https://24-ai.news/hr/news/2026-06-28/amd-rocm-gpu-workload-preemption/ 24 AI hr 2026-06-28T00:00:00Z AMD: Resource Manager automatski istiskuje neaktivne GPU workloadove i vraća resurse u zajednički pool klastera https://24-ai.news/hr/news/2026-06-28/arxiv-cot-training-action-prediction/ 24 AI hr 2026-06-28T00:00:00Z arXiv:2606.26935: dobici CoT treninga slijevaju se u jače predviđanje akcije, a ne u dublje rezoniranje agenata https://24-ai.news/hr/news/2026-06-28/arxiv-evaf-parametric-agent-memory/ 24 AI hr 2026-06-28T00:00:00Z arXiv:2606.26806: EVAF konsolidira pamćenje agenata u parametre modela za postojanost cilja nakon brisanja konteksta https://24-ai.news/hr/news/2026-06-28/arxiv-reasoning-tokens-failure-divergence/ 24 AI hr 2026-06-28T00:00:00Z arXiv:2606.26502: reasoning modeli troše više tokena na zadatke koje pogriješe, suprotno od ljudi koji odustaju https://24-ai.news/hr/news/2026-06-28/github-copilot-agentic-harness-benchmark/ 24 AI hr 2026-06-28T00:00:00Z GitHub: Copilot agentic harness dostiže razinu vendor-harnessa uz manju potrošnju tokena kroz 20+ frontier modela https://24-ai.news/hr/news/2026-06-28/github-mai-code-1-flash-copilot-ga/ 24 AI hr 2026-06-28T00:00:00Z GitHub: MAI-Code-1-Flash, Microsoftov coding model, sada općedostupan u Copilot Business i Enterprise planovima https://24-ai.news/en/news/2026-06-28/amd-rocm-gpu-workload-preemption/ 24 AI en 2026-06-28T00:00:00Z AMD: Resource Manager automatically preempts inactive GPU workloads and returns resources to the cluster shared pool https://24-ai.news/en/news/2026-06-28/arxiv-cot-training-action-prediction/ 24 AI en 2026-06-28T00:00:00Z arXiv:2606.26935: CoT training gains land in stronger action prediction, not deeper agent reasoning https://24-ai.news/en/news/2026-06-28/arxiv-evaf-parametric-agent-memory/ 24 AI en 2026-06-28T00:00:00Z arXiv:2606.26806: EVAF consolidates agent memory into model parameters for goal persistence after context deletion https://24-ai.news/en/news/2026-06-28/arxiv-reasoning-tokens-failure-divergence/ 24 AI en 2026-06-28T00:00:00Z arXiv:2606.26502: reasoning models spend more tokens on tasks they fail, opposite to humans who disengage https://24-ai.news/en/news/2026-06-28/github-copilot-agentic-harness-benchmark/ 24 AI en 2026-06-28T00:00:00Z GitHub: Copilot agentic harness reaches vendor-harness level with lower token consumption across 20+ frontier models https://24-ai.news/en/news/2026-06-28/github-mai-code-1-flash-copilot-ga/ 24 AI en 2026-06-28T00:00:00Z GitHub: MAI-Code-1-Flash, Microsoft's coding model, now generally available in Copilot Business and Enterprise plans https://24-ai.news/de/news/2026-06-28/amd-rocm-gpu-workload-preemption/ 24 AI de 2026-06-28T00:00:00Z AMD: Resource Manager verdrängt automatisch inaktive GPU-Workloads und gibt Ressourcen an den