise Agent Platform Must Own

2026-07-23

解读 #OpenAI Presence#Enterprise AI Agents#Agent Governance#AI Operations#Human Escalation#Agent Evaluation#Enterprise AI
OpenAI Presence turns policies, approved actions, simulations, evaluations, escalation, and controlled updates into one operating loop. Here is what t
Mathematical AI Agents: A Dual-Loop Verification Protocol

2026-07-23

最佳实践 #AI Agents#Mathematical Proofs#Proof Verification#Multi-Agent Systems#AI4Math#Codex#Formal Verification
A practical workflow for AI agents in mathematical proofs: durable state, hostile audits, blind reconstruction, and evidence-gated knowledge.
GigaToken Benchmark Audit: Finding the Real CPU Bottleneck

2026-07-23

最佳实践 #LLM#Tokenization#GigaToken#CPU Performance#Benchmarking#Hugging Face#tiktoken#Inference Optimization
An audit of GigaToken's 1000x claim, independent results, compatibility risks, and when CPU tokenization changes LLM throughput.
AI Agent Sandbox Persistence: What Survives Each Lifecycle?

2026-07-23

最佳实践 #AI Agents#Sandbox#Persistence#Snapshots#State Management#Agent Infrastructure#Security#Durable Storage
AI agent sandbox persistence varies by lifecycle action. Use this five-layer model to verify files, memory, volumes, connections, and side effects.
"原地扩展 Tokenizer：不从头训练升级模型"

2026-07-22

最佳实践 #LLM#Tokenizer#BPE#模型迁移#多语言AI#推理性能#预训练模型
"把 Tokenizer 视为 checkpoint ABI，用可验证的迁移契约原地扩展预训练模型词表，并守住生成质量与真实性能。"
开放权重为何成为中国 AI 的产业策略

2026-07-22

洞察 #开放权重#中国AI#Qwen#DeepSeek#Kimi#AI产业#开源AI
开放权重不是单纯公开模型，而是一套产业分发策略。本文用采用、衍生生态、部署成本、硬件覆盖和商业转化五个指标判断它是否有效。
Why Open Weights Became China's AI Distribution Strategy

2026-07-22

洞察 #Open Weights#China AI#Qwen#DeepSeek#Kimi#AI Strategy#Open Source AI
China's open-weight strategy is a distribution system. Evaluate it through adoption, derivatives, deployment cost, hardware reach, and business conver
Previous Next