<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>AIニュース｜株式会社さいとうテックプラス</title>
    <link>https://saitotechplus.com/pages/ai-news.html</link>
    <description>海外で発表されたAIの動きを毎日自動で収集し、日本の開発者・IT企業に向けて日本語で要約・解説しています。</description>
    <language>ja</language>
    <lastBuildDate>Wed, 05 Aug 2026 00:06:36 GMT</lastBuildDate>
    <atom:link href="https://saitotechplus.com/pages/ai-news/feed.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>Nvidia主導のOSAA、設立から1週間でAI安全提言とOSSツールを公開</title>
      <link>https://saitotechplus.com/pages/ai-news/20260804-nvidia-doesn-t-mess-around-a-week-after-1bnxg7.html</link>
      <guid isPermaLink="true">https://saitotechplus.com/pages/ai-news/20260804-nvidia-doesn-t-mess-around-a-week-after-1bnxg7.html</guid>
      <description>Nvidiaが主導し、設立から1週間で120社以上が参加する業界団体「Open Secure AI Alliance（OSAA）」が、AIセキュリティインシデントの報告や分析に関する提案を取りまとめました。同団体は「Shared AI Findings Exchange（SAFE）」という作業部会を立ち上げ、Black Hatカンファレンスで提案を公開しています。また、各社が持つオープンソースのAIセキュリティ技術を持ち寄り、今後共通基盤を整備する方針です。一方、OpenAIやGoogleは公開書簡には署名したものの団体には不参加で、今後の動向が注目されます。</description>
      <category>オープンソース</category>
      <pubDate>Wed, 05 Aug 2026 00:06:35 GMT</pubDate>
    </item>
    <item>
      <title>スペースX、AI向け計算資源の提供で売上高が前年比3倍の26億ドルに</title>
      <link>https://saitotechplus.com/pages/ai-news/20260804-spacex-has-more-neocloud-revenue-1f7avj.html</link>
      <guid isPermaLink="true">https://saitotechplus.com/pages/ai-news/20260804-spacex-has-more-neocloud-revenue-1f7avj.html</guid>
      <description>宇宙開発企業として知られるスペースXが、AI企業向けに計算資源を提供する「ネオクラウド」事業で収入を大きく伸ばした。四半期決算によると、AI関連の売上高は前年比3倍以上の26億ドルに達したが、同部門は依然として15億ドルの赤字を計上している。本記事では、AI分野に進出する宇宙企業の戦略と、事業全体の収益構造の課題を解説する。また、AnthropicやGoogleとの契約により、ネオクラウド市場での競争が激化していることも紹介する。</description>
      <category>ビジネス・資金</category>
      <pubDate>Wed, 05 Aug 2026 00:05:07 GMT</pubDate>
    </item>
    <item>
      <title>Amazon BedrockにWeb検索が標準搭載、ファウンデーションモデルの接地を効率化する</title>
      <link>https://saitotechplus.com/pages/ai-news/20260804-introducing-web-search-on-amazon-bedrock-for-foundation-1pa4q9.html</link>
      <guid isPermaLink="true">https://saitotechplus.com/pages/ai-news/20260804-introducing-web-search-on-amazon-bedrock-for-foundation-1pa4q9.html</guid>
      <description>AWSは、Amazon Bedrock上でファウンデーションモデルを最新のウェブ情報に接地できる「Web Search on Amazon Bedrock」の一般提供を発表した。従来必要だったサードパーティ製の検索サービス統合を廃し、OpenAI互換APIの呼び出しにツール定義を追加するだけで、引用付きの回答を得られる。現時点では米国の3リージョンで提供され、データがAWS環境外に出ない設計となっている。</description>
      <category>モデル・製品</category>
      <pubDate>Wed, 05 Aug 2026 00:03:19 GMT</pubDate>
    </item>
    <item>
