https://24-ai.news/hr/news/2026-07-21/amd-minimax-m3-atomesh/ 24 AI hr 2026-07-21T00:00:00Z AMD: MiniMax-M3 model od 428 milijardi parametara servira brže od NVIDIA B200 i B300 na Instinct MI355X GPU-ovima https://24-ai.news/hr/news/2026-07-21/arxiv-clinical-llm-evidence-sufficiency/ 24 AI hr 2026-07-21T00:00:00Z arXiv:2607.18086: Evidence-sufficiency prompt smanjuje pretjeranu sigurnost kliničkih LLM-ova, ali i točnost https://24-ai.news/hr/news/2026-07-21/arxiv-mind-jailbreak-text-to-image/ 24 AI hr 2026-07-21T00:00:00Z arXiv:2607.17779: MIND okvir zaobilazi obranu text-to-image modela stopom uspješnosti napada od 95,62% https://24-ai.news/hr/news/2026-07-21/arxiv-sophia-reasoning-self-loops/ 24 AI hr 2026-07-21T00:00:00Z arXiv:2607.18100: SOPHIA activation steering izvlači jezične modele iz reasoning samoponavljajućih petlji https://24-ai.news/hr/news/2026-07-21/arxiv-vrr-stop-verify-repair-agents/ 24 AI hr 2026-07-21T00:00:00Z arXiv:2607.17641: VRR-Stop rješava kada LLM agent treba stati s ponavljanim popravljanjem svog odgovora https://24-ai.news/hr/news/2026-07-21/arxiv-worldcuparena-football-forecast/ 24 AI hr 2026-07-21T00:00:00Z arXiv:2607.18084: WorldCupArena testira jezične modele na predviđanju rezultata nogometnog prvenstva https://24-ai.news/hr/news/2026-07-21/aws-self-distilled-reasoning-nova/ 24 AI hr 2026-07-21T00:00:00Z AWS: Self-Distilled Reasoning vraća matematičku točnost Amazon Nova 2 sa 6% na oko 68% nakon fine-tuninga https://24-ai.news/hr/news/2026-07-21/claude-code-v2-1-217/ 24 AI hr 2026-07-21T00:00:00Z Anthropic: Claude Code v2.1.217 popravlja memory leak kod MCP-a i uvodi limit od 20 konkurentnih subagenata https://24-ai.news/hr/news/2026-07-21/cncf-platform-engineering-agentic/ 24 AI hr 2026-07-21T00:00:00Z CNCF: platform engineering mora tretirati AI agente kao ravnopravne potrošače platforme uz aplikacije https://24-ai.news/hr/news/2026-07-21/gemini-3-6-flash-ga/ 24 AI hr 2026-07-21T00:00:00Z Google: Gemini 3.6 Flash i Gemini 3.5 Flash-Lite dobili opću dostupnost uz 17% manje tokena i novi Cyber model https://24-ai.news/hr/news/2026-07-21/github-copilot-canvases/ 24 AI hr 2026-07-21T00:00:00Z GitHub Copilot: canvasi donose dijeljene interaktivne radne površine za suradnju s AI agentima uživo https://24-ai.news/hr/news/2026-07-21/meta-genesis-mission-sam3-dinov3/ 24 AI hr 2026-07-21T00:00:00Z Meta: SAM 3 i DINOv3 otvoreni za Genesis Mission, segmentacijska analiza skraćena s mjesec dana na 15 minuta https://24-ai.news/hr/news/2026-07-21/nvidia-spectrum-6-gigascale-ai/ 24 AI hr 2026-07-21T00:00:00Z NVIDIA: Spectrum-6 Ethernet switch donosi 102,4 Tb/s i dvostruko veću propusnost za gigascale AI tvornice https://24-ai.news/hr/news/2026-07-21/openai-huggingface-security-incident/ 24 AI hr 2026-07-21T00:00:00Z OpenAI: zajednički rani nalazi s Hugging Faceom o sigurnosnom incidentu tijekom evaluacije AI modela https://24-ai.news/hr/news/2026-07-21/sakana-fugu-cyber-model/ 24 AI hr 2026-07-21T00:00:00Z Sakana AI: Fugu-Cyber postiže 86,9% na CyberGym i 72,1% na CTI-REALM testu, usporedivo s GPT-5.5-Cyber https://24-ai.news/en/news/2026-07-21/amd-minimax-m3-atomesh/ 24 AI en 2026-07-21T00:00:00Z AMD: 428-billion-parameter MiniMax-M3 model serves faster than NVIDIA B200 and B300 on Instinct MI355X GPUs https://24-ai.news/en/news/2026-07-21/arxiv-clinical-llm-evidence-sufficiency/ 24 AI en 2026-07-21T00:00:00Z arXiv:2607.18086: Evidence-sufficiency prompting reduces clinical LLM overconfidence, but also accuracy https://24-ai.news/en/news/2026-07-21/arxiv-mind-jailbreak-text-to-image/ 24 AI en 2026-07-21T00:00:00Z arXiv:2607.17779: MIND framework bypasses text-to-image model defenses with a 95.62% attack success rate https://24-ai.news/en/news/2026-07-21/arxiv-sophia-reasoning-self-loops/ 24 AI en 2026-07-21T00:00:00Z arXiv:2607.18100: SOPHIA activation steering pulls language models out of reasoning self-loops https://24-ai.news/en/news/2026-07-21/arxiv-vrr-stop-verify-repair-agents/ 24 AI en 2026-07-21T00:00:00Z arXiv:2607.17641: VRR-Stop solves when an LLM agent should stop repeatedly fixing its answer