KK Blog

Menu

Archives

2026 / 10

Muse Spark Open Hardware Requirements: Why 42B Active Parameters Still Need an 816B Deployment Plan

2026-10-19

最佳实践 #Muse Spark Open#Meta AI#MoE#Local LLM#Mac Studio#MLX#M5 Ultra#AI Infrastructure
Muse Spark Open computes with 42B active parameters but must keep an 816B model available. This guide budgets weights, KV cache, memory, and topology.
Hybrid Reasoning Agents: Separating Latent Thinking from Training, Harness, and Benchmark Effects

2026-10-19

解读 #HybridAgent#Latent Reasoning#AI Agents#Hybrid Reasoning#Agent Evaluation#Chain-of-Thought#Tool Use
HybridAgent reports higher quality and faster inference. A causal audit separates latent thinking from training, harness, and benchmark effects.
Harness² Self-Improving Agents: An Audit for Overfitting, Leakage, Cost, and Rollback

2026-10-19

最佳实践 #Harness²#Self-Improving Agents#Agent Harness#Benchmark Leakage#Overfitting#Rollback#AI Evaluation
Harness² improves frozen models by editing the agent harness. Production use needs hidden evaluations, leakage controls, cost budgets, and rollback.
SynthID 规模化落地：为水印验证吞吐设计完整架构

2026-10-18

调研 #SynthID#AI水印#内容溯源#验证吞吐#Google DeepMind#C2PA#AI安全
SynthID 的规模化能力取决于验证端：版本路由、不确定态、审计证据、人工复核和每个可辩护结论的真实成本。
SynthID at Scale: Design the Watermark Verification Throughput Architecture

2026-10-18

调研 #SynthID#AI Watermarking#Content Provenance#Verification Throughput#Google DeepMind#C2PA#AI Safety
SynthID scale depends on verification throughput: version routing, uncertainty, audit evidence, review queues, and cost per defensible verdict.
Claude Code 代码审查：把审查意见变成可持续的回归控制

2026-10-18

最佳实践 #Claude Code#代码审查#回归测试#AI编程Agent#CI#软件质量#Anthropic
把 Claude Code 的审查意见转成可复现测试、静态规则、评测集与风险门禁，让同类错误无法静默复发。
Claude Code Code Review: Turn Findings into Durable Regression Controls

2026-10-18

最佳实践 #Claude Code#Code Review#Regression Testing#AI Coding Agents#CI#Software Quality#Anthropic
Use Claude Code review findings to build reproducible tests, lint rules, evals, and risk gates that prevent the same failures from returning.
Claude 设计量子电路：从优化收益到独立验证，需要一套成本账本

2026-10-17

解读 #Claude#量子电路#AI科研#Toffoli门#量子计算#可验证性#算法设计
Claude 优化了八类量子电路任务。本文用成本与验证账本拆开 TQ 收益、失败预算、搜索成本、硬件映射和复现状态。
Claude-Designed Quantum Circuits Need a Cost-to-Verification Ledger

2026-10-17

解读 #Claude#Quantum Circuits#AI for Science#Toffoli Gates#Quantum Computing#Verification#Algorithm Design
Claude improved eight quantum circuit tasks. A cost-to-verification ledger shows what TQ gains prove, where costs move, and reproduction gaps.
Claude 追溯 ECDSA 曲线起源：把历史归因与密码学发现分开验证

2026-10-17

解读 #Claude#ECDSA#NIST曲线#密码学#AI科学发现#可复现性#安全研究
Claude 将 NIST 曲线 seed 追溯到时间戳驱动的 lrand48。本文用五级证据栈拆开可复算机制、历史归因与作者意图。
Previous Next