      <title>MiniMax-H3がMLX対応、Appleシリコンで15秒動画生成を実現</title>
      <link>https://saitotechplus.com/pages/ai-news/20260804-pipenetwork-minimax-h3-mlx-f66dkc.html</link>
      <guid isPermaLink="true">https://saitotechplus.com/pages/ai-news/20260804-pipenetwork-minimax-h3-mlx-f66dkc.html</guid>
      <description>MiniMaxが公開した汎用オムニモーダル生成システム「MiniMax-H3」を、Appleシリコン向け機械学習フレームワークMLXへ移植するPythonパッケージが登場しました。テキストや画像、音声、動画を入力として最大15秒の音声付き動画を生成できます。実際にM5 Max搭載MacBook Proで動作確認が行われ、約115GBのモデルファイルをダウンロードし、1本の動画生成に45分弱かかったと報告されています。</description>
      <category>モデル・製品</category>
      <pubDate>Tue, 04 Aug 2026 21:08:20 GMT</pubDate>
    </item>
    <item>
      <title>AnthropicがAIクラウド新興Voltaと100億ドル契約、6年間の計算能力確保</title>
      <link>https://saitotechplus.com/pages/ai-news/20260804-anthropic-signs-10b-deal-with-ai-cloud-startup-uexnrd.html</link>
      <guid isPermaLink="true">https://saitotechplus.com/pages/ai-news/20260804-anthropic-signs-10b-deal-with-ai-cloud-startup-uexnrd.html</guid>
      <description>Anthropicは、AIクラウド新興企業Voltaと総額100億ドル・期間6年の計算リソース契約を結んだと報じられた。Voltaは今年設立されたばかりで、暗号資産マイニング企業Bitdeerと協力し、ノルウェーにデータセンターを建設する。Nvidiaの最新チップアーキテクチャ「Vera Rubin」を採用し、133メガワットの計算容量を提供する計画だ。Anthropicのクラウド契約拡大の一環とみられる。</description>
      <category>インフラ・半導体</category>
      <pubDate>Tue, 04 Aug 2026 21:07:41 GMT</pubDate>
    </item>
    <item>
      <title>オープンウェイトAI、先端能力に迫るも安全性に大きなギャップ</title>
      <link>https://saitotechplus.com/pages/ai-news/20260804-open-weight-ai-models-are-catching-up-to-17v9q4.html</link>
      <guid isPermaLink="true">https://saitotechplus.com/pages/ai-news/20260804-open-weight-ai-models-are-catching-up-to-17v9q4.html</guid>
      <description>AI安全団体SaferAIの評価で、中国Z.aiのオープンウェイトモデル『GLM-5.2』がサイバー・バイオ分野の能力でOpenAIやAnthropicの最先端モデルに数カ月差まで迫っていることが分かった。しかし、攻撃的な指示を一切拒否しないなど、安全対策には深刻な問題がある。オープンウェイトモデルが現実の脅威になる中、リスク管理の在り方が問われている。</description>
      <category>モデル・製品</category>
      <pubDate>Tue, 04 Aug 2026 21:06:26 GMT</pubDate>
    </item>
    <item>
      <title>OpenAIがChatGPT Workを公開、知識労働向けエージェントの仕組み</title>
      <link>https://saitotechplus.com/pages/ai-news/20260804-unpacking-chatgpt-work-the-agent-for-a-billion-sim9by.html</link>
      <guid isPermaLink="true">https://saitotechplus.com/pages/ai-news/20260804-unpacking-chatgpt-work-the-agent-for-a-billion-sim9by.html</guid>
      <description>2026年7月9日にOpenAIが公開した知識労働向けエージェント「ChatGPT Work」の概要を解説します。Codexを基盤にクラウド上の仮想マシンで動作し、Slackやメール、カレンダーなどと連携して業務を自動化します。また、メモリ管理やファイル管理の設計、既存製品との違い、今後の課題を整理します。</description>
      <category>モデル・製品</category>
      <pubDate>Tue, 04 Aug 2026 21:04:16 GMT</pubDate>
    </item>
    <item>