https://24-ai.news/en/news/2026-07-21/arxiv-worldcuparena-football-forecast/ 24 AI en 2026-07-21T00:00:00Z arXiv:2607.18084: WorldCupArena tests language models on predicting football World Cup results https://24-ai.news/en/news/2026-07-21/aws-self-distilled-reasoning-nova/ 24 AI en 2026-07-21T00:00:00Z AWS: Self-Distilled Reasoning restores Amazon Nova 2 math accuracy from 6% to about 68% after fine-tuning https://24-ai.news/en/news/2026-07-21/claude-code-v2-1-217/ 24 AI en 2026-07-21T00:00:00Z Anthropic: Claude Code v2.1.217 fixes an MCP memory leak and introduces a 20-concurrent-subagent limit https://24-ai.news/en/news/2026-07-21/cncf-platform-engineering-agentic/ 24 AI en 2026-07-21T00:00:00Z CNCF: platform engineering must treat AI agents as equal platform consumers alongside applications https://24-ai.news/en/news/2026-07-21/gemini-3-6-flash-ga/ 24 AI en 2026-07-21T00:00:00Z Google: Gemini 3.6 Flash and Gemini 3.5 Flash-Lite reach general availability with 17% fewer tokens and a new Cyber model https://24-ai.news/en/news/2026-07-21/github-copilot-canvases/ 24 AI en 2026-07-21T00:00:00Z GitHub Copilot: canvases bring shared interactive workspaces for live collaboration with AI agents https://24-ai.news/en/news/2026-07-21/meta-genesis-mission-sam3-dinov3/ 24 AI en 2026-07-21T00:00:00Z Meta: SAM 3 and DINOv3 open-sourced for the Genesis Mission, segmentation analysis cut from a month to 15 minutes https://24-ai.news/en/news/2026-07-21/nvidia-spectrum-6-gigascale-ai/ 24 AI en 2026-07-21T00:00:00Z NVIDIA: Spectrum-6 Ethernet switch delivers 102.4 Tb/s and double the throughput for gigascale AI factories https://24-ai.news/en/news/2026-07-21/openai-huggingface-security-incident/ 24 AI en 2026-07-21T00:00:00Z OpenAI: joint early findings with Hugging Face on a security incident during AI model evaluation https://24-ai.news/en/news/2026-07-21/sakana-fugu-cyber-model/ 24 AI en 2026-07-21T00:00:00Z Sakana AI: Fugu-Cyber scores 86.9% on CyberGym and 72.1% on CTI-REALM, comparable to GPT-5.5-Cyber https://24-ai.news/de/news/2026-07-21/amd-minimax-m3-atomesh/ 24 AI de 2026-07-21T00:00:00Z AMD: MiniMax-M3-Modell mit 428 Milliarden Parametern bedient schneller als NVIDIA B200 und B300 auf Instinct-MI355X-GPUs https://24-ai.news/de/news/2026-07-21/arxiv-clinical-llm-evidence-sufficiency/ 24 AI de 2026-07-21T00:00:00Z arXiv:2607.18086: Evidence-Sufficiency-Prompting senkt die Selbstüberschätzung klinischer LLMs, aber auch die Genauigkeit https://24-ai.news/de/news/2026-07-21/arxiv-mind-jailbreak-text-to-image/ 24 AI de 2026-07-21T00:00:00Z arXiv:2607.17779: MIND-Framework umgeht die Abwehr von Text-to-Image-Modellen mit einer Angriffserfolgsquote von 95,62 % https://24-ai.news/de/news/2026-07-21/arxiv-sophia-reasoning-self-loops/ 24 AI de 2026-07-21T00:00:00Z arXiv:2607.18100: SOPHIA-Activation-Steering befreit Sprachmodelle aus sich selbst wiederholenden Reasoning-Schleifen https://24-ai.news/de/news/2026-07-21/arxiv-vrr-stop-verify-repair-agents/ 24 AI de 2026-07-21T00:00:00Z arXiv:2607.17641: VRR-Stop löst, wann ein LLM-Agent das wiederholte Reparieren seiner Antwort stoppen sollte https://24-ai.news/de/news/2026-07-21/arxiv-worldcuparena-football-forecast/ 24 AI de 2026-07-21T00:00:00Z arXiv:2607.18084: WorldCupArena testet Sprachmodelle bei der Vorhersage von Fußball-WM-Ergebnissen https://24-ai.news/de/news/2026-07-21/aws-self-distilled-reasoning-nova/ 24 AI de 2026-07-21T00:00:00Z AWS: Self-Distilled Reasoning stellt mathematische Genauigkeit von Amazon Nova 2 nach Fine-Tuning von 6 % auf rund 68 % wieder her https://24-ai.news/de/news/2026-07-21/claude-code-v2-1-217/ 24 AI de 2026-07-21T00:00:00Z Anthropic: Claude Code v2.1.217 behebt Memory-Leak bei MCP und führt Limit von 20 gleichzeitigen Subagenten ein https://24-ai.news/de/news/2026-07-21/cncf-platform-engineering-agentic/ 24 AI de 2026-07-21T00:00:00Z CNCF: Platform Engineering muss KI-Agenten als gleichberechtigte Plattform-Konsumenten neben Anwendungen behandeln https://24-ai.news/de/news/2026-07-21/gemini-3-6-flash-ga/ 24 AI de 2026-07-21T00:00:00Z Google: Gemini 3.6 Flash und Gemini 3.5 Flash-Lite erreichen