https://24-ai.news/hr/news/2026-06-23/anthropic-claude-code-2-1-186/ 24 AI hr 2026-06-23T00:00:00Z Anthropic: Claude Code v2.1.186 donosi MCP login/logout, filtriranje workflowa i upravljanje skillovima https://24-ai.news/hr/news/2026-06-23/arxiv-nrt-bench-multiturn-redteaming/ 24 AI hr 2026-06-23T00:00:00Z arXiv:2606.20408: NRT-Bench — benchmark multi-turn red-teaminga LLM agenata u sigurnosno-kritičnim sustavima https://24-ai.news/hr/news/2026-06-23/arxiv-overprivileged-tools-llm-agents/ 24 AI hr 2026-06-23T00:00:00Z arXiv:2606.20023: Kada su niže privilegije dovoljne — LLM agenti biraju previše moćne alate https://24-ai.news/hr/news/2026-06-23/arxiv-tool-programs-agentic-web/ 24 AI hr 2026-06-23T00:00:00Z arXiv:2606.19992: Beyond Static Endpoints — ToolPro uvodi izvršive programe kao sučelje za agentske web servise https://24-ai.news/hr/news/2026-06-23/aws-agentcore-payments-x402/ 24 AI hr 2026-06-23T00:00:00Z AWS: Bedrock AgentCore Payments — AI agenti plaćaju API-je autonomno putem x402 protokola https://24-ai.news/hr/news/2026-06-23/aws-nova-aerial-imagery-search/ 24 AI hr 2026-06-23T00:00:00Z AWS: semantičko pretraživanje zračnih snimaka uz Amazon Nova Multimodal Embeddings (Vexcel) https://24-ai.news/hr/news/2026-06-23/cncf-observability-agentic-systems/ 24 AI hr 2026-06-23T00:00:00Z CNCF: zašto klasična observability ne radi za agentic i LLM sustave https://24-ai.news/hr/news/2026-06-23/github-copilot-claude-agent-jetbrains/ 24 AI hr 2026-06-23T00:00:00Z GitHub: Claude Code kao agent-provider u JetBrains IDE-ovima — public preview i cloud agent GA https://24-ai.news/hr/news/2026-06-23/ibm-openai-frontier-cyber-defense/ 24 AI hr 2026-06-23T00:00:00Z IBM i OpenAI: frontier AI u kibernetičkoj obrani poduzeća protiv prijetnji brzinom stroja https://24-ai.news/hr/news/2026-06-23/ibm-wimbledon-2026-ai-experiences/ 24 AI hr 2026-06-23T00:00:00Z IBM: Wimbledon 2026 dobiva AI osobnog vodiča za gledatelje i modernizirane digitalne platforme https://24-ai.news/hr/news/2026-06-23/nist-continuous-monitoring-ai-security/ 24 AI hr 2026-06-23T00:00:00Z NIST: Matematički dokaz podupire prelazak na kontinuirano praćenje sigurnosti AI sustava https://24-ai.news/hr/news/2026-06-23/nvidia-cudax-science-software/ 24 AI hr 2026-06-23T00:00:00Z NVIDIA: CUDA-X biblioteke cuPhoton, DAQIRI i ALCHEMI ubrzavaju astronomiju, kemiju i materijale https://24-ai.news/hr/news/2026-06-23/nvidia-jupiter-exascale-europe-science/ 24 AI hr 2026-06-23T00:00:00Z NVIDIA: JUPITER — prvo europsko exascale superračunalo postavlja znanstvene rekorde na ISC 2026 https://24-ai.news/hr/news/2026-06-23/nvidia-vera-cpu-los-alamos-science/ 24 AI hr 2026-06-23T00:00:00Z NVIDIA: Vera CPU u Los Alamosu — 7× brži agentic AI za nuklearnu znanost i 3 nova superračunala https://24-ai.news/hr/news/2026-06-23/openai-daybreak-codex-security-cyber/ 24 AI hr 2026-06-23T00:00:00Z OpenAI: Daybreak — Codex Security i GPT-5.5-Cyber stižu u borbu protiv ranjivosti u velikom razmjeru https://24-ai.news/en/news/2026-06-23/anthropic-claude-code-2-1-186/ 24 AI en 2026-06-23T00:00:00Z Anthropic: Claude Code v2.1.186 brings MCP login/logout, workflow filtering and skills management https://24-ai.news/en/news/2026-06-23/arxiv-nrt-bench-multiturn-redteaming/ 24 AI en 2026-06-23T00:00:00Z arXiv:2606.20408: NRT-Bench — a multi-turn red-teaming benchmark for LLM agents in safety-critical systems https://24-ai.news/en/news/2026-06-23/arxiv-overprivileged-tools-llm-agents/ 24 AI en 2026-06-23T00:00:00Z arXiv:2606.20023: When lower privileges suffice — LLM agents choose overly powerful tools https://24-ai.news/en/news/2026-06-23/arxiv-tool-programs-agentic-web/ 24 AI en 2026-06-23T00:00:00Z arXiv:2606.19992: Beyond Static Endpoints — ToolPro introduces executable programs as an interface for agentic web services https://24-ai.news/en/news/2026-06-23/aws-agentcore-payments-x402/ 24 AI en 2026-06-23T00:00:00Z AWS: Bedrock AgentCore Payments — AI agents pay APIs autonomously via the x402 protocol https://24-ai.news/en/news/2026-06-23/aws-nova-aerial-imagery-search/ 24 AI en 2026-06-23T00:00:00Z AWS: semantic search of aerial imagery with Amazon Nova Multimodal Embeddings (Vexcel) https://24-ai.news/en/news/2026-06-23/cncf-observability-agentic-systems/ 24 AI en 2026-06-23T00:00:00Z CNCF: why classical observability does not work for agentic and LLM systems https://24-ai.news/en/news/2026-06-23/github-copilot-claude-agent-jetbrains/ 24 AI en 2026-06-23T00:00:00Z GitHub: Claude Code as an agent provider in JetBrains IDEs — public preview and cloud agent GA https://24-ai.news/en/news/2026-06-23/ibm-openai-frontier-cyber-defense/ 24 AI en 2026-06-23T00:00:00Z IBM and OpenAI: frontier AI in enterprise cyber defense against machine-speed threats https://24-ai.news/en/news/2026-06-23/ibm-wimbledon-2026-ai-experiences/ 24 AI en 2026-06-23T00:00:00Z IBM: Wimbledon 2026 gets an AI personal guide for fans and modernized digital platforms https://24-ai.news/en/news/2026-06-23/nist-continuous-monitoring-ai-security/ 24 AI en 2026-06-23T00:00:00Z NIST: Mathematical proof supports the shift to continuous monitoring of AI system security https://24-ai.news/en/news/2026-06-23/nvidia-cudax-science-software/ 24 AI en 2026-06-23T00:00:00Z NVIDIA: CUDA-X libraries cuPhoton, DAQIRI, and ALCHEMI accelerate astronomy, chemistry, and materials science https://24-ai.news/en/news/2026-06-23/nvidia-jupiter-exascale-europe-science/ 