https://24-ai.news/hr/news/2026-04-28/aisi-ai-models-sabotage-safety-research/ 24 AI hr 2026-04-28T00:00:00Z AISI testirao četiri Claude modela na sabotaži AI safety istraživanja: nema spontane sabotaže, ali Mythos Preview pokazao 65% nesklad između razmišljanja i postupaka https://24-ai.news/hr/news/2026-04-28/arxiv-esrrsim-strategic-reasoning-risks/ 24 AI hr 2026-04-28T00:00:00Z ESRRSim okvir mjeri strateško zaključivanje u 11 reasoning modela: stope detekcije rizika variraju 14,45–72,72%, otkriva i međugeneracijsku evaluation awareness https://24-ai.news/hr/news/2026-04-28/aws-strands-agents-sagemaker-mlflow/ 24 AI hr 2026-04-28T00:00:00Z AWS objavio kako graditi Strands Agente sa SageMaker AI modelima i MLflow observability: SageMakerAIModel provider, autolog tracing i A/B testiranje varijanti https://24-ai.news/hr/news/2026-04-28/github-copilot-usage-based-billing-credits/ 24 AI hr 2026-04-28T00:00:00Z GitHub Copilot prelazi na usage-based naplatu od 1. lipnja: krediti zamjenjuju premium request jedinice, Pro plan dobiva $10 mjesečnih AI Credits https://24-ai.news/hr/news/2026-04-28/langchain-langsmith-eu-ai-act-compliance/ 24 AI hr 2026-04-28T00:00:00Z LangChain i LangSmith ciljaju EU AI Act: alati za sukladnost prema člancima 9, 10, 12-15 i 72 do roka 2. kolovoza 2026. https://24-ai.news/hr/news/2026-04-28/onnx-v1-21-0-opset-26-2bit-types/ 24 AI hr 2026-04-28T00:00:00Z ONNX v1.21.0 izlazi s Opset 26: novi CumProd i BitCast operatori, podrška za 2-bitne tipove i Python 3.14 free-threading eksperiment https://24-ai.news/hr/news/2026-04-28/openai-fedramp-moderate-government/ 24 AI hr 2026-04-28T00:00:00Z OpenAI dobio FedRAMP Moderate autorizaciju: ChatGPT Enterprise i API otvoreni za sigurno usvajanje u američkim federalnim agencijama https://24-ai.news/hr/news/2026-04-28/openai-microsoft-amended-partnership/ 24 AI hr 2026-04-28T00:00:00Z OpenAI i Microsoft objavili izmijenjeni ugovor: nova faza partnerstva s dugoročnom jasnoćom i pojednostavljenom strukturom https://24-ai.news/hr/news/2026-04-28/openai-privacy-filter-pii-detection-15b/ 24 AI hr 2026-04-28T00:00:00Z OpenAI objavio Privacy Filter: 1.5B parametara, Apache 2.0 licenca, 128K kontekst i state-of-the-art detekcija osmero kategorija PII-a u jednom prolazu https://24-ai.news/hr/news/2026-04-28/openai-symphony-codex-orchestration-spec/ 24 AI hr 2026-04-28T00:00:00Z OpenAI objavio Symphony: open-source specifikaciju za orkestraciju Codex agenata koja issue trackere pretvara u 'always-on' inženjerske sustave https://24-ai.news/en/news/2026-04-28/aisi-ai-models-sabotage-safety-research/ 24 AI en 2026-04-28T00:00:00Z AISI tested four Claude models for AI safety research sabotage: no spontaneous sabotage detected, but Mythos Preview shows 65% reasoning-action discrepancy https://24-ai.news/en/news/2026-04-28/arxiv-esrrsim-strategic-reasoning-risks/ 24 AI en 2026-04-28T00:00:00Z ESRRSim framework measures strategic reasoning in 11 