allgemeine Verfügbarkeit mit 17 % weniger Token und neuem Cyber-Modell https://24-ai.news/de/news/2026-07-21/github-copilot-canvases/ 24 AI de 2026-07-21T00:00:00Z GitHub Copilot: Canvases bringen gemeinsame interaktive Arbeitsflächen für die Live-Zusammenarbeit mit KI-Agenten https://24-ai.news/de/news/2026-07-21/meta-genesis-mission-sam3-dinov3/ 24 AI de 2026-07-21T00:00:00Z Meta: SAM 3 und DINOv3 für Genesis Mission geöffnet, Segmentierungsanalyse von einem Monat auf 15 Minuten verkürzt https://24-ai.news/de/news/2026-07-21/nvidia-spectrum-6-gigascale-ai/ 24 AI de 2026-07-21T00:00:00Z NVIDIA: Spectrum-6-Ethernet-Switch liefert 102,4 Tb/s und doppelten Durchsatz für Gigascale-KI-Fabriken https://24-ai.news/de/news/2026-07-21/openai-huggingface-security-incident/ 24 AI de 2026-07-21T00:00:00Z OpenAI: gemeinsame frühe Erkenntnisse mit Hugging Face zu einem Sicherheitsvorfall bei der KI-Modellevaluierung https://24-ai.news/de/news/2026-07-21/sakana-fugu-cyber-model/ 24 AI de 2026-07-21T00:00:00Z Sakana AI: Fugu-Cyber erzielt 86,9 % bei CyberGym und 72,1 % bei CTI-REALM, vergleichbar mit GPT-5.5-Cyber https://24-ai.news/zh/news/2026-07-21/amd-minimax-m3-atomesh/ 24 AI zh-cn 2026-07-21T00:00:00Z AMD:4280亿参数的MiniMax-M3模型在Instinct MI355X GPU上的服务速度超过NVIDIA B200和B300 https://24-ai.news/zh/news/2026-07-21/arxiv-clinical-llm-evidence-sufficiency/ 24 AI zh-cn 2026-07-21T00:00:00Z arXiv:2607.18086: 证据充分性提示降低了临床LLM的过度自信,但也降低了准确率 https://24-ai.news/zh/news/2026-07-21/arxiv-mind-jailbreak-text-to-image/ 24 AI zh-cn 2026-07-21T00:00:00Z arXiv:2607.17779: MIND框架以95.62%的攻击成功率绕过text-to-image模型的防御 https://24-ai.news/zh/news/2026-07-21/arxiv-sophia-reasoning-self-loops/ 24 AI zh-cn 2026-07-21T00:00:00Z arXiv:2607.18100: SOPHIA激活转向技术让语言模型摆脱推理自我循环 https://24-ai.news/zh/news/2026-07-21/arxiv-vrr-stop-verify-repair-agents/ 24 AI zh-cn 2026-07-21T00:00:00Z arXiv:2607.17641: VRR-Stop解决LLM智能体何时应停止反复修复自身回答的问题 https://24-ai.news/zh/news/2026-07-21/arxiv-worldcuparena-football-forecast/ 24 AI zh-cn 2026-07-21T00:00:00Z arXiv:2607.18084: WorldCupArena测试语言模型预测足球世界杯赛果的能力 https://24-ai.news/zh/news/2026-07-21/aws-self-distilled-reasoning-nova/ 24 AI zh-cn 2026-07-21T00:00:00Z AWS:Self-Distilled Reasoning让Amazon Nova 2数学准确率在微调后从6%恢复至约68% https://24-ai.news/zh/news/2026-07-21/claude-code-v2-1-217/ 24 AI zh-cn 2026-07-21T00:00:00Z Anthropic:Claude Code v2.1.217修复MCP内存泄漏问题,并引入20个并发子代理的限制 https://24-ai.news/zh/news/2026-07-21/cncf-platform-engineering-agentic/ 24 AI zh-cn 2026-07-21T00:00:00Z CNCF:平台工程必须将AI智能体视为与应用程序平等的平台消费者 https://24-ai.news/zh/news/2026-07-21/gemini-3-6-flash-ga/ 24 AI zh-cn 2026-07-21T00:00:00Z 谷歌:Gemini 3.6 Flash与Gemini 3.5 Flash-Lite正式全面上市,输出token减少17%,并推出全新Cyber模型 https://24-ai.news/zh/news/2026-07-21/github-copilot-canvases/ 24 AI zh-cn 2026-07-21T00:00:00Z GitHub Copilot:canvas为与AI智能体实时协作带来共享交互式工作空间 https://24-ai.news/zh/news/2026-07-21/meta-genesis-mission-sam3-dinov3/ 24 AI zh-cn 2026-07-21T00:00:00Z Meta:SAM 3与DINOv3向Genesis Mission开放,分割分析时间从一个月缩短至15分钟 https://24-ai.news/zh/news/2026-07-21/nvidia-spectrum-6-gigascale-ai/ 24 AI zh-cn 2026-07-21T00:00:00Z NVIDIA:Spectrum-6以太网交换机为超大规模AI工厂带来102.4 Tb/s容量与两倍带宽 https://24-ai.news/zh/news/2026-07-21/openai-huggingface-security-incident/ 24 AI zh-cn 2026-07-21T00:00:00Z OpenAI:与Hugging Face联合发布AI模型评估过程中安全事件的早期调查结果 https://24-ai.news/zh/news/2026-07-21/sakana-fugu-cyber-model/ 24 AI zh-cn 2026-07-21T00:00:00Z Sakana AI:Fugu-Cyber在CyberGym测试中达86.9%,在CTI-REALM测试中达72.1%,与GPT-5.5-Cyber相当 https://24-ai.news/ja/news/2026-07-21/amd-minimax-m3-atomesh/ 24 AI ja 2026-07-21T00:00:00Z AMD、4280億パラメータのMiniMax-M3モデルがInstinct MI355X GPUでNVIDIA B200・B300を上回る速度でサービング https://24-ai.news/ja/news/2026-07-21/arxiv-clinical-llm-evidence-sufficiency/ 24 AI ja 2026-07-21T00:00:00Z arXiv:2607.18086: エビデンス充足性プロンプトが臨床LLMの過信を減らすが、精度も低下させる https://24-ai.news/ja/news/2026-07-21/arxiv-mind-jailbreak-text-to-image/ 24 AI ja 2026-07-21T00:00:00Z arXiv:2607.17779: MINDフレームワーク、95.62%の攻撃成功率でtext-to-imageモデルの防御を突破 https://24-ai.news/ja/news/2026-07-21/arxiv-sophia-reasoning-self-loops/ 24 AI ja 2026-07-21T00:00:00Z arXiv:2607.18100: SOPHIAのアクティベーションステアリングが言語モデルを推論の自己ループから救い出す https://24-ai.news/ja/news/2026-07-21/arxiv-vrr-stop-verify-repair-agents/ 24 AI ja 2026-07-21T00:00:00Z arXiv:2607.17641: VRR-Stop、LLMエージェントが自己回答の反復修正をいつ止めるべきかを解決 https://24-ai.news/ja/news/2026-07-21/arxiv-worldcuparena-football-forecast/ 24 AI ja 2026-07-21T00:00:00Z arXiv:2607.18084: WorldCupArenaがサッカーワールドカップの結果予測で言語モデルをテスト https://24-ai.news/ja/news/2026-07-21/aws-self-distilled-reasoning-nova/ 24 AI ja 2026-07-21T00:00:00Z AWS、Self-Distilled ReasoningでAmazon Nova 2の数学的精度をファインチューニング後6%から約68%へ回復 https://24-ai.news/ja/news/2026-07-21/claude-code-v2-1-217/ 24 AI ja 2026-07-21T00:00:00Z Anthropic、Claude Code v2.1.217でMCPのメモリリークを修正、同時実行サブエージェント数を20に制限 https://24-ai.news/ja/news/2026-07-21/cncf-platform-engineering-agentic/ 24 AI ja 2026-07-21T00:00:00Z CNCF、プラットフォームエンジニアリングはAIエージェントをアプリケーションと対等なプラットフォーム消費者として扱うべき https://24-ai.news/ja/news/2026-07-21/gemini-3-6-flash-ga/ 24 AI ja 2026-07-21T00:00:00Z Google、Gemini 3.6 FlashとGemini 3.5 Flash-Liteが一般提供開始、出力トークン17%削減と新Cyberモデルを発表 https://24-ai.news/ja/news/2026-07-21/github-copilot-canvases/ 24 AI ja 2026-07-21T00:00:00Z GitHub Copilot、canvasがAIエージェントとのリアルタイム協働を実現する共有インタラクティブワークスペースを提供 https://24-ai.news/ja/news/2026-07-21/meta-genesis-mission-sam3-dinov3/ 24 AI ja 2026-07-21T00:00:00Z Meta、SAM 3とDINOv3をGenesis Missionに開放、セグメンテーション分析を1ヶ月から15分に短縮 https://24-ai.news/ja/news/2026-07-21/nvidia-spectrum-6-gigascale-ai/ 24 AI ja 2026-07-21T00:00:00Z NVIDIA、Spectrum-6イーサネットスイッチがギガスケールAI向けに102.4 Tb/sと2倍の帯域幅を実現 https://24-ai.news/ja/news/2026-07-21/openai-huggingface-security-incident/ 24 AI ja 2026-07-21T00:00:00Z OpenAI、AIモデル評価中のセキュリティインシデントについてHugging Faceと共同で初期調査結果を発表 https://24-ai.news/ja/news/2026-07-21/sakana-fugu-cyber-model/ 24 AI ja 2026-07-21T00:00:00Z Sakana AI、Fugu-CyberがCyberGymで86.9%、CTI-REALMで72.1%を達成、GPT-5.5-Cyberに匹敵 https://24-ai.news/ko/news/2026-07-21/amd-minimax-m3-atomesh/ 24 AI ko 2026-07-21T00:00:00Z AMD, 4,280억 파라미터 MiniMax-M3 모델이 Instinct MI355X GPU에서 NVIDIA B200·B300보다 빠르게 서빙 https://24-ai.news/ko/news/2026-07-21/arxiv-clinical-llm-evidence-sufficiency/ 24 AI ko 2026-07-21T00:00:00Z arXiv:2607.18086: 증거 충분성 프롬프트, 임상 LLM의 과신은 줄이지만 정확도도 낮춘다 https://24-ai.news/ko/news/2026-07-21/arxiv-mind-jailbreak-text-to-image/ 24 AI ko 2026-07-21T00:00:00Z arXiv:2607.17779: MIND 프레임워크, 공격 성공률 95.62%로 text-to-image 모델 방어 우회 https://24-ai.news/ko/news/2026-07-21/arxiv-sophia-reasoning-self-loops/ 24 AI ko 2026-07-21T00:00:00Z arXiv:2607.18100: SOPHIA 활성화 스티어링, 언어 모델을 추론 자기 반복 루프에서 구출 https://24-ai.news/ko/news/2026-07-21/arxiv-vrr-stop-verify-repair-agents/ 24 AI ko 2026-07-21T00:00:00Z arXiv:2607.17641: VRR-Stop, LLM 에이전트가 자기 답변 반복 수정을 언제 멈춰야 하는지 해결 https://24-ai.news/ko/news/2026-07-21/arxiv-worldcuparena-football-forecast/ 24 AI ko 2026-07-21T00:00:00Z arXiv:2607.18084: WorldCupArena, 축구 월드컵 결과 예측으로 언어 모델 테스트 https://24-ai.news/ko/news/2026-07-21/aws-self-distilled-reasoning-nova/ 24 AI ko 2026-07-21T00:00:00Z AWS, Self-Distilled Reasoning로 Amazon Nova 2 수학 정확도를 파인튜닝 후 6%에서 약 68%로 회복 https://24-ai.news/ko/news/2026-07-21/claude-code-v2-1-217/ 24 AI ko 2026-07-21T00:00:00Z Anthropic, Claude Code v2.1.217에서 MCP 메모리 누수 수정 및 동시 서브에이전트 20개 제한 도입 https://24-ai.news/ko/news/2026-07-21/cncf-platform-engineering-agentic/ 24 AI ko 2026-07-21T00:00:00Z CNCF, 플랫폼 엔지니어링은 AI 에이전트를 애플리케이션과 동등한 플랫폼 소비자로 다뤄야 한다 https://24-ai.news/ko/news/2026-07-21/gemini-3-6-flash-ga/ 24 AI ko 2026-07-21T00:00:00Z 구글, Gemini 3.6 Flash와 Gemini 3.5 Flash-Lite 정식 출시—출력 토큰 17% 감소 및 신규 Cyber 모델 공개 https://24-ai.news/ko/news/2026-07-21/github-copilot-canvases/ 24 AI ko 2026-07-21T00:00:00Z GitHub Copilot, 캔버스로 AI 에이전트와의 실시간 협업을 위한 공유 인터랙티브 작업 공간 제공 https://24-ai.news/ko/news/2026-07-21/meta-genesis-mission-sam3-dinov3/ 24 AI ko 2026-07-21T00:00:00Z 메타, SAM 3와 DINOv3를 Genesis Mission에 공개—분할 분석 시간 한 달에서 15분으로 단축 https://24-ai.news/ko/news/2026-07-21/nvidia-spectrum-6-gigascale-ai/ 24 AI ko 2026-07-21T00:00:00Z NVIDIA, Spectrum-6 이더넷 스위치로 기가스케일 AI에 102.4 Tb/s와 두 배의 대역폭 제공 https://24-ai.news/ko/news/2026-07-21/openai-huggingface-security-incident/ 24 AI ko 2026-07-21T00:00:00Z OpenAI, AI 모델 평가 중 발생한 보안 사고에 