24 AI en 2026-06-23T00:00:00Z NVIDIA: JUPITER — Europe's first exascale supercomputer sets scientific records at ISC 2026 https://24-ai.news/en/news/2026-06-23/nvidia-vera-cpu-los-alamos-science/ 24 AI en 2026-06-23T00:00:00Z NVIDIA: Vera CPU at Los Alamos — 7× faster agentic AI for nuclear science and 3 new supercomputers https://24-ai.news/en/news/2026-06-23/openai-daybreak-codex-security-cyber/ 24 AI en 2026-06-23T00:00:00Z OpenAI: Daybreak — Codex Security and GPT-5.5-Cyber arrive to fight vulnerabilities at scale https://24-ai.news/de/news/2026-06-23/anthropic-claude-code-2-1-186/ 24 AI de 2026-06-23T00:00:00Z Anthropic: Claude Code v2.1.186 bringt MCP-Login/Logout, Workflow-Filter und Skill-Verwaltung https://24-ai.news/de/news/2026-06-23/arxiv-nrt-bench-multiturn-redteaming/ 24 AI de 2026-06-23T00:00:00Z arXiv:2606.20408: NRT-Bench — Benchmark für Multi-Turn-Red-Teaming von KI-Agenten in sicherheitskritischen Systemen https://24-ai.news/de/news/2026-06-23/arxiv-overprivileged-tools-llm-agents/ 24 AI de 2026-06-23T00:00:00Z arXiv:2606.20023: Wenn niedrigere Rechte ausreichen — KI-Agenten wählen zu mächtige Werkzeuge https://24-ai.news/de/news/2026-06-23/arxiv-tool-programs-agentic-web/ 24 AI de 2026-06-23T00:00:00Z arXiv:2606.19992: Beyond Static Endpoints — ToolPro führt ausführbare Programme als Schnittstelle für agentische Webdienste ein https://24-ai.news/de/news/2026-06-23/aws-agentcore-payments-x402/ 24 AI de 2026-06-23T00:00:00Z AWS: Bedrock AgentCore Payments — KI-Agenten zahlen APIs autonom per x402-Protokoll https://24-ai.news/de/news/2026-06-23/aws-nova-aerial-imagery-search/ 24 AI de 2026-06-23T00:00:00Z AWS: Semantische Suche in Luftaufnahmen mit Amazon Nova Multimodal Embeddings (Vexcel) https://24-ai.news/de/news/2026-06-23/cncf-observability-agentic-systems/ 24 AI de 2026-06-23T00:00:00Z CNCF: Warum klassische Observability für agentische KI-Systeme und LLMs nicht funktioniert https://24-ai.news/de/news/2026-06-23/github-copilot-claude-agent-jetbrains/ 24 AI de 2026-06-23T00:00:00Z GitHub: Claude Code als Agent-Provider in JetBrains-IDEs — Public Preview und Cloud-Agent GA https://24-ai.news/de/news/2026-06-23/ibm-openai-frontier-cyber-defense/ 24 AI de 2026-06-23T00:00:00Z IBM und OpenAI: Frontier-KI in der Unternehmens-Cyberabwehr gegen Bedrohungen mit Maschinengeschwindigkeit https://24-ai.news/de/news/2026-06-23/ibm-wimbledon-2026-ai-experiences/ 24 AI de 2026-06-23T00:00:00Z IBM: Wimbledon 2026 erhält einen KI-gestützten persönlichen Begleiter für Zuschauer und modernisierte digitale Plattformen https://24-ai.news/de/news/2026-06-23/nist-continuous-monitoring-ai-security/ 24 AI de 2026-06-23T00:00:00Z NIST: Mathematischer Beweis unterstützt Übergang zur kontinuierlichen Überwachung der KI-Sicherheit https://24-ai.news/de/news/2026-06-23/nvidia-cudax-science-software/ 24 AI de 2026-06-23T00:00:00Z NVIDIA: CUDA-X-Bibliotheken cuPhoton, DAQIRI und ALCHEMI beschleunigen Astronomie, Chemie und Materialwissenschaft https://24-ai.news/de/news/2026-06-23/nvidia-jupiter-exascale-europe-science/ 24 AI de 2026-06-23T00:00:00Z NVIDIA: JUPITER — Europas erstes Exascale-Supercomputer stellt Wissenschaftsrekorde auf der ISC 2026 auf https://24-ai.news/de/news/2026-06-23/nvidia-vera-cpu-los-alamos-science/ 24 AI de 2026-06-23T00:00:00Z NVIDIA: Vera CPU in Los Alamos — 7× schnellere agentische KI für Nuklearwissenschaft und 3 neue Supercomputer https://24-ai.news/de/news/2026-06-23/openai-daybreak-codex-security-cyber/ 24 AI de 2026-06-23T00:00:00Z OpenAI: Daybreak — Codex Security und GPT-5.5-Cyber treten im Kampf gegen Schwachstellen im großen Maßstab an https://24-ai.news/zh/news/2026-06-23/anthropic-claude-code-2-1-186/ 24 AI zh 2026-06-23T00:00:00Z Anthropic：Claude Code v2.1.186 新增 MCP 登录/登出、工作流过滤与技能管理功能 https://24-ai.news/zh/news/2026-06-23/arxiv-nrt-bench-multiturn-redteaming/ 24 AI zh 2026-06-23T00:00:00Z arXiv:2606.20408: NRT-Bench——用于安全关键系统中 LLM 智能体多轮红队测试的基准 https://24-ai.news/zh/news/2026-06-23/arxiv-overprivileged-tools-llm-agents/ 24 AI zh 2026-06-23T00:00:00Z arXiv:2606.20023: 低权限已够用——LLM 智能体仍倾向选择高权限工具 https://24-ai.news/zh/news/2026-06-23/arxiv-tool-programs-agentic-web/ 24 AI zh 2026-06-23T00:00:00Z arXiv:2606.19992: 超越静态端点——ToolPro 引入可执行程序作为智能体 Web 服务接口 https://24-ai.news/zh/news/2026-06-23/aws-agentcore-payments-x402/ 24 AI zh 2026-06-23T00:00:00Z AWS：Bedrock AgentCore Payments——AI 智能体通过 x402 协议自主支付 API 费用 https://24-ai.news/zh/news/2026-06-23/aws-nova-aerial-imagery-search/ 24 AI zh 2026-06-23T00:00:00Z AWS：借助 Amazon Nova 多模态嵌入实现航拍图像语义搜索（Vexcel 案例） https://24-ai.news/zh/news/2026-06-23/cncf-observability-agentic-systems/ 24 AI zh 2026-06-23T00:00:00Z CNCF：为何传统可观测性不适用于智能体和 LLM 系统 https://24-ai.news/zh/news/2026-06-23/github-copilot-claude-agent-jetbrains/ 24 AI zh 2026-06-23T00:00:00Z GitHub：Claude Code 作为 JetBrains IDE 中的智能体提供商——公开预览与云智能体 GA https://24-ai.news/zh/news/2026-06-23/ibm-openai-frontier-cyber-defense/ 24 AI zh 2026-06-23T00:00:00Z IBM 与 OpenAI：前沿 AI 助力企业网络防御，应对机器速度威胁 https://24-ai.news/zh/news/2026-06-23/ibm-wimbledon-2026-ai-experiences/ 24 AI zh 2026-06-23T00:00:00Z IBM：温布尔登2026年引入 AI 个人向导与升级数字平台 https://24-ai.news/zh/news/2026-06-23/nist-continuous-monitoring-ai-security/ 24 AI zh 2026-06-23T00:00:00Z NIST：数学证明支持向 AI 系统持续安全监控模式过渡 https://24-ai.news/zh/news/2026-06-23/nvidia-cudax-science-software/ 24 AI zh 2026-06-23T00:00:00Z