Cluster-Pool zurück https://24-ai.news/de/news/2026-06-28/arxiv-cot-training-action-prediction/ 24 AI de 2026-06-28T00:00:00Z arXiv:2606.26935: CoT-Trainingsgewinne fließen in stärkere Aktionsvorhersage, nicht in tieferes Agenten-Reasoning https://24-ai.news/de/news/2026-06-28/arxiv-evaf-parametric-agent-memory/ 24 AI de 2026-06-28T00:00:00Z arXiv:2606.26806: EVAF konsolidiert Agentenspeicher in Modellparameter für Zielbeständigkeit nach Kontextlöschung https://24-ai.news/de/news/2026-06-28/arxiv-reasoning-tokens-failure-divergence/ 24 AI de 2026-06-28T00:00:00Z arXiv:2606.26502: Reasoning-Modelle verbrauchen mehr Tokens bei Fehlern – im Gegensatz zu Menschen, die aufgeben https://24-ai.news/de/news/2026-06-28/github-copilot-agentic-harness-benchmark/ 24 AI de 2026-06-28T00:00:00Z GitHub: Copilot Agentic Harness erreicht Vendor-Niveau mit geringerem Token-Verbrauch über 20+ Frontier-Modelle https://24-ai.news/de/news/2026-06-28/github-mai-code-1-flash-copilot-ga/ 24 AI de 2026-06-28T00:00:00Z GitHub: MAI-Code-1-Flash, Microsofts Coding-Modell, jetzt allgemein verfügbar in Copilot Business und Enterprise https://24-ai.news/zh/news/2026-06-28/amd-rocm-gpu-workload-preemption/ 24 AI zh 2026-06-28T00:00:00Z AMD: Resource Manager 自动抢占空闲 GPU 工作负载，将资源归还集群共享池 https://24-ai.news/zh/news/2026-06-28/arxiv-cot-training-action-prediction/ 24 AI zh 2026-06-28T00:00:00Z arXiv:2606.26935: CoT 训练收益流向更强的动作预测，而非更深的智能体推理 https://24-ai.news/zh/news/2026-06-28/arxiv-evaf-parametric-agent-memory/ 24 AI zh 2026-06-28T00:00:00Z arXiv:2606.26806: EVAF 将智能体记忆固化入模型参数，在上下文清除后保持目标持久性 https://24-ai.news/zh/news/2026-06-28/arxiv-reasoning-tokens-failure-divergence/ 24 AI zh 2026-06-28T00:00:00Z arXiv:2606.26502: 推理模型在出错任务上消耗更多 token，与人类放弃的行为相反 https://24-ai.news/zh/news/2026-06-28/github-copilot-agentic-harness-benchmark/ 24 AI zh 2026-06-28T00:00:00Z GitHub: Copilot 智能体 harness 在 20 余款前沿模型上达到供应商 harness 水平，且 token 消耗更低 https://24-ai.news/zh/news/2026-06-28/github-mai-code-1-flash-copilot-ga/ 24 AI zh 2026-06-28T00:00:00Z GitHub: MAI-Code-1-Flash，微软编程模型，现已在 Copilot Business 和 Enterprise 方案中正式发布 https://24-ai.news/ja/news/2026-06-28/amd-rocm-gpu-workload-preemption/ 24 AI ja 2026-06-28T00:00:00Z AMD: Resource Managerが非アクティブなGPUワークロードを自動的にプリエンプションしクラスタの共有プールにリソースを返還 https://24-ai.news/ja/news/2026-06-28/arxiv-cot-training-action-prediction/ 24 AI ja 2026-06-28T00:00:00Z arXiv:2606.26935：CoTトレーニングの恩恵はより深い推論ではなくより強い行動予測として現れる https://24-ai.news/ja/news/2026-06-28/arxiv-evaf-parametric-agent-memory/ 24 AI ja 2026-06-28T00:00:00Z arXiv:2606.26806：EVAFがエージェントの記憶をモデルパラメータに統合しコンテキスト削除後も目標を維持 https://24-ai.news/ja/news/2026-06-28/arxiv-reasoning-tokens-failure-divergence/ 24 AI ja 2026-06-28T00:00:00Z arXiv:2606.26502：推論モデルは失敗したタスクにより多くのトークンを消費し、諦める人間とは逆の傾向を示す https://24-ai.news/ja/news/2026-06-28/github-copilot-agentic-harness-benchmark/ 