대해 Hugging Face와 공동 초기 조사 결과 발표 https://24-ai.news/ko/news/2026-07-21/sakana-fugu-cyber-model/ 24 AI ko 2026-07-21T00:00:00Z Sakana AI, Fugu-Cyber가 CyberGym에서 86.9%, CTI-REALM에서 72.1% 달성—GPT-5.5-Cyber와 견줄 만한 수준 https://24-ai.news/hr/news/2026-07-20/amd-geak-v3-kernel-optimization/ 24 AI hr 2026-07-20T00:00:00Z AMD: GEAK v3 open-source agent ubrzava GPU kernele 3,02× na Instinct MI300X, MI355X i RDNA4 arhitekturi https://24-ai.news/hr/news/2026-07-20/anthropic-rare-disease-grants/ 24 AI hr 2026-07-20T00:00:00Z Anthropic: do 50.000 dolara u Claude kreditima za istraživanje rijetkih genetskih bolesti kroz AI for Science https://24-ai.news/hr/news/2026-07-20/arxiv-coding-agents-arc-agi-3/ 24 AI hr 2026-07-20T00:00:00Z arXiv:2607.15439: Verifikacijski coding agenti potpuno rješavaju ARC-AGI-3 uz oko 99% RHAE rezultata https://24-ai.news/hr/news/2026-07-20/arxiv-craft-rubric-capability-diagnosis/ 24 AI hr 2026-07-20T00:00:00Z arXiv:2607.16122: CRAFT metoda klasterira kriterije ocjenjivanja i otkriva slabosti kod jezičnih modela https://24-ai.news/hr/news/2026-07-20/arxiv-dsworld-data-science-agents/ 24 AI hr 2026-07-20T00:00:00Z arXiv:2607.15901: DSWorld world model ubrzava agente za data science čak 14 puta u treningu i inferenciji https://24-ai.news/hr/news/2026-07-20/arxiv-harmonizing-ai-safety-thresholds/ 24 AI hr 2026-07-20T00:00:00Z arXiv:2607.16112: Znanstvenici predlažu usklađivanje sigurnosnih pragova kapabilnosti frontier AI kompanija https://24-ai.news/hr/news/2026-07-20/arxiv-s1-omni-scientific-model/ 24 AI hr 2026-07-20T00:00:00Z arXiv:2607.15686: S1-Omni objedinjuje znanost u jedan multimodalni model i nadmašuje GPT-5.5 i Gemini https://24-ai.news/hr/news/2026-07-20/arxiv-seerguard-gui-agent-safety/ 24 AI hr 2026-07-20T00:00:00Z arXiv:2607.15550: SeerGuard sigurnosnim svjetskim modelom predviđa posljedice akcija mobilnih GUI agenata https://24-ai.news/hr/news/2026-07-20/claude-code-v2-1-215-216/ 24 AI hr 2026-07-20T00:00:00Z Anthropic: Claude Code v2.1.215 i v2.1.216 popravljaju performanse, OAuth grešku i worktree izolaciju agenta https://24-ai.news/hr/news/2026-07-20/cncf-goodput-vs-throughput-llm/ 24 AI hr 2026-07-20T00:00:00Z CNCF: Goodput metrika otkriva gotovo 10× lošiju p95 latenciju kod throughput-optimiziranog LLM servinga https://24-ai.news/hr/news/2026-07-20/eu-ai-act-article-50-transparency/ 24 AI hr 2026-07-20T00:00:00Z Europska komisija: smjernice za transparentnost AI sustava prema članku 50 stupaju na snagu 2. kolovoza https://24-ai.news/hr/news/2026-07-20/github-code-quality-ga/ 24 AI hr 2026-07-20T00:00:00Z GitHub: Code Quality dostiže Generally Available status uz CodeQL analizu i Copilot Autofix popravke https://24-ai.news/hr/news/2026-07-20/github-sponsors-100-million/ 24 AI hr 2026-07-20T00:00:00Z GitHub Sponsors: globalna zajednica open-source developera prikupila je 100 milijuna dolara donacija https://24-ai.news/hr/news/2026-07-20/langchain-issuebench-agent-eval/ 24 AI hr 2026-07-20T00:00:00Z LangChain: IssueBench interno mjeri prepoznavanje 15 kategorija grešaka unutar agent trace-ovima sustava https://24-ai.news/hr/news/2026-07-20/openai-long-horizon-model-safety/ 24 AI hr 2026-07-20T00:00:00Z OpenAI: naučene lekcije i nove zaštitne mjere za sigurnost dugotrajno-djelujućih (long-horizon) AI modela https://24-ai.news/en/news/2026-07-20/amd-geak-v3-kernel-optimization/ 24 AI en 2026-07-20T00:00:00Z AMD: GEAK v3 Open-Source Agent Speeds Up GPU Kernels 3.02× on Instinct MI300X, MI355X, and RDNA4 https://24-ai.news/en/news/2026-07-20/anthropic-rare-disease-grants/ 24 AI en 2026-07-20T00:00:00Z Anthropic: Up to $50,000 in Claude Credits for Rare Genetic Disease Research Through AI for Science https://24-ai.news/en/news/2026-07-20/arxiv-coding-agents-arc-agi-3/ 24 AI en 2026-07-20T00:00:00Z arXiv:2607.15439: Verification Coding Agents Fully Solve ARC-AGI-3 with Around 99% RHAE https://24-ai.news/en/news/2026-07-20/arxiv-craft-rubric-capability-diagnosis/ 24 AI en 2026-07-20T00:00:00Z arXiv:2607.16122: CRAFT Method Clusters Grading Rubrics to Reveal Weaknesses in Language Models https://24-ai.news/en/news/2026-07-20/arxiv-dsworld-data-science-agents/ 