NVIDIA：CUDA-X 库 cuPhoton、DAQIRI 和 ALCHEMI 加速天文学、化学与材料研究 https://24-ai.news/zh/news/2026-06-23/nvidia-jupiter-exascale-europe-science/ 24 AI zh 2026-06-23T00:00:00Z NVIDIA：JUPITER——欧洲首台 E 级超算在 ISC 2026 创下科学记录 https://24-ai.news/zh/news/2026-06-23/nvidia-vera-cpu-los-alamos-science/ 24 AI zh 2026-06-23T00:00:00Z NVIDIA：Vera CPU 落户洛斯阿拉莫斯——智能体 AI 性能提升7倍，三台新超算同步推进 https://24-ai.news/zh/news/2026-06-23/openai-daybreak-codex-security-cyber/ 24 AI zh 2026-06-23T00:00:00Z OpenAI：Daybreak——Codex Security 与 GPT-5.5-Cyber 大规模应对安全漏洞 https://24-ai.news/ja/news/2026-06-23/anthropic-claude-code-2-1-186/ 24 AI ja 2026-06-23T00:00:00Z Anthropic: Claude Code v2.1.186 — MCP ログイン/ログアウト、ワークフローフィルター、スキル管理を追加 https://24-ai.news/ja/news/2026-06-23/arxiv-nrt-bench-multiturn-redteaming/ 24 AI ja 2026-06-23T00:00:00Z arXiv:2606.20408: NRT-Bench — 安全クリティカルシステムにおけるLLMエージェントのマルチターン・レッドチーミングベンチマーク https://24-ai.news/ja/news/2026-06-23/arxiv-overprivileged-tools-llm-agents/ 24 AI ja 2026-06-23T00:00:00Z arXiv:2606.20023: 低権限で十分なときに — LLMエージェントは過剰に強力なツールを選ぶ https://24-ai.news/ja/news/2026-06-23/arxiv-tool-programs-agentic-web/ 24 AI ja 2026-06-23T00:00:00Z arXiv:2606.19992: 静的エンドポイントを超えて — ToolProがエージェント型Webサービスに実行可能プログラムを導入 https://24-ai.news/ja/news/2026-06-23/aws-agentcore-payments-x402/ 24 AI ja 2026-06-23T00:00:00Z AWS: Bedrock AgentCore Payments — AIエージェントがx402プロトコルでAPIに自律的に支払う https://24-ai.news/ja/news/2026-06-23/aws-nova-aerial-imagery-search/ 24 AI ja 2026-06-23T00:00:00Z AWS: Amazon Nova Multimodal Embeddingsによる航空写真のセマンティック検索（Vexcel） https://24-ai.news/ja/news/2026-06-23/cncf-observability-agentic-systems/ 24 AI ja 2026-06-23T00:00:00Z CNCF: 従来のオブザーバビリティがエージェント型・LLMシステムに機能しない理由 https://24-ai.news/ja/news/2026-06-23/github-copilot-claude-agent-jetbrains/ 24 AI ja 2026-06-23T00:00:00Z GitHub: JetBrains IDEでClaudeをエージェントプロバイダーとして提供 — パブリックプレビューとクラウドエージェントGA https://24-ai.news/ja/news/2026-06-23/ibm-openai-frontier-cyber-defense/ 24 AI ja 2026-06-23T00:00:00Z IBMとOpenAI: マシンスピードの脅威に対抗するエンタープライズサイバー防衛にフロンティアAIを導入 https://24-ai.news/ja/news/2026-06-23/ibm-wimbledon-2026-ai-experiences/ 24 AI ja 2026-06-23T00:00:00Z IBM: ウィンブルドン2026がAIパーソナルガイドとデジタルプラットフォームの刷新を導入 https://24-ai.news/ja/news/2026-06-23/nist-continuous-monitoring-ai-security/ 24 AI ja 2026-06-23T00:00:00Z NIST: 数学的証明がAIシステムの継続的セキュリティ監視への移行を支持 https://24-ai.news/ja/news/2026-06-23/nvidia-cudax-science-software/ 24 AI ja 2026-06-23T00:00:00Z NVIDIA: CUDA-Xライブラリ cuPhoton・DAQIRI・ALCHEMIが天文学・化学・素材研究を加速 https://24-ai.news/ja/news/2026-06-23/nvidia-jupiter-exascale-europe-science/ 24 AI ja 2026-06-23T00:00:00Z NVIDIA: JUPITER — ヨーロッパ初のエクサスケールスーパーコンピュータがISC 2026で科学記録を更新 https://24-ai.news/ja/news/2026-06-23/nvidia-vera-cpu-los-alamos-science/ 24 AI ja 2026-06-23T00:00:00Z NVIDIA: ロスアラモスにVera CPU登場 — 核科学向けエージェントAIを7倍高速化、新スーパーコンピュータ3台 https://24-ai.news/ja/news/2026-06-23/openai-daybreak-codex-security-cyber/ 24 AI ja 2026-06-23T00:00:00Z OpenAI: Daybreak — Codex SecurityとGPT-5.5-Cyberが大規模な脆弱性対策に登場 https://24-ai.news/ko/news/2026-06-23/anthropic-claude-code-2-1-186/ 24 AI ko 2026-06-23T00:00:00Z Anthropic: Claude Code v2.1.186, MCP 로그인/로그아웃·워크플로 필터·스킬 관리 기능 추가 https://24-ai.news/ko/news/2026-06-23/arxiv-nrt-bench-multiturn-redteaming/ 24 AI ko 2026-06-23T00:00:00Z arXiv:2606.20408: NRT-Bench — 안전 위험 시스템에서 LLM 에이전트의 멀티턴 레드팀 벤치마크 https://24-ai.news/ko/news/2026-06-23/arxiv-overprivileged-tools-llm-agents/ 24 AI ko 2026-06-23T00:00:00Z arXiv:2606.20023: 낮은 권한으로 충분할 때도 — LLM 에이전트는 과도한 권한의 도구를 선택합니다 https://24-ai.news/ko/news/2026-06-23/arxiv-tool-programs-agentic-web/ 24 AI ko 2026-06-23T00:00:00Z arXiv:2606.19992: Beyond Static Endpoints — ToolPro, 에이전틱 웹 서비스의 실행 가능한 프로그램 인터페이스 https://24-ai.news/ko/news/2026-06-23/aws-agentcore-payments-x402/ 24 AI ko 2026-06-23T00:00:00Z AWS: Bedrock AgentCore Payments — AI 에이전트, x402 프로토콜로 API 자율 결제 https://24-ai.news/ko/news/2026-06-23/aws-nova-aerial-imagery-search/ 24 AI ko 2026-06-23T00:00:00Z AWS: Amazon Nova 멀티모달 임베딩으로 항공 이미지 시맨틱 검색 (Vexcel) https://24-ai.news/ko/news/2026-06-23/cncf-observability-agentic-systems/ 24 AI ko 2026-06-23T00:00:00Z CNCF: 기존 옵저버빌리티가 에이전틱·LLM 시스템에 작동하지 않는 이유 https://24-ai.news/ko/news/2026-06-23/github-copilot-claude-agent-jetbrains/ 24 AI ko 2026-06-23T00:00:00Z GitHub: JetBrains IDE에서 Claude Code 에이전트 프로바이더 공개 프리뷰 및 클라우드 에이전트 GA https://24-ai.news/ko/news/2026-06-23/ibm-openai-frontier-cyber-defense/ 24 AI ko 2026-06-23T00:00:00Z IBM과 OpenAI: 머신 속도 위협에 맞서는 프론티어 AI 기업 사이버 방어 https://24-ai.news/ko/news/2026-06-23/ibm-wimbledon-2026-ai-experiences/ 24 AI ko 2026-06-23T00:00:00Z IBM: 윔블던 2026, AI 개인 가이드와 디지털 플랫폼 현대화 https://24-ai.news/ko/news/2026-06-23/nist-continuous-monitoring-ai-security/ 24 AI ko 2026-06-23T00:00:00Z NIST: AI 시스템 보안의 지속적 모니터링 전환을 지지하는 수학적 증명 https://24-ai.news/ko/news/2026-06-23/nvidia-cudax-science-software/ 24 AI ko 2026-06-23T00:00:00Z NVIDIA: CUDA-X 라이브러리 cuPhoton·DAQIRI·ALCHEMI, 천문학·화학·재료과학 가속 https://24-ai.news/ko/news/2026-06-23/nvidia-jupiter-exascale-europe-science/ 24 AI ko 2026-06-23T00:00:00Z NVIDIA: JUPITER — ISC 2026에서 유럽 최초 엑사스케일 슈퍼컴퓨터, 과학 신기록 수립 https://24-ai.news/ko/news/2026-06-23/nvidia-vera-cpu-los-alamos-science/ 24 AI ko 2026-06-23T00:00:00Z NVIDIA: 로스앨러모스의 Vera CPU — 핵과학 에이전틱 AI 7배 가속 및 3개 신규 슈퍼컴퓨터 https://24-ai.news/ko/news/2026-06-23/openai-daybreak-codex-security-cyber/ 24 AI ko 2026-06-23T00:00:00Z OpenAI: Daybreak — Codex Security와 GPT-5.5-Cyber, 대규모 취약점 