24 AI ja 2026-06-28T00:00:00Z GitHub: Copilot agenticハーネスが20以上のフロンティアモデルにわたってベンダーハーネス水準の性能を達成し、より少ないトークン消費を実現 https://24-ai.news/ja/news/2026-06-28/github-mai-code-1-flash-copilot-ga/ 24 AI ja 2026-06-28T00:00:00Z GitHub: MAI-Code-1-Flash、MicrosoftのコーディングモデルがCopilot BusinessおよびEnterpriseプランで一般提供開始 https://24-ai.news/ko/news/2026-06-28/amd-rocm-gpu-workload-preemption/ 24 AI ko 2026-06-28T00:00:00Z AMD: Resource Manager, 비활성 GPU 워크로드를 자동 선점하여 클러스터 공유 풀에 자원 반환 https://24-ai.news/ko/news/2026-06-28/arxiv-cot-training-action-prediction/ 24 AI ko 2026-06-28T00:00:00Z arXiv:2606.26935: CoT 훈련의 이득은 깊은 에이전트 추론이 아닌 행동 예측 강화로 귀결 https://24-ai.news/ko/news/2026-06-28/arxiv-evaf-parametric-agent-memory/ 24 AI ko 2026-06-28T00:00:00Z arXiv:2606.26806: EVAF, 컨텍스트 삭제 후에도 목표를 유지하는 에이전트 파라메트릭 메모리 통합 https://24-ai.news/ko/news/2026-06-28/arxiv-reasoning-tokens-failure-divergence/ 24 AI ko 2026-06-28T00:00:00Z arXiv:2606.26502: 추론 모델은 틀린 문제에 더 많은 토큰 소비 — 포기하는 사람과 반대 패턴 https://24-ai.news/ko/news/2026-06-28/github-copilot-agentic-harness-benchmark/ 24 AI ko 2026-06-28T00:00:00Z GitHub: Copilot 에이전트 하니스, 20개 이상의 프론티어 모델에서 벤더 하니스 수준 달성 및 토큰 사용량 절감 https://24-ai.news/ko/news/2026-06-28/github-mai-code-1-flash-copilot-ga/ 24 AI ko 2026-06-28T00:00:00Z GitHub: MAI-Code-1-Flash, 마이크로소프트의 코딩 모델, 이제 Copilot Business 및 Enterprise 플랜에서 일반 제공 https://24-ai.news/hr/news/2026-06-27/amd-mi355x-mxfp4-quantization/ 24 AI hr 2026-06-27T00:00:00Z AMD: MXFP4/MXFP6 mixed-precision kvantizacija na MI355X — do 29% veći throughput https://24-ai.news/hr/news/2026-06-27/anthropic-api-rate-limits-three-tiers/ 24 AI hr 2026-06-27T00:00:00Z Anthropic: API rate limiti podignuti — Sonnet i Haiku sad na razini Opusa, tri tiera https://24-ai.news/hr/news/2026-06-27/anthropic-economic-index-cadences/ 24 AI hr 2026-06-27T00:00:00Z Anthropic: Economic Index otkriva ritmove i obrasce korištenja AI-ja kroz dan https://24-ai.news/hr/news/2026-06-27/arxiv-autoformalization-agent-policy-code/ 24 AI hr 2026-06-27T00:00:00Z arXiv:2606.26649: Upute agentima pretvorene u formalno verificirani policy-as-code https://24-ai.news/hr/news/2026-06-27/arxiv-capability-frontier-benchmarks/ 24 AI hr 2026-06-27T00:00:00Z arXiv:2606.26836: Benchmarci propuštaju 82% stvarnih sposobnosti AI modela https://24-ai.news/hr/news/2026-06-27/arxiv-co-failure-ceiling-llm-ensembles/ 24 AI hr 2026-06-27T00:00:00Z arXiv:2606.27288: Kad kombiniranje LLM-ova stvarno pomaže — co-failure ceiling na 67 frontier modela https://24-ai.news/hr/news/2026-06-27/arxiv-egg-expert-guided-kernels/ 24 AI hr 2026-06-27T00:00:00Z arXiv:2606.26758: EGG — višeagentni okvir generira GPU kernele 2,13× brže od PyTorcha https://24-ai.news/hr/news/2026-06-27/arxiv-leanguard-fast-content-moderation/ 24 AI hr 2026-06-27T00:00:00Z arXiv:2606.26686: LeanGuard — brza moderacija