      <title>OpenAI、教育プラグイン3種を発表 - ChatGPT Work/Codexで学びを支援</title>
      <link>https://saitotechplus.com/pages/ai-news/20260804-new-ways-to-learn-and-teach-with-chatgpt-1l49zj.html</link>
      <guid isPermaLink="true">https://saitotechplus.com/pages/ai-news/20260804-new-ways-to-learn-and-teach-with-chatgpt-1l49zj.html</guid>
      <description>OpenAIは2026年8月4日、教育機関向けの新プラグインを3種類発表した。K-12（小中高）の教員、大学教員、大学生向けで、教材や既存ツールと連携し、AIエージェントによる多段階の作業を支援する。提供基盤となるChatGPT EduやChatGPT for Teachersの動向、学生コミュニティや研究者向けプログラムの拡充も含めて解説する。</description>
      <category>モデル・製品</category>
      <pubDate>Tue, 04 Aug 2026 18:05:32 GMT</pubDate>
    </item>
    <item>
      <title>NVIDIAがストレージ新技術を発表、cuFileのオープンソース化とStorage-Nextを推進</title>
      <link>https://saitotechplus.com/pages/ai-news/20260804-as-ai-increases-demands-on-memory-storage-steps-1jdof5.html</link>
      <guid isPermaLink="true">https://saitotechplus.com/pages/ai-news/20260804-as-ai-increases-demands-on-memory-storage-steps-1jdof5.html</guid>
      <description>NVIDIAはFMSカンファレンスで、GPUがストレージへ直接アクセスする時代に向けた新たな取り組みを発表しました。cuFile APIのオープンソース化に加え、40以上のベンダーが参加する「Storage-Next」や、安全な高速アクセスを実現する「SCADA」フレームワークを打ち出しています。AIの大規模データ処理を支えるストレージ基盤の変革が加速しそうです。</description>
      <category>インフラ・半導体</category>
      <pubDate>Tue, 04 Aug 2026 18:04:09 GMT</pubDate>
    </item>
    <item>
      <title>自動運転向けオープン推論モデル Alpamayo 2 Superが商用利用可能に</title>
      <link>https://saitotechplus.com/pages/ai-news/20260804-nvidia-alpamayo-2-super-the-frontier-open-model-1s6pps.html</link>
      <guid isPermaLink="true">https://saitotechplus.com/pages/ai-news/20260804-nvidia-alpamayo-2-super-the-frontier-open-model-1s6pps.html</guid>
      <description>NVIDIAは2026年8月4日、自動運転向けのオープン推論モデル「Alpamayo 2 Super」を商用利用可能な形で公開した。ライセンスはLinux FoundationのOpenMDW-1.1で、微調整や派生モデルの作成、商用再配布が認められる。Hugging Faceで公開され、LingoQAなどの自動運転ベンチマークで最高性能を達成している。360度カメラ映像を入力に、軌道計画や因果連鎖の説明、メタアクションなどを出力するマルチタスクモデルだ。</description>
      <category>モデル・製品</category>
      <pubDate>Tue, 04 Aug 2026 18:03:05 GMT</pubDate>
    </item>
    <item>
      <title>AI評価・セキュリティ・小型モデルに動き、Apple対OpenAI訴訟も拡大</title>
      <link>https://saitotechplus.com/pages/ai-news/digest-20260805.html</link>
      <guid isPermaLink="true">https://saitotechplus.com/pages/ai-news/digest-20260805.html</guid>
      <description>8月5日のAI業界は、AI生成物を人間が評価するDesignArenaの資金調達、Apple対OpenAIの訴訟拡大、Linux Foundationによるセキュリティ指針SAFEの提案、Liquid AIによる小型モデル公開と、多様な動きが同時に進行した。信頼性と安全性を軸に、AI開発の新たな段階に入ったことを示す一日だった。</description>
      <category>まとめ</category>
      <pubDate>Tue, 04 Aug 2026 15:04:27 GMT</pubDate>
    </item>
    <item>