대응에 출격 https://24-ai.news/hr/news/2026-06-21/anthropic-claude-code-2-1-185-stall-hint/ 24 AI hr 2026-06-21T00:00:00Z Anthropic (Claude Code GitHub): Claude Code v2.1.185 poboljšava poruku pri zastoju streama https://24-ai.news/hr/news/2026-06-21/arxiv-activation-directions-misalignment/ 24 AI hr 2026-06-21T00:00:00Z arXiv:2606.20225: Aktivacijski smjerovi otkrivaju pogrešno poravnanje LLM-ova s 99,6% točnošću https://24-ai.news/hr/news/2026-06-21/arxiv-contagion-networks-evaluator-bias/ 24 AI hr 2026-06-21T00:00:00Z arXiv:2606.20493: Contagion Networks — kako pristranost evaluatora zarazi cijeli multi-agent sustav https://24-ai.news/hr/news/2026-06-21/arxiv-diffusiongemma-interpretability/ 24 AI hr 2026-06-21T00:00:00Z arXiv:2606.20560: DiffusionGemma jednako čitljiv kao Gemma 4 — praznina od 28,6× srušena na 1,1× https://24-ai.news/hr/news/2026-06-21/arxiv-hreplan-cross-device-agent-recovery/ 24 AI hr 2026-06-21T00:00:00Z arXiv:2606.20487: H-RePlan — hijerarhijski oporavak AI agenata na više uređaja https://24-ai.news/hr/news/2026-06-21/arxiv-humanscale-egocentric-video-robot/ 24 AI hr 2026-06-21T00:00:00Z arXiv:2606.20521: Egocentrični video čovjeka nadmašuje robotske podatke za predobuku utjelovljenih AI modela https://24-ai.news/hr/news/2026-06-21/arxiv-neuroimprint-peft-privacy-backdoor/ 24 AI hr 2026-06-21T00:00:00Z arXiv:2606.20553: NeuroImprint — skriveni backdoor u federiranom fine-tuningu rekonstruira 59–79% podataka https://24-ai.news/hr/news/2026-06-21/arxiv-ssd-speculative-image-generation/ 24 AI hr 2026-06-21T00:00:00Z arXiv:2606.20543: Prostorno spekulativno dekodiranje ubrzava generiranje slika 13,3× https://24-ai.news/hr/news/2026-06-21/arxiv-timeprove-long-video-reasoning/ 24 AI hr 2026-06-21T00:00:00Z arXiv:2606.20561: TimeProVe smanjuje troškove zaključivanja nad dugim videozapisima za 93% https://24-ai.news/hr/news/2026-06-21/arxiv-ultraquant-4bit-kv-cache/ 24 AI hr 2026-06-21T00:00:00Z arXiv:2606.20474: UltraQuant smanjuje latenciju KV predmemorije za 3,47× s 4-bitnom preciznošću https://24-ai.news/hr/news/2026-06-21/arxiv-vimpo-critic-free-rl-reasoning/ 24 AI hr 2026-06-21T00:00:00Z arXiv:2606.20008: VIMPO — pojačano učenje bez kritičara pobjeđuje GRPO na MATH-500 i AIME https://24-ai.news/en/news/2026-06-21/anthropic-claude-code-2-1-185-stall-hint/ 24 AI en 2026-06-21T00:00:00Z Anthropic (Claude Code GitHub): Claude Code v2.1.185 improves stream stall message https://24-ai.news/en/news/2026-06-21/arxiv-activation-directions-misalignment/ 24 AI en 2026-06-21T00:00:00Z arXiv:2606.20225: Activation Directions Detect LLM Misalignment with 99.6% Accuracy https://24-ai.news/en/news/2026-06-21/arxiv-contagion-networks-evaluator-bias/ 24 AI en 2026-06-21T00:00:00Z arXiv:2606.20493: Contagion Networks — how evaluator bias infects an entire multi-agent system https://24-ai.news/en/news/2026-06-21/arxiv-diffusiongemma-interpretability/ 24 AI en 2026-06-21T00:00:00Z arXiv:2606.20560: DiffusionGemma as interpretable as Gemma 4 — 28.6× gap reduced to 1.1× https://24-ai.news/en/news/2026-06-21/arxiv-hreplan-cross-device-agent-recovery/ 24 AI en 2026-06-21T00:00:00Z arXiv:2606.20487: H-RePlan — Hierarchical Recovery of AI Agents Across Multiple Devices https://24-ai.news/en/news/2026-06-21/arxiv-humanscale-egocentric-video-robot/ 24 AI en 2026-06-21T00:00:00Z arXiv:2606.20521: Human Egocentric Video Outperforms Robot Data for Pre-training Embodied AI Models https://24-ai.news/en/news/2026-06-21/arxiv-neuroimprint-peft-privacy-backdoor/ 24 AI en 2026-06-21T00:00:00Z arXiv:2606.20553: NeuroImprint — hidden backdoor in federated fine-tuning reconstructs 59–79% of training data https://24-ai.news/en/news/2026-06-21/arxiv-ssd-speculative-image-generation/ 24 AI en 2026-06-21T00:00:00Z arXiv:2606.20543: Spatially Speculative Decoding accelerates image generation 13.3× https://24-ai.news/en/news/2026-06-21/arxiv-timeprove-long-video-reasoning/ 24 AI en 2026-06-21T00:00:00Z arXiv:2606.20561: TimeProVe Reduces Long-Video Reasoning Inference Costs by 93% https://24-ai.news/en/news/2026-06-21/arxiv-ultraquant-4bit-kv-cache/ 24 AI en 2026-06-21T00:00:00Z arXiv:2606.20474: UltraQuant Reduces KV Cache Latency by 3.47× with 4-Bit Precision https://24-ai.news/en/news/2026-06-21/arxiv-vimpo-critic-free-rl-reasoning/ 24 AI en 2026-06-21T00:00:00Z arXiv:2606.20008: VIMPO — Critic-Free Reinforcement Learning Beats GRPO on MATH-500 and AIME https://24-ai.news/de/news/2026-06-21/anthropic-claude-code-2-1-185-stall-hint/ 24 AI de 2026-06-21T00:00:00Z Anthropic (Claude Code GitHub): Claude Code v2.1.185 verbessert die Stream-Stall-Meldung https://24-ai.news/de/news/2026-06-21/arxiv-activation-directions-misalignment/ 24 AI de 2026-06-21T00:00:00Z arXiv:2606.20225: Aktivierungsrichtungen erkennen Fehlausrichtung von LLMs mit 99,6 % Genauigkeit https://24-ai.news/de/news/2026-06-21/arxiv-contagion-networks-evaluator-bias/ 24 AI de 2026-06-21T00:00:00Z arXiv:2606.20493: Contagion Networks — wie die Verzerrung eines Evaluators das gesamte Multi-Agenten-System infiziert https://24-ai.news/de/news/2026-06-21/arxiv-diffusiongemma-interpretability/ 24 AI de 2026-06-21T00:00:00Z arXiv:2606.20560: DiffusionGemma ebenso interpretierbar wie Gemma 4 — 28,6×-Lücke auf 1,1× reduziert https://24-ai.news/de/news/2026-06-21/arxiv-hreplan-cross-device-agent-recovery/ 24 AI de 2026-06-21T00:00:00Z arXiv:2606.20487: H-RePlan — Hierarchische