models: risk detection rates vary from 14.45% to 72.72%, revealing cross-generational evaluation awareness https://24-ai.news/en/news/2026-04-28/aws-strands-agents-sagemaker-mlflow/ 24 AI en 2026-04-28T00:00:00Z AWS publishes guide for building Strands Agents with SageMaker AI models and MLflow observability: SageMakerAIModel provider, autolog tracing, and A/B variant testing https://24-ai.news/en/news/2026-04-28/github-copilot-usage-based-billing-credits/ 24 AI en 2026-04-28T00:00:00Z GitHub Copilot moves to usage-based billing from June 1: credits replace premium request units, Pro plan receives $10 monthly AI Credits https://24-ai.news/en/news/2026-04-28/langchain-langsmith-eu-ai-act-compliance/ 24 AI en 2026-04-28T00:00:00Z LangChain and LangSmith target EU AI Act: compliance tools mapped to Articles 9, 10, 12-15, and 72 ahead of the August 2, 2026 deadline https://24-ai.news/en/news/2026-04-28/onnx-v1-21-0-opset-26-2bit-types/ 24 AI en 2026-04-28T00:00:00Z ONNX v1.21.0 releases with Opset 26: new CumProd and BitCast operators, 2-bit type support, and Python 3.14 free-threading experiment https://24-ai.news/en/news/2026-04-28/openai-fedramp-moderate-government/ 24 AI en 2026-04-28T00:00:00Z OpenAI receives FedRAMP Moderate authorization: ChatGPT Enterprise and API open for secure adoption by US federal agencies https://24-ai.news/en/news/2026-04-28/openai-microsoft-amended-partnership/ 24 AI en 2026-04-28T00:00:00Z OpenAI and Microsoft announce amended agreement: new partnership phase with long-term clarity and simplified structure https://24-ai.news/en/news/2026-04-28/openai-privacy-filter-pii-detection-15b/ 24 AI en 2026-04-28T00:00:00Z OpenAI releases Privacy Filter: 1.5B parameters, Apache 2.0 license, 128K context, and state-of-the-art detection of eight PII categories in a single pass https://24-ai.news/en/news/2026-04-28/openai-symphony-codex-orchestration-spec/ 24 AI en 2026-04-28T00:00:00Z OpenAI releases Symphony: open-source specification for Codex agent orchestration that turns issue trackers into 'always-on' engineering systems https://24-ai.news/de/news/2026-04-28/aisi-ai-models-sabotage-safety-research/ 24 AI de 2026-04-28T00:00:00Z AISI testete vier Claude-Modelle auf Sabotage von KI-Safety-Forschung: keine spontane Sabotage festgestellt, aber Mythos Preview zeigt 65 % Reasoning-Action-Diskrepanz https://24-ai.news/de/news/2026-04-28/arxiv-esrrsim-strategic-reasoning-risks/ 24 AI de 2026-04-28T00:00:00Z ESRRSim-Framework misst strategisches Reasoning in 11 Reasoning-Modellen: Erkennungsraten variieren von 14,45 % bis 72,72 % und decken generationsübergreifende Evaluation Awareness auf https://24-ai.news/de/news/2026-04-28/aws-strands-agents-sagemaker-mlflow/ 24 AI de 2026-04-28T00:00:00Z AWS veröffentlicht Leitfaden zum Aufbau von Strands Agents mit SageMaker-KI-Modellen und MLflow-Observability: SageMakerAIModel-Provider, Autolog-Tracing und A/B-Variantentests