<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>ai-local-lab</title>
    <link>https://ai-local-lab.com/</link>
    <atom:link href="https://ai-local-lab.com/rss.xml" rel="self" type="application/rss+xml"/>
    <description>PC・自宅GPU・Jetson・Raspberry PiでローカルAIを動かす人のための実測メディア。機材×モデル×量子化のベンチマークDB、再現手順、購入判断ガイドを実測データで提供します。</description>
    <language>ja</language>
    <lastBuildDate>Tue, 07 Jul 2026 00:00:00 GMT</lastBuildDate>
    <item>
      <title>128GBでローカルLLM【DGX Spark・Ryzen AI Max・Mac Studio】</title>
      <link>https://ai-local-lab.com/articles/dgx-spark-vs-ryzen-ai-max-vs-mac-studio/</link>
      <guid isPermaLink="true">https://ai-local-lab.com/articles/dgx-spark-vs-ryzen-ai-max-vs-mac-studio/</guid>
      <pubDate>Tue, 07 Jul 2026 00:00:00 GMT</pubDate>
      <description>大型MoE時代の新定番「128GBクラス一体機」3択を比較。DGX Spark・Ryzen AI Max+ 395・Mac Studioの帯域・容量・価格を整理し、当サイトのA6000実測からの理論推定で生成速度の目安を示します（未実測を明示）。</description>
    </item>
    <item>
      <title>国産ローカルLLMを実測【地政学リスクと日本語エージェントの現実】</title>
      <link>https://ai-local-lab.com/articles/japanese-local-llm/</link>
      <guid isPermaLink="true">https://ai-local-lab.com/articles/japanese-local-llm/</guid>
      <pubDate>Tue, 07 Jul 2026 00:00:00 GMT</pubDate>
      <description>Sarashina・LLM-jp-4・Swallow・ELYZAをA6000で実測。中国系モデルを使えない組織のための国産・非中国系ローカルLLM選定ガイド。速度・電力に加え、日本語ツール呼び出しの意外な実態を一次データで示します。</description>
    </item>
    <item>
      <title>士業のローカルAI活用【守秘義務と両立する文書AI】</title>
      <link>https://ai-local-lab.com/articles/local-ai-legal-professions/</link>
      <guid isPermaLink="true">https://ai-local-lab.com/articles/local-ai-legal-professions/</guid>
      <pubDate>Tue, 07 Jul 2026 00:00:00 GMT</pubDate>
      <description>弁護士・税理士・司法書士など守秘義務を負う士業では、依頼者情報をクラウドAIに送る運用が悩みどころです。依頼者情報を事務所の外に出さないローカルLLMの構成と使いどころを、日本語FC・過剰拒否の自前実測を基に整理します。</description>
    </item>
    <item>
      <title>MCPでツールを増やすとローカルLLMは壊れるか【5/15/30個を7モデル実測】</title>
      <link>https://ai-local-lab.com/articles/local-ai-mcp-tool-scaling/</link>
      <guid isPermaLink="true">https://ai-local-lab.com/articles/local-ai-mcp-tool-scaling/</guid>
      <pubDate>Tue, 07 Jul 2026 00:00:00 GMT</pubDate>
      <description>MCP時代のエージェントはツール定義を数十個モデルに渡します。「ツールが増えると選択精度が落ちる」は日本語ローカルLLMでも本当か——似た名前のトラップを仕込んだ5/15/30個の3条件×7モデル・378試行をA6000で実測しました。</description>
    </item>
    <item>
      <title>建設・測量の現場でローカルAIを使う【通信が細い現場での実践】</title>
      <link>https://ai-local-lab.com/articles/local-ai-construction-survey/</link>
      <guid isPermaLink="true">https://ai-local-lab.com/articles/local-ai-construction-survey/</guid>
      <pubDate>Fri, 03 Jul 2026 00:00:00 GMT</pubDate>
      <description>山間部・造成中の現場・地下構造物など、通信が不安定な建設・測量現場ではクラウドAIが止まります。現場で動くローカルAI機材の構成と、現場日報の要約・マニュアル検索などの使いどころを、当サイトの実測データを基に整理します。</description>
    </item>
    <item>
      <title>警備・防犯カメラをローカルAIで動かす【映像を外に出さない物体検知】</title>
      <link>https://ai-local-lab.com/articles/local-ai-security-camera/</link>
      <guid isPermaLink="true">https://ai-local-lab.com/articles/local-ai-security-camera/</guid>
      <pubDate>Fri, 03 Jul 2026 00:00:00 GMT</pubDate>
      <description>防犯・警備カメラの物体検知（人物・車両検知）は、映像を外部に送らずローカルで完結させやすいタスクです。Raspberry Pi 5+AI HAT2+（Hailo-10H）の実測（YOLOv8m 76.25 FPS）を基に、低コストな構成の作り方を整理します。</description>