      <title>AI生成物に“味覚”を加えるDesignArena、790万ドル調達</title>
      <link>https://saitotechplus.com/pages/ai-news/20260803-designarena-creators-raise-7-9-million-to-bring-owzhrt.html</link>
      <guid isPermaLink="true">https://saitotechplus.com/pages/ai-news/20260803-designarena-creators-raise-7-9-million-to-bring-owzhrt.html</guid>
      <description>AI生成物の品質を人間の評価で改善する「DesignArena」を運営するIntelligence社が、シードラウンドで790万ドルを調達した。このツールは現在約530万人が利用し、AI企業向けにユーザーフィードバックを提供するビジネスとして成長している。すでに年間経常収益（ARR）は6000万ドルに達しており、人間の評価データがAI開発の重要な資源になりつつあることを示している。</description>
      <category>ビジネス・資金</category>
      <pubDate>Tue, 04 Aug 2026 15:03:47 GMT</pubDate>
    </item>
    <item>
      <title>Apple、OpenAIへの営業秘密訴訟で仮処分申請—元従業員の秘密持ち出し疑惑拡大</title>
      <link>https://saitotechplus.com/pages/ai-news/20260804-apple-says-more-ex-employees-may-have-taken-1i8obl.html</link>
      <guid isPermaLink="true">https://saitotechplus.com/pages/ai-news/20260804-apple-says-more-ex-employees-may-have-taken-1i8obl.html</guid>
      <description>Appleは2026年8月4日、OpenAIを相手取った営業秘密訴訟で、AIデバイスなどの開発を差し止める仮処分申請を行った。また、追加の証拠開示を求める中で、当初訴状に名前が挙がっていた元従業員2人以外にも、少なくとも11人の元従業員が秘密情報の持ち出しに関与した可能性があると主張している。OpenAIはこれを「誤った情報に基づく」と否定し、対立は激しさを増している。</description>
      <category>ビジネス・資金</category>
      <pubDate>Tue, 04 Aug 2026 15:03:01 GMT</pubDate>
    </item>
    <item>
      <title>Linux FoundationがAIセキュリティ指針SAFEを提案、120組織が参画するアライアンスが支援</title>
      <link>https://saitotechplus.com/pages/ai-news/20260804-ai-leaders-propose-safe-guidelines-for-cybersecurity-transparency-1m111l.html</link>
      <guid isPermaLink="true">https://saitotechplus.com/pages/ai-news/20260804-ai-leaders-propose-safe-guidelines-for-cybersecurity-transparency-1m111l.html</guid>
      <description>オープンセキュアAIアライアンスは、エージェント型AIのセキュリティインシデントを共有し、防御に生かすための新指針SAFEをLinux Foundationを通じて提案した。NVIDIAやCisco、CrowdStrikeなど120以上の組織が参加し、Black Hatカンファレンスの開幕に合わせて発表された。本記事では、SAFEの概要と、NVIDIAをはじめとする各社の具体的なセキュリティツールの貢献を解説する。</description>
      <category>社会・規制</category>
      <pubDate>Tue, 04 Aug 2026 15:02:27 GMT</pubDate>
    </item>
    <item>
      <title>Liquid AI、小型エージェントモデルLFM2.5-2.6B公開 4倍規模に匹敵</title>
      <link>https://saitotechplus.com/pages/ai-news/20260804-deploy-local-agents-everywhere-with-lfm2-5-2-aodn6g.html</link>
      <guid isPermaLink="true">https://saitotechplus.com/pages/ai-news/20260804-deploy-local-agents-everywhere-with-lfm2-5-2-aodn6g.html</guid>
      <description>Liquid AIは、エッジデバイスで動作する小型エージェントモデル「LFM2.5-2.6B」を公開した。26億パラメータながら、4倍規模のモデルに匹敵する指示追従・ツール使用・マルチステップタスク性能を持つ。訓練には実環境での強化学習を導入し、CPUで毎秒113トークンという高速推論を実現している。</description>