24 AI en 2026-07-20T00:00:00Z arXiv:2607.15901: DSWorld World Model Speeds Up Data Science Agents Up to 14× in Training and Inference https://24-ai.news/en/news/2026-07-20/arxiv-harmonizing-ai-safety-thresholds/ 24 AI en 2026-07-20T00:00:00Z arXiv:2607.16112: Researchers Propose Harmonizing Safety Capability Thresholds Across Frontier AI Companies https://24-ai.news/en/news/2026-07-20/arxiv-s1-omni-scientific-model/ 24 AI en 2026-07-20T00:00:00Z arXiv:2607.15686: S1-Omni Unifies Science in a Single Multimodal Model, Outperforming GPT-5.5 and Gemini https://24-ai.news/en/news/2026-07-20/arxiv-seerguard-gui-agent-safety/ 24 AI en 2026-07-20T00:00:00Z arXiv:2607.15550: SeerGuard Uses a Safety World Model to Predict Consequences of Mobile GUI Agent Actions https://24-ai.news/en/news/2026-07-20/claude-code-v2-1-215-216/ 24 AI en 2026-07-20T00:00:00Z Anthropic: Claude Code v2.1.215 and v2.1.216 Fix Performance, an OAuth Error, and Agent Worktree Isolation https://24-ai.news/en/news/2026-07-20/cncf-goodput-vs-throughput-llm/ 24 AI en 2026-07-20T00:00:00Z CNCF: Goodput Metric Reveals Nearly 10× Worse p95 Latency in Throughput-Optimized LLM Serving https://24-ai.news/en/news/2026-07-20/eu-ai-act-article-50-transparency/ 24 AI en 2026-07-20T00:00:00Z European Commission: AI Transparency Guidelines Under Article 50 Take Effect August 2 https://24-ai.news/en/news/2026-07-20/github-code-quality-ga/ 24 AI en 2026-07-20T00:00:00Z GitHub: Code Quality Reaches General Availability with CodeQL Analysis and Copilot Autofix Fixes https://24-ai.news/en/news/2026-07-20/github-sponsors-100-million/ 24 AI en 2026-07-20T00:00:00Z GitHub Sponsors: Global Open-Source Developer Community Raises $100 Million in Donations https://24-ai.news/en/news/2026-07-20/langchain-issuebench-agent-eval/ 24 AI en 2026-07-20T00:00:00Z LangChain: IssueBench Internally Measures Recognition of 15 Error Categories in Agent Traces https://24-ai.news/en/news/2026-07-20/openai-long-horizon-model-safety/ 24 AI en 2026-07-20T00:00:00Z OpenAI: Lessons Learned and New Safeguards for Long-Horizon AI Model Safety https://24-ai.news/de/news/2026-07-20/amd-geak-v3-kernel-optimization/ 24 AI de 2026-07-20T00:00:00Z AMD: Open-Source-Agent GEAK v3 beschleunigt GPU-Kernel um 3,02× auf Instinct MI300X, MI355X und RDNA4 https://24-ai.news/de/news/2026-07-20/anthropic-rare-disease-grants/ 24 AI de 2026-07-20T00:00:00Z Anthropic: Bis zu 50.000 US-Dollar in Claude-Guthaben für die Erforschung seltener genetischer Erkrankungen über AI for Science https://24-ai.news/de/news/2026-07-20/arxiv-coding-agents-arc-agi-3/ 24 AI de 2026-07-20T00:00:00Z arXiv:2607.15439: Verifikations-Coding-Agenten lösen ARC-AGI-3 vollständig mit rund 99 % RHAE https://24-ai.news/de/news/2026-07-20/arxiv-craft-rubric-capability-diagnosis/ 24 AI de 2026-07-20T00:00:00Z arXiv:2607.16122: CRAFT-Methode clustert Bewertungskriterien und deckt Schwächen bei Sprachmodellen auf https://24-ai.news/de/news/2026-07-20/arxiv-dsworld-data-science-agents/ 24 AI de 2026-07-20T00:00:00Z arXiv:2607.15901: DSWorld-Weltmodell beschleunigt Data-Science-Agenten um bis zu 14× beim Training und bei der Inferenz https://24-ai.news/de/news/2026-07-20/arxiv-harmonizing-ai-safety-thresholds/ 24 AI de 2026-07-20T00:00:00Z arXiv:2607.16112: Wissenschaftler schlagen Angleichung der Sicherheits-Kapazitätsschwellen von Frontier-KI-Unternehmen vor https://24-ai.news/de/news/2026-07-20/arxiv-s1-omni-scientific-model/ 24 AI de 2026-07-20T00:00:00Z arXiv:2607.15686: S1-Omni vereint Wissenschaft in einem multimodalen Modell und übertrifft GPT-5.5 und Gemini https://24-ai.news/de/news/2026-07-20/arxiv-seerguard-gui-agent-safety/ 24 AI de 2026-07-20T00:00:00Z arXiv:2607.15550: SeerGuard sagt mit einem Sicherheits-Weltmodell die Folgen von Aktionen mobiler GUI-Agenten voraus https://24-ai.news/de/news/2026-07-20/claude-code-v2-1-215-216/ 24 AI de 2026-07-20T00:00:00Z Anthropic: Claude Code v2.1.215 und v2.1.216 beheben Performance-Probleme, einen OAuth-Fehler und die Worktree-Isolation von Agenten https://24-ai.news/de/news/2026-07-20/cncf-goodput-vs-throughput-llm/ 