    </item>
    <item>
      <title>TOPS神話の検証【NPUの公称値はローカルLLM速度を予測しない】</title>
      <link>https://ai-local-lab.com/articles/local-ai-tops-myth/</link>
      <guid isPermaLink="true">https://ai-local-lab.com/articles/local-ai-tops-myth/</guid>
      <pubDate>Fri, 03 Jul 2026 00:00:00 GMT</pubDate>
      <description>Hailo-10H（AI HAT2+・40TOPS）でLLMを実際に動かし実測。画像は最大25倍速くなる一方、LLM生成はHailo純正SDKでも4.75 tok/s・2.5WとCPU実行(10.6 tok/s)より遅い結果に。TOPSがLLM速度を予測しない理由をJetson・CPUとの3機材比較で検証します。</description>
    </item>
    <item>
      <title>バーチカルAIにローカルAIをどう組み込むか【ローカル×クラウドの使い分け】</title>
      <link>https://ai-local-lab.com/articles/vertical-ai/</link>
      <guid isPermaLink="true">https://ai-local-lab.com/articles/vertical-ai/</guid>
      <pubDate>Wed, 01 Jul 2026 00:00:00 GMT</pubDate>
      <description>バーチカルAI（業界特化AI）を実装する際、本サイトが扱う「ローカルAI」をどう位置づけるかを整理します。全部をローカルに寄せるのではなく、業務ごとにローカルとクラウドを組み合わせるハイブリッドが基本。全部ローカルで完結させたい場合の実例として、自前のQLoRAケースも紹介します。</description>
    </item>
    <item>
      <title>AIのグラウンディングをローカルで実践【幻覚を抑える3つの接地】</title>
      <link>https://ai-local-lab.com/articles/local-ai-grounding/</link>
      <guid isPermaLink="true">https://ai-local-lab.com/articles/local-ai-grounding/</guid>
      <pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate>
      <description>ローカルAIの幻覚を抑える「グラウンディング（接地）」を実践解説。RAG・ツール接地・出典付けの3手法を、ローカル埋め込み＋オフラインで自分のデータに接地する方法・必要スペック・限界までまとめます。</description>
    </item>
    <item>
      <title>ローカルLLMの生成速度は「メモリ帯域」で決まる【TOPS神話】</title>
      <link>https://ai-local-lab.com/articles/local-llm-memory-bandwidth-bound/</link>
      <guid isPermaLink="true">https://ai-local-lab.com/articles/local-llm-memory-bandwidth-bound/</guid>
      <pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate>
      <description>ローカルLLMの生成速度は演算性能(TOPS)ではなくメモリ帯域でほぼ決まる。自前の機種横断実測と、1000 TOPSのDGX Sparkが70Bで遅い理由を、帯域律速のroofline「載る・動く・使える」で解説します。</description>
    </item>
    <item>
      <title>新世代エッジHWでローカルLLMは速くなるのか【AI PC/NPU/Thor】</title>
      <link>https://ai-local-lab.com/articles/next-gen-edge-hardware/</link>
      <guid isPermaLink="true">https://ai-local-lab.com/articles/next-gen-edge-hardware/</guid>
      <pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate>
      <description>2026-2027の新世代エッジHW（Copilot+ AI PCのNPU・大容量ユニファイドメモリ機・Jetson Thor）でローカルLLMは本当に速くなるのか。TOPSの誇大とメモリ帯域の現実、結局何を買うべきかを実測の視点で整理します。</description>
    </item>
    <item>
      <title>ハザードマップ×オフラインAI：GIS実務から考える防災の構想</title>
      <link>https://ai-local-lab.com/articles/hazardmap-local-ai/</link>
      <guid isPermaLink="true">https://ai-local-lab.com/articles/hazardmap-local-ai/</guid>
      <pubDate>Tue, 23 Jun 2026 00:00:00 GMT</pubDate>
      <description>QGIS・測量・衛星測位(QZSS)の実務知見を防災へ。ハザードマップ(GIS)とオフラインのローカルAIを組み合わせ、通信途絶時でも手元でリスク判断を助ける構想と実装論点を、実例(妖怪ハザードマップ)とともに整理します。</description>
    </item>
    <item>
      <title>通信が途絶しても動く 災害・現場のローカルAIキット</title>
      <link>https://ai-local-lab.com/articles/local-ai-disaster-kit/</link>
      <guid isPermaLink="true">https://ai-local-lab.com/articles/local-ai-disaster-kit/</guid>
      <pubDate>Tue, 23 Jun 2026 00:00:00 GMT</pubDate>
      <description>通信も電力も限られる災害現場・僻地で、手元だけで動くローカルAIキットの構成と選び方。オフライン速度は検証DBの実測、バッテリ駆動時間は実測電力からの理論推定（実機実測は手順付きでこれから）。南極・測量現場の経験から。</description>
    </item>