Wiederherstellung von KI-Agenten auf mehreren Geräten https://24-ai.news/de/news/2026-06-21/arxiv-humanscale-egocentric-video-robot/ 24 AI de 2026-06-21T00:00:00Z arXiv:2606.20521: Egozentrische Menschenvideos übertreffen Roboterdaten beim Vortraining von verkörperten KI-Modellen https://24-ai.news/de/news/2026-06-21/arxiv-neuroimprint-peft-privacy-backdoor/ 24 AI de 2026-06-21T00:00:00Z arXiv:2606.20553: NeuroImprint — versteckte Hintertür im föderalen Fine-Tuning rekonstruiert 59–79 % der Trainingsdaten https://24-ai.news/de/news/2026-06-21/arxiv-ssd-speculative-image-generation/ 24 AI de 2026-06-21T00:00:00Z arXiv:2606.20543: Räumliches spekulatives Dekodieren beschleunigt Bildgenerierung um das 13,3-Fache https://24-ai.news/de/news/2026-06-21/arxiv-timeprove-long-video-reasoning/ 24 AI de 2026-06-21T00:00:00Z arXiv:2606.20561: TimeProVe reduziert Inferenzkosten bei langen Videos um 93 % https://24-ai.news/de/news/2026-06-21/arxiv-ultraquant-4bit-kv-cache/ 24 AI de 2026-06-21T00:00:00Z arXiv:2606.20474: UltraQuant reduziert KV-Cache-Latenz um das 3,47-Fache mit 4-Bit-Präzision https://24-ai.news/de/news/2026-06-21/arxiv-vimpo-critic-free-rl-reasoning/ 24 AI de 2026-06-21T00:00:00Z arXiv:2606.20008: VIMPO — Kritiker-freies Reinforcement Learning übertrifft GRPO bei MATH-500 und AIME https://24-ai.news/zh/news/2026-06-21/anthropic-claude-code-2-1-185-stall-hint/ 24 AI zh 2026-06-21T00:00:00Z Anthropic（Claude Code GitHub）：Claude Code v2.1.185 改进流卡顿提示信息 https://24-ai.news/zh/news/2026-06-21/arxiv-activation-directions-misalignment/ 24 AI zh 2026-06-21T00:00:00Z arXiv:2606.20225: 激活方向以99.6%的准确率检测大语言模型的对齐偏差 https://24-ai.news/zh/news/2026-06-21/arxiv-contagion-networks-evaluator-bias/ 24 AI zh 2026-06-21T00:00:00Z arXiv:2606.20493: Contagion Networks —评估者偏差如何感染整个多智能体系统 https://24-ai.news/zh/news/2026-06-21/arxiv-diffusiongemma-interpretability/ 24 AI zh 2026-06-21T00:00:00Z arXiv:2606.20560: DiffusionGemma 可解释性与 Gemma 4 相当——28.6× 差距缩小至 1.1× https://24-ai.news/zh/news/2026-06-21/arxiv-hreplan-cross-device-agent-recovery/ 24 AI zh 2026-06-21T00:00:00Z arXiv:2606.20487: H-RePlan — 跨设备AI智能体的分层故障恢复 https://24-ai.news/zh/news/2026-06-21/arxiv-humanscale-egocentric-video-robot/ 24 AI zh 2026-06-21T00:00:00Z arXiv:2606.20521: 人类自我中心视频在具身AI模型预训练中优于机器人数据 https://24-ai.news/zh/news/2026-06-21/arxiv-neuroimprint-peft-privacy-backdoor/ 24 AI zh 2026-06-21T00:00:00Z arXiv:2606.20553: NeuroImprint——联邦微调中的隐藏后门可重建59–79%的训练数据 https://24-ai.news/zh/news/2026-06-21/arxiv-ssd-speculative-image-generation/ 24 AI zh 2026-06-21T00:00:00Z arXiv:2606.20543: 空间投机解码将图像生成速度提升13.3倍 https://24-ai.news/zh/news/2026-06-21/arxiv-timeprove-long-video-reasoning/ 24 AI zh 2026-06-21T00:00:00Z arXiv:2606.20561: TimeProVe将长视频推理推断成本降低93% https://24-ai.news/zh/news/2026-06-21/arxiv-ultraquant-4bit-kv-cache/ 24 AI zh 2026-06-21T00:00:00Z arXiv:2606.20474: UltraQuant通过4位精度将KV缓存延迟降低3.47倍 https://24-ai.news/zh/news/2026-06-21/arxiv-vimpo-critic-free-rl-reasoning/ 24 AI zh 2026-06-21T00:00:00Z arXiv:2606.20008: VIMPO — 无评论家强化学习在MATH-500和AIME上超越GRPO https://24-ai.news/ja/news/2026-06-21/anthropic-claude-code-2-1-185-stall-hint/ 24 AI ja 2026-06-21T00:00:00Z Anthropic（Claude Code GitHub）：Claude Code v2.1.185 がストリーム停止メッセージを改善しました https://24-ai.news/ja/news/2026-06-21/arxiv-activation-directions-misalignment/ 24 AI ja 2026-06-21T00:00:00Z arXiv:2606.20225: 活性化方向がLLMの誤整合を99.6%の精度で検出 https://24-ai.news/ja/news/2026-06-21/arxiv-contagion-networks-evaluator-bias/ 24 AI ja 2026-06-21T00:00:00Z arXiv:2606.20493: Contagion Networks — 評価者のバイアスがマルチエージェントシステム全体に伝染する仕組み https://24-ai.news/ja/news/2026-06-21/arxiv-diffusiongemma-interpretability/ 24 AI ja 2026-06-21T00:00:00Z arXiv:2606.20560: DiffusionGemma の解釈可能性が Gemma 4 と同等に——28.6× の差が 1.1× へ縮小 https://24-ai.news/ja/news/2026-06-21/arxiv-hreplan-cross-device-agent-recovery/ 24 AI ja 2026-06-21T00:00:00Z arXiv:2606.20487: H-RePlan — 複数デバイスにまたがるAIエージェントの階層的障害回復 https://24-ai.news/ja/news/2026-06-21/arxiv-humanscale-egocentric-video-robot/ 24 AI ja 2026-06-21T00:00:00Z arXiv:2606.20521: 人間の一人称視点動画はロボットデータを上回る——身体的AIモデルの事前学習において https://24-ai.news/ja/news/2026-06-21/arxiv-neuroimprint-peft-privacy-backdoor/ 24 AI ja 2026-06-21T00:00:00Z arXiv:2606.20553: NeuroImprint——連合ファインチューニングの隠れバックドアが訓練データの59–79%を再構築 https://24-ai.news/ja/news/2026-06-21/arxiv-ssd-speculative-image-generation/ 24 AI ja 2026-06-21T00:00:00Z arXiv:2606.20543: 空間的投機的デコーディングが画像生成を13.3倍高速化 https://24-ai.news/ja/news/2026-06-21/arxiv-timeprove-long-video-reasoning/ 24 AI ja 2026-06-21T00:00:00Z arXiv:2606.20561: TimeProVeが長時間動画の推論コストを93%削減 https://24-ai.news/ja/news/2026-06-21/arxiv-ultraquant-4bit-kv-cache/ 24 AI ja 2026-06-21T00:00:00Z arXiv:2606.20474: UltraQuantが4ビット精度でKVキャッシュのレイテンシを3.47倍削減 https://24-ai.news/ja/news/2026-06-21/arxiv-vimpo-critic-free-rl-reasoning/ 24 AI ja 2026-06-21T00:00:00Z arXiv:2606.20008: VIMPO — クリティック不要の強化学習がMATH-500とAIMEでGRPOを上回る https://24-ai.news/ko/news/2026-06-21/anthropic-claude-code-2-1-185-stall-hint/ 24 AI ko 2026-06-21T00:00:00Z Anthropic（Claude Code GitHub）：Claude Code v2.1.185, 스트림 중단 메시지 개선 https://24-ai.news/ko/news/2026-06-21/arxiv-activation-directions-misalignment/ 24 AI ko 2026-06-21T00:00:00Z arXiv:2606.20225: 활성화 방향이 99.6% 정확도로 LLM 정렬 오류를 감지합니다 https://24-ai.news/ko/news/2026-06-21/arxiv-contagion-networks-evaluator-bias/ 24 AI ko 2026-06-21T00:00:00Z arXiv:2606.20493: Contagion Networks — 