https://24-ai.news/de/news/2026-04-28/github-copilot-usage-based-billing-credits/ 24 AI de 2026-04-28T00:00:00Z GitHub Copilot wechselt ab 1. Juni zu nutzungsbasierter Abrechnung: Credits ersetzen Premium-Request-Einheiten, Pro-Plan erhält monatlich 10 $ KI-Credits https://24-ai.news/de/news/2026-04-28/langchain-langsmith-eu-ai-act-compliance/ 24 AI de 2026-04-28T00:00:00Z LangChain und LangSmith zielen auf den EU AI Act: Compliance-Tools zu Artikeln 9, 10, 12-15 und 72 vor der Frist am 2. August 2026 https://24-ai.news/de/news/2026-04-28/onnx-v1-21-0-opset-26-2bit-types/ 24 AI de 2026-04-28T00:00:00Z ONNX v1.21.0 erscheint mit Opset 26: neue CumProd- und BitCast-Operatoren, 2-Bit-Typen-Unterstützung und Python-3.14-Free-Threading-Experiment https://24-ai.news/de/news/2026-04-28/openai-fedramp-moderate-government/ 24 AI de 2026-04-28T00:00:00Z OpenAI erhält FedRAMP Moderate-Autorisierung: ChatGPT Enterprise und API für sichere Einführung in US-Bundesbehörden geöffnet https://24-ai.news/de/news/2026-04-28/openai-microsoft-amended-partnership/ 24 AI de 2026-04-28T00:00:00Z OpenAI und Microsoft kündigen geänderten Vertrag an: neue Partnerschaftsphase mit langfristiger Klarheit und vereinfachter Struktur https://24-ai.news/de/news/2026-04-28/openai-privacy-filter-pii-detection-15b/ 24 AI de 2026-04-28T00:00:00Z OpenAI veröffentlicht Privacy Filter: 1,5 Mrd. Parameter, Apache-2.0-Lizenz, 128K-Kontext und State-of-the-Art-Erkennung von acht PII-Kategorien in einem Durchlauf https://24-ai.news/de/news/2026-04-28/openai-symphony-codex-orchestration-spec/ 24 AI de 2026-04-28T00:00:00Z OpenAI veröffentlicht Symphony: Open-Source-Spezifikation für die Orchestrierung von Codex-Agenten, die Issue-Tracker in „Always-On”-Ingenieurlösungen verwandelt https://24-ai.news/zh/news/2026-04-28/aisi-ai-models-sabotage-safety-research/ 24 AI zh 2026-04-28T00:00:00Z AISI测试四款Claude模型对AI安全研究的破坏行为：未见自发破坏，但Mythos Preview显示65%推理与行动不一致 https://24-ai.news/zh/news/2026-04-28/arxiv-esrrsim-strategic-reasoning-risks/ 24 AI zh 2026-04-28T00:00:00Z ESRRSim框架测量11个推理模型的战略推理：风险检测率14.45-72.72%，揭示跨代际评估感知 https://24-ai.news/zh/news/2026-04-28/aws-strands-agents-sagemaker-mlflow/ 24 AI zh 2026-04-28T00:00:00Z AWS发布如何使用SageMaker AI模型和MLflow可观测性构建Strands智能体：SageMakerAIModel提供者、自动日志追踪和A/B测试变体 https://24-ai.news/zh/news/2026-04-28/github-copilot-usage-based-billing-credits/ 24 AI zh 2026-04-28T00:00:00Z GitHub Copilot从6月1日起转为按用量计费：AI Credits取代高级请求单元，Pro计划每月获得10美元AI Credits https://24-ai.news/zh/news/2026-04-28/langchain-langsmith-eu-ai-act-compliance/ 24 AI zh 2026-04-28T00:00:00Z LangChain与LangSmith瞄准欧盟AI法案：覆盖第9、10、12-15、72条的合规工具，2026年8月2日截止 https://24-ai.news/zh/news/2026-04-28/onnx-v1-21-0-opset-26-2bit-types/ 24 AI zh 2026-04-28T00:00:00Z ONNX v1.21.0发布，引入Opset 26：新增CumProd和BitCast运算符、2位类型支持及Python 3.14自由线程实验 https://24-ai.news/zh/news/2026-04-28/openai-fedramp-moderate-government/ 24 AI zh 2026-04-28T00:00:00Z OpenAI获得FedRAMP