24 AI de 2026-07-20T00:00:00Z CNCF: Goodput-Metrik zeigt fast 10× schlechtere p95-Latenz bei durchsatzoptimiertem LLM-Serving https://24-ai.news/de/news/2026-07-20/eu-ai-act-article-50-transparency/ 24 AI de 2026-07-20T00:00:00Z Europäische Kommission: Leitlinien zur KI-Transparenz nach Artikel 50 gelten ab 2. August https://24-ai.news/de/news/2026-07-20/github-code-quality-ga/ 24 AI de 2026-07-20T00:00:00Z GitHub: Code Quality erreicht General Availability mit CodeQL-Analyse und Copilot-Autofix-Korrekturen https://24-ai.news/de/news/2026-07-20/github-sponsors-100-million/ 24 AI de 2026-07-20T00:00:00Z GitHub Sponsors: Globale Open-Source-Entwickler-Community sammelt 100 Millionen US-Dollar an Spenden https://24-ai.news/de/news/2026-07-20/langchain-issuebench-agent-eval/ 24 AI de 2026-07-20T00:00:00Z LangChain: IssueBench misst intern die Erkennung von 15 Fehlerkategorien in Agent-Traces https://24-ai.news/de/news/2026-07-20/openai-long-horizon-model-safety/ 24 AI de 2026-07-20T00:00:00Z OpenAI: Gelernte Lektionen und neue Schutzmaßnahmen für die Sicherheit langfristig agierender (Long-Horizon) KI-Modelle https://24-ai.news/zh/news/2026-07-20/amd-geak-v3-kernel-optimization/ 24 AI zh-cn 2026-07-20T00:00:00Z AMD：GEAK v3开源智能体在Instinct MI300X、MI355X及RDNA4架构上将GPU内核加速3.02倍 https://24-ai.news/zh/news/2026-07-20/anthropic-rare-disease-grants/ 24 AI zh-cn 2026-07-20T00:00:00Z Anthropic：通过AI for Science计划为罕见遗传病研究提供最高5万美元Claude额度 https://24-ai.news/zh/news/2026-07-20/arxiv-coding-agents-arc-agi-3/ 24 AI zh-cn 2026-07-20T00:00:00Z arXiv:2607.15439：验证型编程智能体完全攻克ARC-AGI-3，RHAE约达99% https://24-ai.news/zh/news/2026-07-20/arxiv-craft-rubric-capability-diagnosis/ 24 AI zh-cn 2026-07-20T00:00:00Z arXiv:2607.16122：CRAFT方法聚类评分标准并揭示语言模型的薄弱环节 https://24-ai.news/zh/news/2026-07-20/arxiv-dsworld-data-science-agents/ 24 AI zh-cn 2026-07-20T00:00:00Z arXiv:2607.15901：DSWorld世界模型让数据科学智能体训练与推理最高提速14倍 https://24-ai.news/zh/news/2026-07-20/arxiv-harmonizing-ai-safety-thresholds/ 24 AI zh-cn 2026-07-20T00:00:00Z arXiv:2607.16112：科学家提出统一前沿AI公司安全能力阈值的方法 https://24-ai.news/zh/news/2026-07-20/arxiv-s1-omni-scientific-model/ 24 AI zh-cn 2026-07-20T00:00:00Z arXiv:2607.15686：S1-Omni将科学统一为单一多模态模型，超越GPT-5.5和Gemini https://24-ai.news/zh/news/2026-07-20/arxiv-seerguard-gui-agent-safety/ 24 AI zh-cn 2026-07-20T00:00:00Z arXiv:2607.15550：SeerGuard以安全世界模型预测移动GUI智能体动作的后果 https://24-ai.news/zh/news/2026-07-20/claude-code-v2-1-215-216/ 24 AI zh-cn 2026-07-20T00:00:00Z Anthropic：Claude Code v2.1.215与v2.1.216修复性能问题、OAuth错误及智能体worktree隔离漏洞 https://24-ai.news/zh/news/2026-07-20/cncf-goodput-vs-throughput-llm/ 24 AI zh-cn 2026-07-20T00:00:00Z CNCF：Goodput指标揭示以吞吐量为优化目标的LLM服务p95延迟近乎慢10倍 https://24-ai.news/zh/news/2026-07-20/eu-ai-act-article-50-transparency/ 24 AI zh-cn 2026-07-20T00:00:00Z 欧盟委员会：AI Act第50条透明度指南将于8月2日正式生效 https://24-ai.news/zh/news/2026-07-20/github-code-quality-ga/ 24 AI zh-cn 2026-07-20T00:00:00Z GitHub：Code Quality结合CodeQL分析与Copilot Autofix修复正式GA发布 https://24-ai.news/zh/news/2026-07-20/github-sponsors-100-million/ 24 AI zh-cn 2026-07-20T00:00:00Z GitHub Sponsors：全球开源开发者社区累计筹得1亿美元捐款 https://24-ai.news/zh/news/2026-07-20/langchain-issuebench-agent-eval/ 24 AI zh-cn 2026-07-20T00:00:00Z LangChain：IssueBench内部基准衡量智能体轨迹中15类错误的识别能力 https://24-ai.news/zh/news/2026-07-20/openai-long-horizon-model-safety/ 24 AI zh-cn 2026-07-20T00:00:00Z OpenAI：长时程（long-horizon）AI模型安全的经验教训与新防护措施 https://24-ai.news/ja/news/2026-07-20/amd-geak-v3-kernel-optimization/ 24 AI ja 2026-07-20T00:00:00Z AMD：GEAK v3オープンソースエージェント、Instinct MI300X、MI355X、RDNA4アーキテクチャでGPUカーネルを3.02倍高速化 https://24-ai.news/ja/news/2026-07-20/anthropic-rare-disease-grants/ 24 AI ja 2026-07-20T00:00:00Z Anthropic：AI for Scienceを通じ希少遺伝性疾患研究に最大5万ドルのClaudeクレジットを提供 https://24-ai.news/ja/news/2026-07-20/arxiv-coding-agents-arc-agi-3/ 24 AI ja 2026-07-20T00:00:00Z arXiv:2607.15439：検証型コーディングエージェントがARC-AGI-3を完全攻略、RHAE約99%を達成 https://24-ai.news/ja/news/2026-07-20/arxiv-craft-rubric-capability-diagnosis/ 24 AI ja 2026-07-20T00:00:00Z arXiv:2607.16122：CRAFT手法が評価基準をクラスタリングし言語モデルの弱点を発見 https://24-ai.news/ja/news/2026-07-20/arxiv-dsworld-data-science-agents/ 