      <category>モデル・製品</category>
      <pubDate>Tue, 04 Aug 2026 15:02:05 GMT</pubDate>
    </item>
    <item>
      <title>パランティアCEOがAI業界を「マルクス主義的」と批判、決算は好調</title>
      <link>https://saitotechplus.com/pages/ai-news/20260803-after-killer-quarter-palantir-ceo-alex-karp-calls-fizy5v.html</link>
      <guid isPermaLink="true">https://saitotechplus.com/pages/ai-news/20260803-after-killer-quarter-palantir-ceo-alex-karp-calls-fizy5v.html</guid>
      <description>パランティア・テクノロジーズは2026年第2四半期に売上高19億ドル、前年同期比93%増の好決算を発表した。同時にCEOのアレックス・カープ氏は、大規模言語モデルを開発するAI企業を「マルクス主義的」と批判し、企業がAIベンダーにデータを渡すリスクを警告した。この発言の背景には、AI市場の急成長と、エンタープライズ向けAIの在り方を巡る業界の対立がある。</description>
      <category>ビジネス・資金</category>
      <pubDate>Tue, 04 Aug 2026 12:07:00 GMT</pubDate>
    </item>
    <item>
      <title>海底ケーブルに代わる宇宙レーザー通信、スタートアップEONがシード調達</title>
      <link>https://saitotechplus.com/pages/ai-news/20260804-eon-wants-to-move-the-data-superhighway-from-15maq1.html</link>
      <guid isPermaLink="true">https://saitotechplus.com/pages/ai-news/20260804-eon-wants-to-move-the-data-superhighway-from-15maq1.html</guid>
      <description>新興企業のEndeavor Optical Networks（EON）は2026年8月4日、宇宙レーザーを使ったデータセンター間通信の計画を発表し、General CatalystとAndreessen Horowitzから1,075万ドルのシード資金を調達した。衛星を約20基配置し、海底ケーブルを代替する高速通信を目指す。初期目標は2.4Tbpsで、デモ衛星を2027年末に打ち上げる予定だ。</description>
      <category>インフラ・半導体</category>
      <pubDate>Tue, 04 Aug 2026 12:06:26 GMT</pubDate>
    </item>
    <item>
      <title>Ai2が衛星データ基盤モデルOlmoEarthを大陸規模推論できるプラットフォームとして公開</title>
      <link>https://saitotechplus.com/pages/ai-news/20260728-the-olmoearth-platform-geospatial-inference-at-planetary-scale-1wtwkc.html</link>
      <guid isPermaLink="true">https://saitotechplus.com/pages/ai-news/20260728-the-olmoearth-platform-geospatial-inference-at-planetary-scale-1wtwkc.html</guid>
      <description>AI2は地球観測向け基盤モデル「OlmoEarth」を、大陸規模の地理空間推論を実行できるプラットフォームとして公開した。約10テラバイトの衛星データで事前学習したモデルを、データ取得から推論、後処理まで一貫して自動化する。森林減少や食料安全保障、山火事リスク評価など環境分野の組織が自前の大規模インフラなしで利用できるのが特徴だ。</description>
      <category>インフラ・半導体</category>
      <pubDate>Tue, 04 Aug 2026 12:05:50 GMT</pubDate>
    </item>
    <item>
      <title>コーディングAIが科学計算を変える OpenAIが8件の実践報告を公開</title>
      <link>https://saitotechplus.com/pages/ai-news/20260728-scientific-computing-in-the-age-of-agentic-ai-bij8w4.html</link>
      <guid isPermaLink="true">https://saitotechplus.com/pages/ai-news/20260728-scientific-computing-in-the-age-of-agentic-ai-bij8w4.html</guid>
      <description>OpenAIは2026年7月28日、AIエージェントを使って科学計算ソフトウェアを開発・保守した8件の実地調査レポートを公開した。生命科学を中心に、Codex単独またはClaude Codeと組み合わせて活用。研究用ツールの近代化や性能改善が加速する一方、エージェントの成果を検証する人間の役割と長期的な維持管理の重要性も浮き彫りになっている。</description>