평가자 편향이 전체 멀티에이전트 시스템을 감염시키는 방식 https://24-ai.news/ko/news/2026-06-21/arxiv-diffusiongemma-interpretability/ 24 AI ko 2026-06-21T00:00:00Z arXiv:2606.20560: DiffusionGemma의 해석 가능성이 Gemma 4와 동등——28.6× 격차가 1.1×로 축소 https://24-ai.news/ko/news/2026-06-21/arxiv-hreplan-cross-device-agent-recovery/ 24 AI ko 2026-06-21T00:00:00Z arXiv:2606.20487: H-RePlan — 다중 디바이스 AI 에이전트의 계층적 장애 복구 https://24-ai.news/ko/news/2026-06-21/arxiv-humanscale-egocentric-video-robot/ 24 AI ko 2026-06-21T00:00:00Z arXiv:2606.20521: 인간 에고센트릭 비디오가 로봇 데이터를 능가한다 — 구현된 AI 모델 사전 학습에서 https://24-ai.news/ko/news/2026-06-21/arxiv-neuroimprint-peft-privacy-backdoor/ 24 AI ko 2026-06-21T00:00:00Z arXiv:2606.20553: NeuroImprint——연합 파인튜닝의 숨겨진 백도어가 훈련 데이터의 59–79%를 재구성 https://24-ai.news/ko/news/2026-06-21/arxiv-ssd-speculative-image-generation/ 24 AI ko 2026-06-21T00:00:00Z arXiv:2606.20543: 공간 투기적 디코딩으로 이미지 생성 13.3배 가속 https://24-ai.news/ko/news/2026-06-21/arxiv-timeprove-long-video-reasoning/ 24 AI ko 2026-06-21T00:00:00Z arXiv:2606.20561: TimeProVe, 장시간 동영상 추론 비용 93% 절감 https://24-ai.news/ko/news/2026-06-21/arxiv-ultraquant-4bit-kv-cache/ 24 AI ko 2026-06-21T00:00:00Z arXiv:2606.20474: UltraQuant, 4비트 정밀도로 KV 캐시 지연 시간 3.47배 단축 https://24-ai.news/ko/news/2026-06-21/arxiv-vimpo-critic-free-rl-reasoning/ 24 AI ko 2026-06-21T00:00:00Z arXiv:2606.20008: VIMPO — 비평가 없는 강화학습이 MATH-500과 AIME에서 GRPO를 능가하다 https://24-ai.news/hr/news/2026-06-20/amd-rocm-matrix3d-world-generation/ 24 AI hr 2026-06-20T00:00:00Z AMD: ROCm optimizacija Matrix3D-a za 3D svjetove ubrzala renderiranje do 54 posto na Instinct GPU-ima https://24-ai.news/hr/news/2026-06-20/anthropic-claude-code-2-1-183-safety/ 24 AI hr 2026-06-20T00:00:00Z Anthropic: Claude Code v2.1.183 blokira destruktivne git i infrastrukturne naredbe u auto modu https://24-ai.news/hr/news/2026-06-20/arxiv-llm-psychology-profiles-artifact/ 24 AI hr 2026-06-20T00:00:00Z arXiv:2606.20205: psihološki profili jezičnih modela uglavnom su mjerni artefakt, a ne stabilna osobnost https://24-ai.news/hr/news/2026-06-20/arxiv-mixed-compliance-demonstrations/ 24 AI hr 2026-06-20T00:00:00Z arXiv:2606.20508: što jezični modeli uče iz miješanih demonstracija sigurnog i štetnog ponašanja https://24-ai.news/hr/news/2026-06-20/arxiv-multi-lcb-livecodebench-12-languages/ 24 AI hr 2026-06-20T00:00:00Z arXiv:2606.20517: Multi-LCB proširuje LiveCodeBench na 12 programskih jezika i otkriva Python overfitting kod 24 modela https://24-ai.news/hr/news/2026-06-20/arxiv-sevra-selective-verification/ 24 AI hr 2026-06-20T00:00:00Z arXiv:2606.19808: SEVRA selektivnom verifikacijom štedi do 91 posto tokena u rasuđivanju modela https://24-ai.news/hr/news/2026-06-20/arxiv-softskill-behavioral-compression/ 24 AI hr 2026-06-20T00:00:00Z arXiv:2606.20333: SoftSkill komprimira skill-dokumente u 32 latentna tokena i diže LiveMath za 42,1 boda https://24-ai.news/hr/news/2026-06-20/aws-sagemaker-inference-metrics-dashboard/ 24 AI hr 2026-06-20T00:00:00Z AWS: SageMaker dobiva preko 100 detaljnih inference metrika i Insights nadzornu ploču na CloudWatchu https://24-ai.news/hr/news/2026-06-20/eu-ai-office-europa-frontier-challenge/ 24 AI hr 2026-06-20T00:00:00Z EU AI Office: konzorcij EUROPA gradi otvoreni europski frontier model s preko 400 milijardi parametara za svih 24 jezika EU-a https://24-ai.news/hr/news/2026-06-20/github-copilot-changelog-opus-46-sunset/ 24 AI hr 2026-06-20T00:00:00Z GitHub: Copilot gasi Opus 4.6 (fast) 29. lipnja, dodaje AGENTS.md u code review i polje ai_credits_used u API https://24-ai.news/hr/news/2026-06-20/github-qubot-data-analytics-agent/ 24 AI hr 2026-06-20T00:00:00Z GitHub: interni analitički agent Qubot smanjio vrijeme rješavanja upita za oko 66 posto https://24-ai.news/hr/news/2026-06-20/nvidia-agentic-marketing-cannes-lions/ 24 AI hr 2026-06-20T00:00:00Z NVIDIA: partneri na Cannes Lions 2026 pokazali agentski AI za marketing, Criteo udvostručio brzinu treninga na Blackwellu https://24-ai.news/hr/news/2026-06-20/uk-aisi-engineering-playbook/ 24 AI hr 2026-06-20T00:00:00Z UK AISI: Engineering Playbook otvara infrastrukturu za evaluaciju frontier modela u pet slojeva https://24-ai.news/en/news/2026-06-20/amd-rocm-matrix3d-world-generation/ 24 AI en 2026-06-20T00:00:00Z AMD: ROCm Optimization of Matrix3D for 3D Worlds Speeds Up Rendering by up to 54 Percent on Instinct GPUs https://24-ai.news/en/news/2026-06-20/anthropic-claude-code-2-1-183-safety/ 24 AI en 2026-06-20T00:00:00Z Anthropic: Claude Code v2.1.183 Blocks Destructive Git and Infrastructure Commands in Auto Mode https://24-ai.news/en/news/2026-06-20/arxiv-llm-psychology-profiles-artifact/ 24 AI en 2026-06-20T00:00:00Z arXiv:2606.20205: Psychological Profiles of LLMs Are Largely a Measurement Artifact, Not a Stable Personality https://24-ai.news/en/news/2026-06-20/arxiv-mixed-compliance-demonstrations/ 24 AI en 2026-06-20T00:00:00Z arXiv:2606.20508: What Language Models Learn from Mixed Demonstrations of Safe and Harmful Behavior https://24-ai.news/en/news/2026-06-20/arxiv-multi-lcb-livecodebench-12-languages/ 24 AI en 2026-06-20T00:00:00Z arXiv:2606.20517: Multi-LCB Extends LiveCodeBench to 12 Programming Languages and Reveals Python Overfitting in 24 Models https://24-ai.news/en/news/2026-06-20/arxiv-sevra-selective-verification/ 24 AI en 2026-06-20T00:00:00Z arXiv:2606.19808: SEVRA Saves up to 91 Percent of Tokens Through Selective Verification in Model