24 AI ja 2026-07-20T00:00:00Z arXiv:2607.15901：DSWorldワールドモデルがデータサイエンスエージェントの訓練と推論を最大14倍高速化 https://24-ai.news/ja/news/2026-07-20/arxiv-harmonizing-ai-safety-thresholds/ 24 AI ja 2026-07-20T00:00:00Z arXiv:2607.16112：科学者、フロンティアAI企業の安全能力しきい値の調和を提案 https://24-ai.news/ja/news/2026-07-20/arxiv-s1-omni-scientific-model/ 24 AI ja 2026-07-20T00:00:00Z arXiv:2607.15686：S1-Omniが科学を単一のマルチモーダルモデルに統合し、GPT-5.5とGeminiを凌駕 https://24-ai.news/ja/news/2026-07-20/arxiv-seerguard-gui-agent-safety/ 24 AI ja 2026-07-20T00:00:00Z arXiv:2607.15550：SeerGuardが安全世界モデルでモバイルGUIエージェントの行動の結果を予測 https://24-ai.news/ja/news/2026-07-20/claude-code-v2-1-215-216/ 24 AI ja 2026-07-20T00:00:00Z Anthropic：Claude Code v2.1.215とv2.1.216がパフォーマンス、OAuthエラー、エージェントのworktree分離を修正 https://24-ai.news/ja/news/2026-07-20/cncf-goodput-vs-throughput-llm/ 24 AI ja 2026-07-20T00:00:00Z CNCF：GoodputメトリクスがThroughput最適化されたLLMサービングでp95レイテンシがほぼ10倍悪化することを明らかに https://24-ai.news/ja/news/2026-07-20/eu-ai-act-article-50-transparency/ 24 AI ja 2026-07-20T00:00:00Z 欧州委員会：AI Act第50条の透明性ガイドライン、8月2日に発効 https://24-ai.news/ja/news/2026-07-20/github-code-quality-ga/ 24 AI ja 2026-07-20T00:00:00Z GitHub：CodeQL分析とCopilot Autofix修正を組み合わせたCode QualityがGeneral Availableに https://24-ai.news/ja/news/2026-07-20/github-sponsors-100-million/ 24 AI ja 2026-07-20T00:00:00Z GitHub Sponsors：世界のオープンソース開発者コミュニティが1億ドルの寄付を達成 https://24-ai.news/ja/news/2026-07-20/langchain-issuebench-agent-eval/ 24 AI ja 2026-07-20T00:00:00Z LangChain：IssueBenchが内部でエージェントトレース内の15カテゴリのエラー認識を測定 https://24-ai.news/ja/news/2026-07-20/openai-long-horizon-model-safety/ 24 AI ja 2026-07-20T00:00:00Z OpenAI：長時間（long-horizon）AIモデルの安全性における教訓と新たな防護策 https://24-ai.news/ko/news/2026-07-20/amd-geak-v3-kernel-optimization/ 24 AI ko 2026-07-20T00:00:00Z AMD: GEAK v3 오픈소스 에이전트, Instinct MI300X·MI355X 및 RDNA4 아키텍처에서 GPU 커널 3.02배 가속 https://24-ai.news/ko/news/2026-07-20/anthropic-rare-disease-grants/ 24 AI ko 2026-07-20T00:00:00Z Anthropic: AI for Science를 통해 희귀 유전질환 연구에 최대 5만 달러 Claude 크레딧 지원 https://24-ai.news/ko/news/2026-07-20/arxiv-coding-agents-arc-agi-3/ 24 AI ko 2026-07-20T00:00:00Z arXiv:2607.15439: 검증형 코딩 에이전트, ARC-AGI-3 완전 정복—RHAE 약 99% 달성 https://24-ai.news/ko/news/2026-07-20/arxiv-craft-rubric-capability-diagnosis/ 24 AI ko 2026-07-20T00:00:00Z arXiv:2607.16122: CRAFT 방법, 채점 기준을 군집화해 언어 모델의 약점 발견 https://24-ai.news/ko/news/2026-07-20/arxiv-dsworld-data-science-agents/ 24 AI ko 2026-07-20T00:00:00Z arXiv:2607.15901: DSWorld 세계 모델, 데이터 사이언스 에이전트의 훈련과 추론을 최대 14배 가속 https://24-ai.news/ko/news/2026-07-20/arxiv-harmonizing-ai-safety-thresholds/ 24 AI ko 2026-07-20T00:00:00Z arXiv:2607.16112: 과학자들, 프론티어 AI 기업의 안전 역량 임계값 조율 제안 https://24-ai.news/ko/news/2026-07-20/arxiv-s1-omni-scientific-model/ 24 AI ko 2026-07-20T00:00:00Z arXiv:2607.15686: S1-Omni, 과학을 하나의 멀티모달 모델로 통합해 GPT-5.5와 Gemini를 능가 https://24-ai.news/ko/news/2026-07-20/arxiv-seerguard-gui-agent-safety/ 24 AI ko 2026-07-20T00:00:00Z arXiv:2607.15550: SeerGuard, 안전 세계 모델로 모바일 GUI 에이전트 행동의 결과 예측 https://24-ai.news/ko/news/2026-07-20/claude-code-v2-1-215-216/ 24 AI ko 2026-07-20T00:00:00Z Anthropic: Claude Code v2.1.215 및 v2.1.216, 성능·OAuth 오류·에이전트 worktree 격리 수정 https://24-ai.news/ko/news/2026-07-20/cncf-goodput-vs-throughput-llm/ 24 AI ko 2026-07-20T00:00:00Z CNCF: Goodput 지표, 처리량 최적화된 LLM 서빙의 p95 지연이 거의 10배 나쁘다는 사실 드러내 https://24-ai.news/ko/news/2026-07-20/eu-ai-act-article-50-transparency/ 24 AI ko 2026-07-20T00:00:00Z 유럽연합 집행위원회: AI Act 제50조 투명성 가이드라인, 8월 2일 발효 https://24-ai.news/ko/news/2026-07-20/github-code-quality-ga/ 24 AI ko 2026-07-20T00:00:00Z GitHub: CodeQL 분석과 Copilot Autofix 수정을 결합한 Code Quality, 정식 출시(GA) https://24-ai.news/ko/news/2026-07-20/github-sponsors-100-million/ 24 AI ko 2026-07-20T00:00:00Z GitHub Sponsors: 전 세계 오픈소스 개발자 커뮤니티, 누적 후원금 1억 달러 달성 https://24-ai.news/ko/news/2026-07-20/langchain-issuebench-agent-eval/ 24 AI ko 2026-07-20T00:00:00Z LangChain: IssueBench, 에이전트 트레이스 내 15개 오류 범주 인식을 내부적으로 측정 https://24-ai.news/ko/news/2026-07-20/openai-long-horizon-model-safety/ 24 AI ko 2026-07-20T00:00:00Z OpenAI: 장기 지속형(long-horizon) AI 모델 안전성에서 얻은 교훈과 새로운 보호 조치