      <category>開発・実装</category>
      <pubDate>Tue, 04 Aug 2026 12:04:15 GMT</pubDate>
    </item>
    <item>
      <title>Claude Mythosが暗号解析で成果、HAWKとAESの弱点を発見</title>
      <link>https://saitotechplus.com/pages/ai-news/20260728-discovering-cryptographic-weaknesses-with-claude-1p5z41.html</link>
      <guid isPermaLink="true">https://saitotechplus.com/pages/ai-news/20260728-discovering-cryptographic-weaknesses-with-claude-1p5z41.html</guid>
      <description>Anthropicの研究者らが、AIモデルClaude Mythosを使って暗号方式HAWKとラウンド数を減らしたAESに数学的弱点を発見した。Simon Willison氏のブログで紹介されたこの研究は、現行システムへの実用的影響はないとされるが、AIが暗号解析をどこまで進められるかを示す事例として注目される。合わせて、LLMの暗号解析能力を評価するベンチマーク「CryptanalysisBench」も発表された。</description>
      <category>研究</category>
      <pubDate>Tue, 04 Aug 2026 09:03:38 GMT</pubDate>
    </item>
    <item>
      <title>Amazon Bedrock AgentCoreとMCPサーバーで自律BIを設定だけで実現</title>
      <link>https://saitotechplus.com/pages/ai-news/20260729-generate-autonomous-business-insights-with-ai-agent-and-1jvyk2.html</link>
      <guid isPermaLink="true">https://saitotechplus.com/pages/ai-news/20260729-generate-autonomous-business-insights-with-ai-agent-and-1jvyk2.html</guid>
      <description>Amazon Web Services（AWS）は、ブログ「Generate Autonomous Business Insights with AI Agent and MCP Servers」で、Amazon Bedrock AgentCoreとMCPサーバーを組み合わせた自律型BIの参照アーキテクチャを公開した。工場管理者が複数の業務システムを手作業で横断する代わりに、自然言語で質問すると、エージェントがデータの所在を判別し、必要なデータソースから回答を合成して返す。コードを書かず、既存システムをMCPサーバーで接続する構成となっている。実装はGitHubで公開されており、AWS環境で試…</description>
      <category>開発・実装</category>
      <pubDate>Tue, 04 Aug 2026 09:01:48 GMT</pubDate>
    </item>
    <item>
      <title>Bedrock AgentCore IdentityがPrivate Key JWT認証に対応、KMS署名でOAuth連携を強化</title>
      <link>https://saitotechplus.com/pages/ai-news/20260729-authenticate-with-private-key-jwt-using-amazon-bedrock-g3ug7j.html</link>
      <guid isPermaLink="true">https://saitotechplus.com/pages/ai-news/20260729-authenticate-with-private-key-jwt-using-amazon-bedrock-g3ug7j.html</guid>
      <description>Amazon Bedrock AgentCore Identityは、エージェントがOAuth 2.0のクライアント認証を行う際に、共有シークレットの代わりに署名付きJWTを使うPrivate Key JWT方式をサポートしました。署名はAWS KMSの非対称鍵で行われ、秘密鍵がKMSの外に出ないのが特徴です。本記事では、仕組みから設定手順、CloudTrailによる監査までを解説します。</description>
      <category>開発・実装</category>
      <pubDate>Tue, 04 Aug 2026 09:00:59 GMT</pubDate>
    </item>
    <item>
      <title>Microsoft WordのCopilotで自己複製型プロンプトインジェクションが可能に</title>
      <link>https://saitotechplus.com/pages/ai-news/20260729-ai-worming-through-word-yoq4sz.html</link>