Reasoning https://24-ai.news/en/news/2026-06-20/arxiv-softskill-behavioral-compression/ 24 AI en 2026-06-20T00:00:00Z arXiv:2606.20333: SoftSkill Compresses Skill Documents into 32 Latent Tokens and Boosts LiveMath by 42.1 Points https://24-ai.news/en/news/2026-06-20/aws-sagemaker-inference-metrics-dashboard/ 24 AI en 2026-06-20T00:00:00Z AWS: SageMaker Gets Over 100 Detailed Inference Metrics and an Insights Dashboard on CloudWatch https://24-ai.news/en/news/2026-06-20/eu-ai-office-europa-frontier-challenge/ 24 AI en 2026-06-20T00:00:00Z EU AI Office: The EUROPA Consortium Is Building an Open European Frontier Model with Over 400 Billion Parameters for All 24 EU Languages https://24-ai.news/en/news/2026-06-20/github-copilot-changelog-opus-46-sunset/ 24 AI en 2026-06-20T00:00:00Z GitHub: Copilot Retires Opus 4.6 (fast) on June 29, Adds AGENTS.md to Code Review and ai_credits_used Field to API https://24-ai.news/en/news/2026-06-20/github-qubot-data-analytics-agent/ 24 AI en 2026-06-20T00:00:00Z GitHub: Internal Analytics Agent Qubot Reduced Query Resolution Time by Around 66 Percent https://24-ai.news/en/news/2026-06-20/nvidia-agentic-marketing-cannes-lions/ 24 AI en 2026-06-20T00:00:00Z NVIDIA: Partners at Cannes Lions 2026 Showcased Agentic AI for Marketing, Criteo Doubled Training Speed on Blackwell https://24-ai.news/en/news/2026-06-20/uk-aisi-engineering-playbook/ 24 AI en 2026-06-20T00:00:00Z UK AISI: Engineering Playbook Opens Frontier Model Evaluation Infrastructure in Five Layers https://24-ai.news/de/news/2026-06-20/amd-rocm-matrix3d-world-generation/ 24 AI de 2026-06-20T00:00:00Z AMD: ROCm-Optimierung von Matrix3D für 3D-Welten beschleunigt Rendering auf Instinct-GPUs um bis zu 54 Prozent https://24-ai.news/de/news/2026-06-20/anthropic-claude-code-2-1-183-safety/ 24 AI de 2026-06-20T00:00:00Z Anthropic: Claude Code v2.1.183 blockiert destruktive Git- und Infrastrukturbefehle im Auto-Modus https://24-ai.news/de/news/2026-06-20/arxiv-llm-psychology-profiles-artifact/ 24 AI de 2026-06-20T00:00:00Z arXiv:2606.20205: Psychologische Profile von Sprachmodellen sind größtenteils ein Messartefakt, keine stabile Persönlichkeit https://24-ai.news/de/news/2026-06-20/arxiv-mixed-compliance-demonstrations/ 24 AI de 2026-06-20T00:00:00Z arXiv:2606.20508: Was Sprachmodelle aus gemischten Demonstrations sicheren und schädlichen Verhaltens lernen https://24-ai.news/de/news/2026-06-20/arxiv-multi-lcb-livecodebench-12-languages/ 24 AI de 2026-06-20T00:00:00Z arXiv:2606.20517: Multi-LCB erweitert LiveCodeBench auf 12 Programmiersprachen und deckt Python-Overfitting bei 24 Modellen auf https://24-ai.news/de/news/2026-06-20/arxiv-sevra-selective-verification/ 24 AI de 2026-06-20T00:00:00Z arXiv:2606.19808: SEVRA spart durch selektive Verifikation bis zu 91 Prozent der Tokens beim Modell-Reasoning https://24-ai.news/de/news/2026-06-20/arxiv-softskill-behavioral-compression/ 24 AI de 2026-06-20T00:00:00Z arXiv:2606.20333: SoftSkill komprimiert Skill-Dokumente in 32 latente Tokens und steigert LiveMath um 42,1 Punkte https://24-ai.news/de/news/2026-06-20/aws-sagemaker-inference-metrics-dashboard/ 24 AI de 2026-06-20T00:00:00Z AWS: SageMaker erhält über 100 detaillierte Inference-Metriken und Insights-Dashboard auf CloudWatch https://24-ai.news/de/news/2026-06-20/eu-ai-office-europa-frontier-challenge/ 24 AI de 2026-06-20T00:00:00Z EU AI Office: Konsortium EUROPA baut offenes europäisches Frontier-Modell mit über 400 Milliarden Parametern für alle 24 EU-Sprachen https://24-ai.news/de/news/2026-06-20/github-copilot-changelog-opus-46-sunset/ 24 AI de 2026-06-20T00:00:00Z GitHub: Copilot stellt Opus 4.6 (fast) am 29. Juni ein, fügt AGENTS.md zu Code Review hinzu und ai_credits_used-Feld zur API https://24-ai.news/de/news/2026-06-20/github-qubot-data-analytics-agent/ 24 AI de 2026-06-20T00:00:00Z GitHub: Interner Analyse-Agent Qubot verkürzte Abfragezeit um rund 66 Prozent https://24-ai.news/de/news/2026-06-20/nvidia-agentic-marketing-cannes-lions/ 24 AI de 2026-06-20T00:00:00Z NVIDIA: Partner auf Cannes Lions 2026 zeigten agentische KI für Marketing, Criteo verdoppelte Trainingsgeschwindigkeit auf Blackwell https://24-ai.news/de/news/2026-06-20/uk-aisi-engineering-playbook/ 24 AI de 2026-06-20T00:00:00Z UK AISI: Engineering Playbook öffnet Frontier-Modell-Evaluierungsinfrastruktur in fünf Schichten https://24-ai.news/zh/news/2026-06-20/amd-rocm-matrix3d-world-generation/ 24 AI zh 2026-06-20T00:00:00Z AMD：ROCm优化Matrix3D框架，在Instinct GPU上将3D世界渲染速度提升最高54% https://24-ai.news/zh/news/2026-06-20/anthropic-claude-code-2-1-183-safety/ 24 AI zh 2026-06-20T00:00:00Z Anthropic：Claude Code v2.1.183在自动模式下屏蔽破坏性git和基础设施命令 https://24-ai.news/zh/news/2026-06-20/arxiv-llm-psychology-profiles-artifact/ 24 AI zh 2026-06-20T00:00:00Z arXiv:2606.20205：语言模型的心理特征画像主要是测量偏差，而非稳定人格 https://24-ai.news/zh/news/2026-06-20/arxiv-mixed-compliance-demonstrations/ 24 AI zh 2026-06-20T00:00:00Z arXiv:2606.20508：语言模型从混合无害与有害示范中学到了什么 https://24-ai.news/zh/news/2026-06-20/arxiv-multi-lcb-livecodebench-12-languages/ 24 AI zh 2026-06-20T00:00:00Z arXiv:2606.20517：Multi-LCB将LiveCodeBench扩展至12种编程语言，揭示24个模型的Python过拟合 https://24-ai.news/zh/news/2026-06-20/arxiv-sevra-selective-verification/ 24 AI zh 2026-06-20T00:00:00Z arXiv:2606.19808：SEVRA通过选择性验证可节省模型推理最多91%的token https://24-ai.news/zh/news/2026-06-20/arxiv-softskill-behavioral-compression/ 24 AI zh 