Moderate授权：ChatGPT Enterprise和API向美国联邦机构开放 https://24-ai.news/zh/news/2026-04-28/openai-microsoft-amended-partnership/ 24 AI zh 2026-04-28T00:00:00Z OpenAI与微软公布修订协议：简化合作关系，明确长期框架 https://24-ai.news/zh/news/2026-04-28/openai-privacy-filter-pii-detection-15b/ 24 AI zh 2026-04-28T00:00:00Z OpenAI发布Privacy Filter：15亿参数、Apache 2.0许可证、128K上下文，单次扫描检测八类PII，达到SOTA水平 https://24-ai.news/zh/news/2026-04-28/openai-symphony-codex-orchestration-spec/ 24 AI zh 2026-04-28T00:00:00Z OpenAI发布Symphony：将Issue跟踪器转化为「常驻」工程系统的Codex智能体开源编排规范 https://24-ai.news/ja/news/2026-04-28/aisi-ai-models-sabotage-safety-research/ 24 AI ja 2026-04-28T00:00:00Z AISIがClaudeモデル4種のAI安全研究妨害を評価：自発的妨害なし、しかしMythos Previewは65%で推論と行動に乖離 https://24-ai.news/ja/news/2026-04-28/arxiv-esrrsim-strategic-reasoning-risks/ 24 AI ja 2026-04-28T00:00:00Z ESRRSimフレームワークが11の推論モデルの戦略的推論を測定：リスク検出率14.45〜72.72%、世代間評価意識も明らかに https://24-ai.news/ja/news/2026-04-28/aws-strands-agents-sagemaker-mlflow/ 24 AI ja 2026-04-28T00:00:00Z AWSがSageMaker AIモデルとMLflow可観測性を使ったStrands Agentsの構築方法を公開：SageMakerAIModelプロバイダー、自動ログトレース、A/Bテスト変体 https://24-ai.news/ja/news/2026-04-28/github-copilot-usage-based-billing-credits/ 24 AI ja 2026-04-28T00:00:00Z GitHub Copilotが6月1日から従量課金へ移行：AI CreditsがPremiumリクエスト単位を置き換え、Proプランは月10ドルのAI Creditsを取得 https://24-ai.news/ja/news/2026-04-28/langchain-langsmith-eu-ai-act-compliance/ 24 AI ja 2026-04-28T00:00:00Z LangChainとLangSmithがEU AI法への対応を発表：第9条、10条、12〜15条、72条をカバーするコンプライアンスツール、期限は2026年8月2日 https://24-ai.news/ja/news/2026-04-28/onnx-v1-21-0-opset-26-2bit-types/ 24 AI ja 2026-04-28T00:00:00Z ONNX v1.21.0リリース、Opset 26を導入：新しいCumProdとBitCastオペレーター、2ビット型サポート、Python 3.14フリースレッディング実験 https://24-ai.news/ja/news/2026-04-28/openai-fedramp-moderate-government/ 24 AI ja 2026-04-28T00:00:00Z OpenAIがFedRAMP Moderate認定を取得：ChatGPT EnterpriseとAPIが米国連邦機関に開放 https://24-ai.news/ja/news/2026-04-28/openai-microsoft-amended-partnership/ 24 AI ja 2026-04-28T00:00:00Z OpenAIとMicrosoftが改定契約を発表：パートナーシップを簡素化し長期的な明確性を確保 https://24-ai.news/ja/news/2026-04-28/openai-privacy-filter-pii-detection-15b/ 24 AI ja 2026-04-28T00:00:00Z OpenAIがPrivacy Filterを発表：15億パラメーター、Apache 2.0ライセンス、128Kコンテキスト、1回のスキャンで8カテゴリーのPIIを検出しSOTA達成 https://24-ai.news/ja/news/2026-04-28/openai-symphony-codex-orchestration-spec/ 24 AI ja 2026-04-28T00:00:00Z OpenAIがSymphonyを発表：IssueトラッカーをAlways-Onエンジニアリングシステムに変えるCodexエージェント編成オープンソース仕様 https://24-ai.news/ko/news/2026-04-28/aisi-ai-models-sabotage-safety-research/ 24 AI ko 2026-04-28T00:00:00Z AISI, Claude 모델 4종의 AI 안전 연구 방해 행위 테스트: 자발적 방해 없음, 그러나 Mythos Preview는 65%에서 추론-행동 불일치 나타내 https://24-ai.news/ko/news/2026-04-28/arxiv-esrrsim-strategic-reasoning-risks/ 24 AI ko 2026-04-28T00:00:00Z ESRRSim 프레임워크, 11개 추론 모델의 전략적 추론 측정: 위험 감지율 14.45-72.72%, 세대 간 평가 인식도 드러내 https://24-ai.news/ko/news/2026-04-28/aws-strands-agents-sagemaker-mlflow/ 24 AI ko 2026-04-28T00:00:00Z AWS, SageMaker AI 모델과 MLflow 관찰 가능성을 이용한 Strands 에이전트 구축 방법 공개: SageMakerAIModel 공급자, 자동 