      <guid isPermaLink="true">https://saitotechplus.com/pages/ai-news/20260729-ai-worming-through-word-yoq4sz.html</guid>
      <description>Håkon Måløy氏が、Microsoft WordのCopilotを標的とした新しいプロンプトインジェクション手法を発表しました。隠し指示を文書に埋め込み、Copilotに操作させることで、新しい文書へ指示をコピーして自己複製するワーム型の攻撃が可能になります。Microsoftには責任付きで開示済みですが、現時点では完全な対策は提供されていません。</description>
      <category>研究</category>
      <pubDate>Tue, 04 Aug 2026 06:05:04 GMT</pubDate>
    </item>
    <item>
      <title>GPT-Realtimeでヤマダデンキの接客ノウハウを24時間AI化、2週間で約3万人が利用</title>
      <link>https://saitotechplus.com/pages/ai-news/20260730-how-avatarin-built-a-24-7-retail-agent-12mb6h.html</link>
      <guid isPermaLink="true">https://saitotechplus.com/pages/ai-news/20260730-how-avatarin-built-a-24-7-retail-agent-12mb6h.html</guid>
      <description>avatarinは、OpenAIのGPT-Realtimeを使い、ヤマダデンキ向けの音声対応ショッピングエージェント「Kurashi-Marugoto AI Agent」を開発した。2週間の公開キャンペーンでは約3万人が利用し、事後アンケートの92%が肯定的な回答だった。従来のチャットボットを超えた自然な対話と、販売員のノウハウを組み込んだ設計が特徴である。</description>
      <category>開発・実装</category>
      <pubDate>Tue, 04 Aug 2026 06:03:47 GMT</pubDate>
    </item>
    <item>
      <title>Qwen 3.8 Max（2.4T）発表、オープンウェイトで自律コーディング強化</title>
      <link>https://saitotechplus.com/pages/ai-news/20260804-ainews-qwen-3-8-max-2-4t-and-1abbf3.html</link>
      <guid isPermaLink="true">https://saitotechplus.com/pages/ai-news/20260804-ainews-qwen-3-8-max-2-4t-and-1abbf3.html</guid>
      <description>Alibabaは2026年8月4日、新しいフラッグシップモデル「Qwen3.8-Max」を発表した。2.4TパラメータのMoEモデルで、コーディングや長時間のエージェントワーク、マルチモーダル推論に重点を置く。API価格は入力2ドル/出力6ドル（100万トークンあたり）で、オープンウェイトは来週公開予定。小型のQwen3.8-27Bも同時にオープンウェイトとなる。第三者による報告では、アクティブパラメータ95B、1Mトークンのコンテキスト窓を持ち、PaperBench 93.0などを記録したとされる。</description>
      <category>モデル・製品</category>
      <pubDate>Tue, 04 Aug 2026 06:02:19 GMT</pubDate>
    </item>
    <item>
      <title>Amazon Bedrock、最大5モデルを一括最適化するプロンプト移行支援機能</title>
      <link>https://saitotechplus.com/pages/ai-news/20260730-migrate-your-prompts-to-new-models-and-optimize-k6hiiy.html</link>
      <guid isPermaLink="true">https://saitotechplus.com/pages/ai-news/20260730-migrate-your-prompts-to-new-models-and-optimize-k6hiiy.html</guid>
      <description>AWSは2026年7月30日、Amazon Bedrockの新機能「Advanced Prompt Optimization」を公開した。プロンプトの書き換えと評価を自動で反復し、最大5モデルの品質・レイテンシー・コストを比較できる。従来は数日から数週間かかっていたプロンプト移行作業を、指標駆動のワークフローに置き換えるのが狙いだ。</description>
      <category>開発・実装</category>
      <pubDate>Tue, 04 Aug 2026 03:14:19 GMT</pubDate>
    </item>
    <item>
      <title>経験を進化する知識に変える、LLM推論時学習「EvoLib」</title>
      <link>https://saitotechplus.com/pages/ai-news/20260730-evolib-turning-experience-into-evolving-knowledge-1l7u63.html</link>