sadržaja bez chain-of-thought dostiže teške reasonere https://24-ai.news/hr/news/2026-06-27/arxiv-semantic-early-stopping-agents/ 24 AI hr 2026-06-27T00:00:00Z arXiv:2606.27009: Semantičko rano zaustavljanje smanjuje trošak agentskih petlji za 38% https://24-ai.news/hr/news/2026-06-27/aws-stripe-ai-agents-financial-compliance/ 24 AI hr 2026-06-27T00:00:00Z AWS: Stripe pokrenuo 100+ AI agenata za financijsku usklađenost — lekcije iz produkcije https://24-ai.news/hr/news/2026-06-27/claude-code-2-1-195/ 24 AI hr 2026-06-27T00:00:00Z Anthropic: Claude Code v2.1.195 — onemogućavanje miša, poboljšan voice dictation i popravci hookova https://24-ai.news/hr/news/2026-06-27/github-desktop-3-6-worktrees-copilot/ 24 AI hr 2026-06-27T00:00:00Z GitHub: Desktop 3.6 donosi Git worktrees i AI razrješavanje merge konflikata https://24-ai.news/hr/news/2026-06-27/google-gemini-nano-frozen-mtp-pixel/ 24 AI hr 2026-06-27T00:00:00Z Google: Gemini Nano na Pixelu 50%+ brži uz zamrznutu multi-token predikciju https://24-ai.news/hr/news/2026-06-27/langchain-deep-agents-prompt-caching/ 24 AI hr 2026-06-27T00:00:00Z LangChain: prompt caching u Deep Agents okviru smanjuje latenciju dugotrajnih agenata https://24-ai.news/hr/news/2026-06-27/openai-gpt-5-6-sol-preview/ 24 AI hr 2026-06-27T00:00:00Z OpenAI: GPT-5.6 Sol najavljen u previewu — kodiranje, znanost i kibersigurnost https://24-ai.news/en/news/2026-06-27/amd-mi355x-mxfp4-quantization/ 24 AI en 2026-06-27T00:00:00Z AMD: MXFP4/MXFP6 mixed-precision quantization on MI355X delivers up to 29% higher throughput https://24-ai.news/en/news/2026-06-27/anthropic-api-rate-limits-three-tiers/ 24 AI en 2026-06-27T00:00:00Z Anthropic: API rate limits raised — Sonnet and Haiku now match Opus across three tiers https://24-ai.news/en/news/2026-06-27/anthropic-economic-index-cadences/ 24 AI en 2026-06-27T00:00:00Z Anthropic: Economic Index reveals daily rhythms and patterns of AI usage https://24-ai.news/en/news/2026-06-27/arxiv-autoformalization-agent-policy-code/ 24 AI en 2026-06-27T00:00:00Z arXiv:2606.26649: Agent instructions converted into formally verified policy-as-code https://24-ai.news/en/news/2026-06-27/arxiv-capability-frontier-benchmarks/ 24 AI en 2026-06-27T00:00:00Z arXiv:2606.26836: Benchmarks miss 82% of AI models' real capabilities https://24-ai.news/en/news/2026-06-27/arxiv-co-failure-ceiling-llm-ensembles/ 24 AI en 2026-06-27T00:00:00Z arXiv:2606.27288: When combining LLMs really helps — co-failure ceiling across 67 frontier models https://24-ai.news/en/news/2026-06-27/arxiv-egg-expert-guided-kernels/ 24 AI en 2026-06-27T00:00:00Z arXiv:2606.26758: EGG — multi-agent framework generates GPU kernels 2.13x faster than PyTorch https://24-ai.news/en/news/2026-06-27/arxiv-leanguard-fast-content-moderation/ 24 AI en 2026-06-27T00:00:00Z arXiv:2606.26686: LeanGuard — fast content moderation without chain-of-thought matches heavy reasoners https://24-ai.news/en/news/2026-06-27/arxiv-semantic-early-stopping-agents/ 24 AI en 2026-06-27T00:00:00Z arXiv:2606.27009: Semantic early stopping cuts agent loop token cost by 38% https://24-ai.news/en/news/2026-06-27/aws-stripe-ai-agents-financial-compliance/ 24 AI en 2026-06-27T00:00:00Z AWS: Stripe launched 100+ AI agents for financial compliance — lessons from production https://24-ai.news/en/news/2026-06-27/claude-code-2-1-195/ 24 AI en 2026-06-27T00:00:00Z Anthropic: Claude Code v2.1.195 — mouse disabling, improved voice dictation, and hook fixes https://24-ai.news/en/news/2026-06-27/github-desktop-3-6-worktrees-copilot/ 24 AI en 2026-06-27T00:00:00Z GitHub: Desktop 3.6 brings Git worktrees and AI merge conflict resolution https://24-ai.news/en/news/2026-06-27/google-gemini-nano-frozen-mtp-pixel/ 24 AI en 2026-06-27T00:00:00Z Google: Gemini Nano on Pixel is 50%+ faster with frozen multi-token prediction https://24-ai.news/en/news/2026-06-27/langchain-deep-agents-prompt-caching/ 24 AI en 2026-06-27T00:00:00Z LangChain: prompt caching in Deep Agents framework reduces long-running agent latency https://24-ai.news/en/news/2026-06-27/openai-gpt-5-6-sol-preview/ 24 AI en 2026-06-27T00:00:00Z OpenAI: GPT-5.6 Sol announced in preview — coding, science, and cybersecurity https://24-ai.news/de/news/2026-06-27/amd-mi355x-mxfp4-quantization/ 24 AI de 2026-06-27T00:00:00Z AMD: MXFP4/MXFP6 Mixed-Precision-Quantisierung auf MI355X — bis zu 29% mehr Durchsatz https://24-ai.news/de/news/2026-06-27/anthropic-api-rate-limits-three-tiers/ 24 AI de 2026-06-27T00:00:00Z Anthropic: API-Rate-Limits angehoben — Sonnet und Haiku jetzt auf Opus-Niveau, drei Stufen https://24-ai.news/de/news/2026-06-27/anthropic-economic-index-cadences/ 24 AI de 2026-06-27T00:00:00Z Anthropic: Economic Index zeigt Tagesrhythmen und Nutzungsmuster bei KI https://24-ai.news/de/news/2026-06-27/arxiv-autoformalization-agent-policy-code/ 24 AI de 2026-06-27T00:00:00Z arXiv:2606.26649: Agentenanweisungen werden zu formal verifiziertem Policy-as-Code https://24-ai.news/de/news/2026-06-27/arxiv-capability-frontier-benchmarks/ 24 AI de 2026-06-27T00:00:00Z arXiv:2606.26836: Benchmarks übersehen 82% der tatsächlichen KI-Modellfähigkeiten https://24-ai.news/de/news/2026-06-27/arxiv-co-failure-ceiling-llm-ensembles/ 24 AI de 2026-06-27T00:00:00Z arXiv:2606.27288: Wann die Kombination von Sprachmodellen wirklich hilft — Co-Failure Ceiling bei 67 Frontier-Modellen https://24-ai.news/de/news/2026-06-27/arxiv-egg-expert-guided-kernels/ 24 AI de 2026-06-27T00:00:00Z arXiv:2606.26758: EGG — Mehrагenten-Framework generiert GPU-Kernel 2,13× schneller als PyTorch https://24-ai.news/de/news/2026-06-27/arxiv-leanguard-fast-content-moderation/ 24 AI