로그 추적, A/B 테스트 변체 https://24-ai.news/ko/news/2026-04-28/github-copilot-usage-based-billing-credits/ 24 AI ko 2026-04-28T00:00:00Z GitHub Copilot, 6월 1일부터 사용량 기반 요금제로 전환: AI Credits가 프리미엄 요청 단위 대체, Pro 플랜은 월 10달러 AI Credits 제공 https://24-ai.news/ko/news/2026-04-28/langchain-langsmith-eu-ai-act-compliance/ 24 AI ko 2026-04-28T00:00:00Z LangChain과 LangSmith, EU AI법 대응 발표: 제9, 10, 12-15, 72조 커버 컴플라이언스 도구, 2026년 8월 2일 마감 https://24-ai.news/ko/news/2026-04-28/onnx-v1-21-0-opset-26-2bit-types/ 24 AI ko 2026-04-28T00:00:00Z ONNX v1.21.0 출시, Opset 26 도입: 새로운 CumProd와 BitCast 연산자, 2비트 타입 지원 및 Python 3.14 자유 스레딩 실험 https://24-ai.news/ko/news/2026-04-28/openai-fedramp-moderate-government/ 24 AI ko 2026-04-28T00:00:00Z OpenAI, FedRAMP Moderate 인증 취득: ChatGPT Enterprise와 API가 미국 연방 기관에 개방 https://24-ai.news/ko/news/2026-04-28/openai-microsoft-amended-partnership/ 24 AI ko 2026-04-28T00:00:00Z OpenAI와 Microsoft, 수정 계약 발표: 파트너십 간소화 및 장기적 명확성 확보 https://24-ai.news/ko/news/2026-04-28/openai-privacy-filter-pii-detection-15b/ 24 AI ko 2026-04-28T00:00:00Z OpenAI Privacy Filter 발표: 15억 파라미터, Apache 2.0, 128K 컨텍스트, 단일 스캔으로 8개 PII 카테고리 감지하여 SOTA 달성 https://24-ai.news/ko/news/2026-04-28/openai-symphony-codex-orchestration-spec/ 24 AI ko 2026-04-28T00:00:00Z OpenAI, Symphony 발표: Issue 트래커를 '상시 가동' 엔지니어링 시스템으로 전환하는 Codex 에이전트 오픈소스 오케스트레이션 사양 https://24-ai.news/hr/news/2026-04-27/arxiv-agentic-ai-science-automation/ 24 AI hr 2026-04-27T00:00:00Z ArXiv 2604.21910: agentic AI automatizira znanstveni workflow s 83 % točnosti, 92 % manje data transfera i $0.001 po upitu https://24-ai.news/hr/news/2026-04-27/arxiv-agentic-world-modeling-survey/ 24 AI hr 2026-04-27T00:00:00Z ArXiv 2604.22748: Survey 42 autora uvodi 'levels × laws' taksonomiju za world modele u AI agentima — sinteza 400+ radova https://24-ai.news/hr/news/2026-04-27/arxiv-biominer-protein-ligand-drug-discovery/ 24 AI hr 2026-04-27T00:00:00Z ArXiv 2604.21508 BioMiner: multimodalni AI vadi protein-ligand bioaktivnost iz literature, 5,59× brže od ručnog rada https://24-ai.news/hr/news/2026-04-27/arxiv-ocp-distributed-inference-failures/ 24 AI hr 2026-04-27T00:00:00Z ArXiv 2604.21361: Open Compute Project mapira time/causality kvarove u distribuiranim AI inference sustavima — 5 ms clock skew lomi observability https://24-ai.news/hr/news/2026-04-27/arxiv-separable-expert-gdpr-llm-personalization/ 24 AI hr 2026-04-27T00:00:00Z ArXiv 2604.21571 'Separable Expert': arhitektura za LLM personalizaciju koja omogućuje GDPR pravo na zaborav bez retreniranja https://24-ai.news/hr/news/2026-04-27/arxiv-superminds-test-collective-intelligence/ 24 AI hr 2026-04-27T00:00:00Z ArXiv 2604.22452: Superminds Test pokazuje da kolektivna inteligencija ne nastaje sama u društvu od 2 milijuna AI agenata https://24-ai.news/hr/news/2026-04-27/arxiv-thinking-reasoning-skills-fewer-tokens/ 24 AI hr 2026-04-27T00:00:00Z ArXiv 2604.21764: 'Thinking with Reasoning Skills' smanjuje broj