      <guid isPermaLink="true">https://saitotechplus.com/pages/ai-news/20260730-evolib-turning-experience-into-evolving-knowledge-1l7u63.html</guid>
      <description>Microsoft Researchは、大規模言語モデル（LLM）が推論中に自分の経験から学習できるようにするフレームワーク「EvoLib」を発表した。モデルのパラメータ更新を必要とせず、ブラックボックスのAPIモデルにも適用できる点が特徴だ。過去の試行を再利用可能なスキルや洞察に変換し、継続的に洗練することで、数学的推論やコード生成などのタスクで既存手法を上回る性能を示したという。</description>
      <category>研究</category>
      <pubDate>Tue, 04 Aug 2026 03:12:16 GMT</pubDate>
    </item>
    <item>
      <title>Basetenが130億ドル調達、推論エンジニアリングの最前線とは</title>
      <link>https://saitotechplus.com/pages/ai-news/20260803-the-inference-engineering-masterclass-philip-kiely-ali-taha-4nk9kq.html</link>
      <guid isPermaLink="true">https://saitotechplus.com/pages/ai-news/20260803-the-inference-engineering-masterclass-philip-kiely-ali-taha-4nk9kq.html</guid>
      <description>AIモデルを実運用で高速・安定的に動かす「推論エンジニアリング」が、AIインフラの重要分野として存在感を増している。Basetenは130億ドルの大型調達を完了し、AIインフラの一角を担う企業へと成長。本記事では、同社のエンジニアが語ったキャッシュ活用や量子化、投機的デコードなどの最新技術と、オープンモデルをいち早く本番運用に乗せる方法を解説する。</description>
      <category>開発・実装</category>
      <pubDate>Tue, 04 Aug 2026 00:08:46 GMT</pubDate>
    </item>
    <item>
      <title>ウクライナの格安FPVドローンにAI自律追尾、米Auterion社が5万機分供給</title>
      <link>https://saitotechplus.com/pages/ai-news/20260803-us-company-s-ai-lets-ukraine-s-cheap-df5y3j.html</link>
      <guid isPermaLink="true">https://saitotechplus.com/pages/ai-news/20260803-us-company-s-ai-lets-ukraine-s-cheap-df5y3j.html</guid>
      <description>ウクライナ軍が運用する約400ドルの安価なFPVドローン「Shrike」に、米Auterion社が開発したAI自律追尾システム「Skynode S」が搭載され始めた。このシステムはカメラ映像のみで移動目標を追尾するため、GPSや無線通信に依存せず、電子妨害下でも攻撃を継続できる。今後数ヶ月で5万機分が納入される予定で、ウクライナの戦闘方法をさらに変える可能性がある。</description>
      <category>モデル・製品</category>
      <pubDate>Tue, 04 Aug 2026 00:05:12 GMT</pubDate>
    </item>
    <item>
      <title>AlibabaがQwen3.8-Max発表、エージェント型AIでGPT-5.6超えを主張</title>
      <link>https://saitotechplus.com/pages/ai-news/20260803-qwen3-8-max-arrives-with-a-bold-claim-r4yqea.html</link>
      <guid isPermaLink="true">https://saitotechplus.com/pages/ai-news/20260803-qwen3-8-max-arrives-with-a-bold-claim-r4yqea.html</guid>
      <description>AlibabaのQwenチームは2026年8月2日夜、新しいフラッグシップモデル「Qwen3.8-Max」を発表しました。2.4兆パラメータのMoE（Mixture-of-Experts）構造を持つマルチモーダル大規模言語モデルで、自律的なソフトウェア開発や長期エンタープライズ業務を狙います。同社の公表ベンチマークではOSWorld-Verifiedで86.1を記録し、GPT-5.6 Sol Maxの83.2やFable 5の85.0を上回ったとされます。オープンウェイトは来週公開予定ですが、ライセンス条件は未公開です。</description>
      <category>モデル・製品</category>
      <pubDate>Tue, 04 Aug 2026 00:03:58 GMT</pubDate>
    </item>
  </channel>
</rss>