de 2026-06-27T00:00:00Z arXiv:2606.26686: LeanGuard — schnelle Inhaltsmoderation ohne Chain-of-Thought erreicht schwere Reasoning-Modelle https://24-ai.news/de/news/2026-06-27/arxiv-semantic-early-stopping-agents/ 24 AI de 2026-06-27T00:00:00Z arXiv:2606.27009: Semantisches frühzeitiges Stoppen senkt Kosten von Agenten-Schleifen um 38% https://24-ai.news/de/news/2026-06-27/aws-stripe-ai-agents-financial-compliance/ 24 AI de 2026-06-27T00:00:00Z AWS: Stripe hat über 100 KI-Agenten für die Finanzcompliance eingeführt — Erkenntnisse aus dem Produktionsbetrieb https://24-ai.news/de/news/2026-06-27/claude-code-2-1-195/ 24 AI de 2026-06-27T00:00:00Z Anthropic: Claude Code v2.1.195 — Maus deaktivieren, verbessertes Voice Dictation und Hook-Korrekturen https://24-ai.news/de/news/2026-06-27/github-desktop-3-6-worktrees-copilot/ 24 AI de 2026-06-27T00:00:00Z GitHub: Desktop 3.6 bringt Git Worktrees und KI-gestützte Merge-Konfliktlösung https://24-ai.news/de/news/2026-06-27/google-gemini-nano-frozen-mtp-pixel/ 24 AI de 2026-06-27T00:00:00Z Google: Gemini Nano auf Pixel 50%+ schneller durch eingefrorene Multi-Token-Prediktion https://24-ai.news/de/news/2026-06-27/langchain-deep-agents-prompt-caching/ 24 AI de 2026-06-27T00:00:00Z LangChain: Prompt Caching im Deep Agents-Framework senkt Latenz langlebiger Agenten https://24-ai.news/de/news/2026-06-27/openai-gpt-5-6-sol-preview/ 24 AI de 2026-06-27T00:00:00Z OpenAI: GPT-5.6 Sol im Preview angekündigt — Kodierung, Wissenschaft und Cybersicherheit https://24-ai.news/zh/news/2026-06-27/amd-mi355x-mxfp4-quantization/ 24 AI zh 2026-06-27T00:00:00Z AMD：MI355X上的MXFP4/MXFP6混合精度量化——吞吐量最高提升29% https://24-ai.news/zh/news/2026-06-27/anthropic-api-rate-limits-three-tiers/ 24 AI zh 2026-06-27T00:00:00Z Anthropic：API速率限制提升——Sonnet和Haiku与Opus同等，三个层级 https://24-ai.news/zh/news/2026-06-27/anthropic-economic-index-cadences/ 24 AI zh 2026-06-27T00:00:00Z Anthropic经济指数：揭示全天AI使用的节律与模式 https://24-ai.news/zh/news/2026-06-27/arxiv-autoformalization-agent-policy-code/ 24 AI zh 2026-06-27T00:00:00Z arXiv:2606.26649: AI代理指令转换为形式化验证的策略即代码 https://24-ai.news/zh/news/2026-06-27/arxiv-capability-frontier-benchmarks/ 24 AI zh 2026-06-27T00:00:00Z arXiv:2606.26836: 标准基准遗漏了82%的AI模型真实能力 https://24-ai.news/zh/news/2026-06-27/arxiv-co-failure-ceiling-llm-ensembles/ 24 AI zh 2026-06-27T00:00:00Z arXiv:2606.27288: 组合LLM何时真正有效——67个前沿模型的共同失败上限 https://24-ai.news/zh/news/2026-06-27/arxiv-egg-expert-guided-kernels/ 24 AI zh 2026-06-27T00:00:00Z arXiv:2606.26758: EGG——多代理框架生成GPU内核比PyTorch快2.13倍 https://24-ai.news/zh/news/2026-06-27/arxiv-leanguard-fast-content-moderation/ 24 AI zh 2026-06-27T00:00:00Z arXiv:2606.26686: LeanGuard——无需思维链的快速内容审核媲美重量级推理器 https://24-ai.news/zh/news/2026-06-27/arxiv-semantic-early-stopping-agents/ 24 AI zh 2026-06-27T00:00:00Z arXiv:2606.27009: 