2026-06-20T00:00:00Z arXiv:2606.20333：SoftSkill将技能文档压缩为32个潜在token，LiveMath提升42.1分 https://24-ai.news/zh/news/2026-06-20/aws-sagemaker-inference-metrics-dashboard/ 24 AI zh 2026-06-20T00:00:00Z AWS：SageMaker获得100余项详细推理指标和CloudWatch Insights监控面板 https://24-ai.news/zh/news/2026-06-20/eu-ai-office-europa-frontier-challenge/ 24 AI zh 2026-06-20T00:00:00Z EU AI Office：EUROPA联盟建设支持24种欧盟语言的4000亿参数开放欧洲前沿模型 https://24-ai.news/zh/news/2026-06-20/github-copilot-changelog-opus-46-sunset/ 24 AI zh 2026-06-20T00:00:00Z GitHub：Copilot于6月29日停用Opus 4.6 (fast)，code review新增AGENTS.md，API增加ai_credits_used字段 https://24-ai.news/zh/news/2026-06-20/github-qubot-data-analytics-agent/ 24 AI zh 2026-06-20T00:00:00Z GitHub：内部分析智能体Qubot将查询解决时间缩短约66% https://24-ai.news/zh/news/2026-06-20/nvidia-agentic-marketing-cannes-lions/ 24 AI zh 2026-06-20T00:00:00Z NVIDIA：合作伙伴在2026年戛纳创意节展示营销智能体AI，Criteo在Blackwell上将训练速度提升2倍 https://24-ai.news/zh/news/2026-06-20/uk-aisi-engineering-playbook/ 24 AI zh 2026-06-20T00:00:00Z UK AISI：Engineering Playbook以五层结构开放前沿模型评估基础设施 https://24-ai.news/ja/news/2026-06-20/amd-rocm-matrix3d-world-generation/ 24 AI ja 2026-06-20T00:00:00Z AMD：ROCm最適化でMatrix3DのInstinct GPU向け3D世界生成が最大54%高速化 https://24-ai.news/ja/news/2026-06-20/anthropic-claude-code-2-1-183-safety/ 24 AI ja 2026-06-20T00:00:00Z Anthropic：Claude Code v2.1.183がオートモードで破壊的なgit・インフラコマンドをブロック https://24-ai.news/ja/news/2026-06-20/arxiv-llm-psychology-profiles-artifact/ 24 AI ja 2026-06-20T00:00:00Z arXiv:2606.20205：LLMの心理プロファイルは安定した人格ではなく測定アーティファクトである https://24-ai.news/ja/news/2026-06-20/arxiv-mixed-compliance-demonstrations/ 24 AI ja 2026-06-20T00:00:00Z arXiv:2606.20508：LLMは無害・有害デモンストレーションの混合から何を学ぶか https://24-ai.news/ja/news/2026-06-20/arxiv-multi-lcb-livecodebench-12-languages/ 24 AI ja 2026-06-20T00:00:00Z arXiv:2606.20517：Multi-LCBがLiveCodeBenchを12言語に拡張、24モデルのPython過学習を発見 https://24-ai.news/ja/news/2026-06-20/arxiv-sevra-selective-verification/ 24 AI ja 2026-06-20T00:00:00Z arXiv:2606.19808：SEVRAの選択的検証でモデル推論のトークンを最大91%削減 https://24-ai.news/ja/news/2026-06-20/arxiv-softskill-behavioral-compression/ 24 AI ja 2026-06-20T00:00:00Z arXiv:2606.20333：SoftSkillがスキル文書を32の潜在トークンに圧縮、LiveMathを42.1ポイント向上 https://24-ai.news/ja/news/2026-06-20/aws-sagemaker-inference-metrics-dashboard/ 24 AI ja 2026-06-20T00:00:00Z AWS：SageMakerが100以上の詳細な推論メトリクスとCloudWatch Insightsダッシュボードを取得 https://24-ai.news/ja/news/2026-06-20/eu-ai-office-europa-frontier-challenge/ 24 AI ja 2026-06-20T00:00:00Z EU AI Office：EUROPAコンソーシアム、EU24言語対応の4000億パラメータ欧州オープンフロンティアモデルを構築へ https://24-ai.news/ja/news/2026-06-20/github-copilot-changelog-opus-46-sunset/ 24 AI ja 2026-06-20T00:00:00Z GitHub：CopilotがOpus 4.6 (fast)を6月29日に廃止、code reviewにAGENTS.md追加、APIにai_credits_usedフィールド https://24-ai.news/ja/news/2026-06-20/github-qubot-data-analytics-agent/ 24 AI ja 2026-06-20T00:00:00Z GitHub：内部分析エージェント「Qubot」がクエリ解決時間を約66%短縮 https://24-ai.news/ja/news/2026-06-20/nvidia-agentic-marketing-cannes-lions/ 24 AI ja 2026-06-20T00:00:00Z NVIDIA：カンヌライオンズ2026でパートナーがマーケティング向けエージェントAIを披露、CriteoはBlackwellでトレーニング速度2倍 https://24-ai.news/ja/news/2026-06-20/uk-aisi-engineering-playbook/ 24 AI ja 2026-06-20T00:00:00Z UK AISI：Engineering Playbookでフロンティアモデル評価インフラを5層構造で公開 https://24-ai.news/ko/news/2026-06-20/amd-rocm-matrix3d-world-generation/ 24 AI ko 2026-06-20T00:00:00Z AMD：ROCm으로 Matrix3D 최적화, Instinct GPU에서 3D 세계 렌더링 최대 54% 가속 https://24-ai.news/ko/news/2026-06-20/anthropic-claude-code-2-1-183-safety/ 24 AI ko 2026-06-20T00:00:00Z Anthropic：Claude Code v2.1.183, 자동 모드에서 파괴적 git·인프라 명령 차단 https://24-ai.news/ko/news/2026-06-20/arxiv-llm-psychology-profiles-artifact/ 24 AI ko 2026-06-20T00:00:00Z arXiv:2606.20205：LLM의 심리 프로필은 안정된 인격이 아닌 측정 아티팩트 https://24-ai.news/ko/news/2026-06-20/arxiv-mixed-compliance-demonstrations/ 24 AI ko 2026-06-20T00:00:00Z arXiv:2606.20508：언어 모델은 무해·유해 데모의 혼합에서 무엇을 학습하는가 https://24-ai.news/ko/news/2026-06-20/arxiv-multi-lcb-livecodebench-12-languages/ 24 AI ko 2026-06-20T00:00:00Z arXiv:2606.20517：Multi-LCB, LiveCodeBench를 12개 프로그래밍 언어로 확장하고 24개 모델의 Python 과적합 발견 https://24-ai.news/ko/news/2026-06-20/arxiv-sevra-selective-verification/ 24 AI ko 2026-06-20T00:00:00Z arXiv:2606.19808：SEVRA, 선택적 검증으로 모델 추론 토큰 최대 91% 절감 https://24-ai.news/ko/news/2026-06-20/arxiv-softskill-behavioral-compression/ 24 AI ko 2026-06-20T00:00:00Z arXiv:2606.20333：SoftSkill이 스킬 문서를 32개 잠재 토큰으로 압축, LiveMath 42.1점 향상 https://24-ai.news/ko/news/2026-06-20/aws-sagemaker-inference-metrics-dashboard/ 24 AI ko 2026-06-20T00:00:00Z AWS：SageMaker에 100개 이상의 상세 추론 지표와 CloudWatch Insights 대시보드 추가 https://24-ai.news/ko/news/2026-06-20/eu-ai-office-europa-frontier-challenge/ 24 AI ko 2026-06-20T00:00:00Z EU AI Office：EUROPA 컨소시엄, EU 24개 언어 지원 4000억 파라미터 오픈 유럽 프런티어 모델 구축 https://24-ai.news/ko/news/2026-06-20/github-copilot-changelog-opus-46-sunset/ 24 AI ko 2026-06-20T00:00:00Z GitHub：Copilot이 6월 29일 Opus 4.6 (fast) 종료, code review에 AGENTS.md 추가, API에 ai_credits_used 필드 https://24-ai.news/ko/news/2026-06-20/github-qubot-data-analytics-agent/ 24 AI ko 2026-06-20T00:00:00Z GitHub：내부 분석 에이전트 Qubot이 쿼리 해결 시간 약 66% 단축 https://24-ai.news/ko/news/2026-06-20/nvidia-agentic-marketing-cannes-lions/ 24 AI ko 2026-06-20T00:00:00Z NVIDIA：2026 칸 라이언스에서 파트너들이 마케팅용 에이전트 AI 공개, Criteo는 Blackwell에서 훈련 속도 2배 https://24-ai.news/ko/news/2026-06-20/uk-aisi-engineering-playbook/ 24 AI ko 2026-06-20T00:00:00Z UK AISI：Engineering Playbook으로 프런티어 모델 평가 인프라를 5개 레이어로 공개