    <item>
      <title>ローカルAIでループエンジニアリング【3機種で総実時間を実測】</title>
      <link>https://ai-local-lab.com/articles/local-ai-loop-engineering/</link>
      <guid isPermaLink="true">https://ai-local-lab.com/articles/local-ai-loop-engineering/</guid>
      <pubDate>Tue, 23 Jun 2026 00:00:00 GMT</pubDate>
      <description>エージェントの「ループ＝足場」をローカルで設計・運用する実践ガイド。同じ多段ループをA6000・RTX4060ノート・Mac mini M4で実測。成否は機材非依存でも総実時間は機材で激変——暴走ガード・文脈管理・VRAM適合などループ設計の勘所を一次データで示します。</description>
    </item>
    <item>
      <title>ローカルAI FT実例：妖怪ハザードマップQLoRA（Qwen2.5-14B）</title>
      <link>https://ai-local-lab.com/articles/local-ai-finetune-case-youkai/</link>
      <guid isPermaLink="true">https://ai-local-lab.com/articles/local-ai-finetune-case-youkai/</guid>
      <pubDate>Mon, 22 Jun 2026 00:00:00 GMT</pubDate>
      <description>Qwen2.5-14BをQLoRAで14万件学習し、災害リスクを妖怪が説明する構造化JSON出力を99.97%成功させた自前の実例。FTは『振る舞い・形式』に、知識は外部に——という分離の生きたケーススタディです。</description>
    </item>
    <item>
      <title>ローカルAIの最適化の順序：プロンプト→RAG→ファインチューニング</title>
      <link>https://ai-local-lab.com/articles/local-ai-optimization-order/</link>
      <guid isPermaLink="true">https://ai-local-lab.com/articles/local-ai-optimization-order/</guid>
      <pubDate>Mon, 22 Jun 2026 00:00:00 GMT</pubDate>
      <description>ローカルAIを賢くしたい時、いきなりファインチューニングは失敗のもと。プロンプト→RAG→FT→蒸留の順で試すのが鉄則です。FTは知識ではなく『振る舞い・形式』に効く——使い分けと判定フローを整理します。</description>
    </item>
    <item>
      <title>ローカルAI PoCを成功させる完全ガイド【段階導入・実測・業種別】</title>
      <link>https://ai-local-lab.com/articles/local-ai-poc-success/</link>
      <guid isPermaLink="true">https://ai-local-lab.com/articles/local-ai-poc-success/</guid>
      <pubDate>Mon, 22 Jun 2026 00:00:00 GMT</pubDate>
      <description>ローカルAIのPoCは段階的に小さく始め、用途で族を選び、機材に合わせるのが成功の核。自前実測56件（A6000〜Mac〜Jetson〜Pi5）と段階導入・業種別・PoC事例で、投資を最小化しつつ失敗を避ける進め方を示します。</description>
    </item>
    <item>
      <title>ローカルLLMを開発国×得意分野で選ぶ【A6000実測・日本語評価】</title>
      <link>https://ai-local-lab.com/articles/local-llm-by-origin-and-strength/</link>
      <guid isPermaLink="true">https://ai-local-lab.com/articles/local-llm-by-origin-and-strength/</guid>
      <pubDate>Mon, 22 Jun 2026 00:00:00 GMT</pubDate>
      <description>中国勢が席巻するローカルLLM。仏Mistral・UAE Falcon・EU TeukenなどA6000で自前実測し、開発国ごとの違い・実測速度/電力・日本語ファンクションコール能力・VRAM/RAM/量子化の関係まで横断比較します。</description>
    </item>
    <item>
      <title>AIでSSDが壊れる：書き込み寿命の罠とHDD階層化【実体験】</title>
      <link>https://ai-local-lab.com/articles/ai-storage-ssd-endurance/</link>
      <guid isPermaLink="true">https://ai-local-lab.com/articles/ai-storage-ssd-endurance/</guid>
      <pubDate>Sat, 20 Jun 2026 00:00:00 GMT</pubDate>
      <description>AI学習・生成はSSDを想像以上に酷使します。書き込み寿命（TBW）で“書き潰れる”仕組み、実機を数日で壊した運営者の実体験、そして外付けHDDで階層化してSSDを守る現実的な構成を、出典付きで整理します。</description>
    </item>
    <item>
      <title>GLM-5.2はローカルで動くか：必要スペックを試算【最新ニュース】</title>
      <link>https://ai-local-lab.com/articles/glm-5-2-local-requirements/</link>
      <guid isPermaLink="true">https://ai-local-lab.com/articles/glm-5-2-local-requirements/</guid>
      <pubDate>Sat, 20 Jun 2026 00:00:00 GMT</pubDate>
      <description>Z.aiが2026年6月に公開した753BのオープンモデルGLM-5.2。MITで自由に使えますが、自宅やワークステーションで動かすには何が要るのか。必要メモリを量子化別に試算し、現実的な選択肢を正直に整理します。</description>
    </item>
  </channel>
</rss>