语义早停将代理循环成本降低38% https://24-ai.news/zh/news/2026-06-27/aws-stripe-ai-agents-financial-compliance/ 24 AI zh 2026-06-27T00:00:00Z AWS：Stripe部署100+个AI代理用于金融合规——来自生产环境的经验教训 https://24-ai.news/zh/news/2026-06-27/claude-code-2-1-195/ 24 AI zh 2026-06-27T00:00:00Z Anthropic：Claude Code v2.1.195——禁用鼠标、改进语音听写和钩子修复 https://24-ai.news/zh/news/2026-06-27/github-desktop-3-6-worktrees-copilot/ 24 AI zh 2026-06-27T00:00:00Z GitHub：Desktop 3.6引入Git工作树和AI合并冲突解决 https://24-ai.news/zh/news/2026-06-27/google-gemini-nano-frozen-mtp-pixel/ 24 AI zh 2026-06-27T00:00:00Z Google：冻结多令牌预测使Pixel上的Gemini Nano推理速度提升50%以上 https://24-ai.news/zh/news/2026-06-27/langchain-deep-agents-prompt-caching/ 24 AI zh 2026-06-27T00:00:00Z LangChain：Deep Agents框架中的提示缓存降低长期代理延迟 https://24-ai.news/zh/news/2026-06-27/openai-gpt-5-6-sol-preview/ 24 AI zh 2026-06-27T00:00:00Z OpenAI：GPT-5.6 Sol预览发布——编码、科学与网络安全 https://24-ai.news/ja/news/2026-06-27/amd-mi355x-mxfp4-quantization/ 24 AI ja 2026-06-27T00:00:00Z AMD: MI355XでMXFP4/MXFP6混合精度量子化 — スループット最大29%向上 https://24-ai.news/ja/news/2026-06-27/anthropic-api-rate-limits-three-tiers/ 24 AI ja 2026-06-27T00:00:00Z Anthropic：APIレート制限を引き上げ — SonnetとHaikuがOpusと同水準に、3段階のティア制 https://24-ai.news/ja/news/2026-06-27/anthropic-economic-index-cadences/ 24 AI ja 2026-06-27T00:00:00Z Anthropic：Economic IndexがAI利用の日常的リズムとパターンを明らかに https://24-ai.news/ja/news/2026-06-27/arxiv-autoformalization-agent-policy-code/ 24 AI ja 2026-06-27T00:00:00Z arXiv:2606.26649: エージェントへの指示を形式検証済みのpolicy-as-codeに変換 https://24-ai.news/ja/news/2026-06-27/arxiv-capability-frontier-benchmarks/ 24 AI ja 2026-06-27T00:00:00Z arXiv:2606.26836: 標準ベンチマークはAIモデルの実力の82%を見落としている https://24-ai.news/ja/news/2026-06-27/arxiv-co-failure-ceiling-llm-ensembles/ 24 AI ja 2026-06-27T00:00:00Z arXiv:2606.27288: LLMアンサンブルが実際に効果的な場合——67のフロンティアモデルでco-failure ceiling検証 https://24-ai.news/ja/news/2026-06-27/arxiv-egg-expert-guided-kernels/ 24 AI ja 2026-06-27T00:00:00Z arXiv:2606.26758: EGG——マルチエージェントフレームワークがPyTorch比2.13倍高速なGPUカーネルを生成 https://24-ai.news/ja/news/2026-06-27/arxiv-leanguard-fast-content-moderation/ 24 AI ja 2026-06-27T00:00:00Z arXiv:2606.26686: LeanGuard——chain-of-thoughtなしの高速コンテンツモデレーションが重量級推論モデルに匹敵 https://24-ai.news/ja/news/2026-06-27/arxiv-semantic-early-stopping-agents/ 24 AI ja 2026-06-27T00:00:00Z arXiv:2606.27009: 意味的早期停止でエージェントループのトークンコストを38%削減 https://24-ai.news/ja/news/2026-06-27/aws-stripe-ai-agents-financial-compliance/ 24 AI ja 2026-06-27T00:00:00Z AWS：Stripeが金融コンプライアンス向けに100以上のAIエージェントを本番稼働——現場からの教訓 https://24-ai.news/ja/news/2026-06-27/claude-code-2-1-195/ 24 AI ja 2026-06-27T00:00:00Z Anthropic：Claude Code v2.1.195——マウス無効化、音声ディクテーション改善、フックの修正 https://24-ai.news/ja/news/2026-06-27/github-desktop-3-6-worktrees-copilot/ 24 AI ja 2026-06-27T00:00:00Z GitHub：Desktop 3.6がGit worktreesとAIによるマージコンフリクト解決を導入 https://24-ai.news/ja/news/2026-06-27/google-gemini-nano-frozen-mtp-pixel/ 24 AI ja 2026-06-27T00:00:00Z Google：凍結マルチトークン予測でPixelのGemini Nanoが50%以上高速化 https://24-ai.news/ja/news/2026-06-27/langchain-deep-agents-prompt-caching/ 24 AI ja 2026-06-27T00:00:00Z LangChain：Deep AgentsフレームワークのプロンプトキャッシングでLLMエージェントの遅延を削減 https://24-ai.news/ja/news/2026-06-27/openai-gpt-5-6-sol-preview/ 24 AI ja 2026-06-27T00:00:00Z OpenAI：GPT-5.6 Solをプレビューで発表——コーディング・科学・サイバーセキュリティ https://24-ai.news/ko/news/2026-06-27/amd-mi355x-mxfp4-quantization/ 24 AI ko 2026-06-27T00:00:00Z AMD: MI355X에서 MXFP4/MXFP6 혼합 정밀도 양자화로 처리량 29% 향상 https://24-ai.news/ko/news/2026-06-27/anthropic-api-rate-limits-three-tiers/ 24 AI ko 2026-06-27T00:00:00Z Anthropic: API 속도 제한 상향 — Sonnet·Haiku가 Opus 수준으로, 3단계 티어 도입 https://24-ai.news/ko/news/2026-06-27/anthropic-economic-index-cadences/ 24 AI ko 2026-06-27T00:00:00Z Anthropic: Economic Index, 하루 중 AI 활용 리듬과 패턴 분석 https://24-ai.news/ko/news/2026-06-27/arxiv-autoformalization-agent-policy-code/ 24 AI ko 2026-06-27T00:00:00Z arXiv:2606.26649: AI 에이전트 지침을 형식적으로 검증된 정책 코드로 자동 변환 https://24-ai.news/ko/news/2026-06-27/arxiv-capability-frontier-benchmarks/ 24 AI ko 2026-06-27T00:00:00Z arXiv:2606.26836: 표준 벤치마크, AI 모델 실제 능력의 82%를 놓치고 있다 https://24-ai.news/ko/news/2026-06-27/arxiv-co-failure-ceiling-llm-ensembles/ 24 AI ko 2026-06-27T00:00:00Z arXiv:2606.27288: LLM 앙상블이 실제로 도움이 될 때 — 67개 프론티어 모델의 공동 실패 천장 https://24-ai.news/ko/news/2026-06-27/arxiv-egg-expert-guided-kernels/ 24 AI ko 2026-06-27T00:00:00Z arXiv:2606.26758: EGG — 다중 에이전트 프레임워크, PyTorch보다 2.13배 빠른 GPU 커널 생성 https://24-ai.news/ko/news/2026-06-27/arxiv-leanguard-fast-content-moderation/ 24 AI ko 2026-06-27T00:00:00Z arXiv:2606.26686: LeanGuard — CoT 없이 빠른 콘텐츠 모더레이션으로 무거운 추론 모델과 대등 https://24-ai.news/ko/news/2026-06-27/arxiv-semantic-early-stopping-agents/ 24 AI ko 2026-06-27T00:00:00Z arXiv:2606.27009: 의미론적 조기 중단으로 에이전트 루프 비용 38% 절감 https://24-ai.news/ko/news/2026-06-27/aws-stripe-ai-agents-financial-compliance/ 24 AI ko 2026-06-27T00:00:00Z AWS: Stripe, 금융 규정 준수를 위해 100개 이상의 AI 에이전트 출시 — 프로덕션 교훈 https://24-ai.news/ko/news/2026-06-27/claude-code-2-1-195/ 24 AI ko 2026-06-27T00:00:00Z Anthropic: Claude Code v2.1.195 — 마우스 비활성화, 개선된 음성 받아쓰기, 훅 수정 https://24-ai.news/ko/news/2026-06-27/github-desktop-3-6-worktrees-copilot/ 24 AI ko 2026-06-27T00:00:00Z GitHub: Desktop 3.6, Git 워크트리와 AI 머지 충돌 해결 기능 출시 https://24-ai.news/ko/news/2026-06-27/google-gemini-nano-frozen-mtp-pixel/ 24 AI ko 2026-06-27T00:00:00Z Google: 고정된 멀티 토큰 예측으로 Pixel에서 Gemini Nano 50%+ 속도 향상 https://24-ai.news/ko/news/2026-06-27/langchain-deep-agents-prompt-caching/ 24 AI ko 2026-06-27T00:00:00Z LangChain: Deep Agents 프레임워크의 프롬프트 캐싱으로 장기 실행 에이전트 지연 시간 감소 https://24-ai.news/ko/news/2026-06-27/openai-gpt-5-6-sol-preview/ 24 AI ko 2026-06-27T00:00:00Z OpenAI: GPT-5.6 Sol 프리뷰 발표 — 코딩, 과학, 사이버보안