reasoning tokena uz veću točnost na ACL 2026 Industry Tracku https://24-ai.news/hr/news/2026-04-27/deepmind-korea-partnership-ai-campus/ 24 AI hr 2026-04-27T00:00:00Z Google DeepMind i Južna Koreja osnivaju AI Campus u Seulu deset godina nakon AlphaGo meča https://24-ai.news/hr/news/2026-04-27/github-app-installation-tokens-novi-format/ 24 AI hr 2026-04-27T00:00:00Z GitHub mijenja format App installation tokena: s 40 na ~520 znakova, breakage rizik za CI/CD pipelineove https://24-ai.news/hr/news/2026-04-27/github-copilot-gpt-5-5-ga/ 24 AI hr 2026-04-27T00:00:00Z GitHub Copilot dobio GPT-5.5 GA: dostupan na svim glavnim IDE-ovima uz 7,5× premium multiplier https://24-ai.news/hr/news/2026-04-27/openai-our-principles-agi-deklaracija/ 24 AI hr 2026-04-27T00:00:00Z OpenAI objavio dokument 'Our principles': pet temeljnih načela koja vode put prema AGI-u https://24-ai.news/en/news/2026-04-27/arxiv-agentic-ai-science-automation/ 24 AI en 2026-04-27T00:00:00Z ArXiv 2604.21910: Agentic AI automates scientific workflow with 83% accuracy, 92% less data transfer and $0.001 per query https://24-ai.news/en/news/2026-04-27/arxiv-agentic-world-modeling-survey/ 24 AI en 2026-04-27T00:00:00Z arXiv 2604.22748: Survey by 42 authors introduces 'levels × laws' taxonomy for world models in AI agents — synthesis of 400+ papers https://24-ai.news/en/news/2026-04-27/arxiv-biominer-protein-ligand-drug-discovery/ 24 AI en 2026-04-27T00:00:00Z ArXiv 2604.21508 BioMiner: multimodal AI extracts protein-ligand bioactivity from literature, 5.59× faster than manual work https://24-ai.news/en/news/2026-04-27/arxiv-ocp-distributed-inference-failures/ 24 AI en 2026-04-27T00:00:00Z ArXiv 2604.21361: Open Compute Project maps time/causality failures in distributed AI inference systems — 5 ms clock skew breaks observability https://24-ai.news/en/news/2026-04-27/arxiv-separable-expert-gdpr-llm-personalization/ 24 AI en 2026-04-27T00:00:00Z ArXiv 2604.21571 'Separable Expert': architecture for LLM personalization enabling GDPR right to erasure without retraining https://24-ai.news/en/news/2026-04-27/arxiv-superminds-test-collective-intelligence/ 24 AI en 2026-04-27T00:00:00Z arXiv 2604.22452: Superminds Test shows collective intelligence does not emerge spontaneously in a society of 2 million AI agents https://24-ai.news/en/news/2026-04-27/arxiv-thinking-reasoning-skills-fewer-tokens/ 24 AI en 2026-04-27T00:00:00Z ArXiv 2604.21764: 'Thinking with Reasoning Skills' reduces reasoning tokens while improving accuracy — ACL 2026 Industry Track https://24-ai.news/en/news/2026-04-27/deepmind-korea-partnership-ai-campus/ 24 AI en 2026-04-27T00:00:00Z Google DeepMind and South Korea establish AI Campus in Seoul ten years after the AlphaGo match https://24-ai.news/en/news/2026-04-27/github-app-installation-tokens-novi-format/ 24 AI en 2026-04-27T00:00:00Z GitHub changes App installation token format: from 40 to ~520 characters, breakage risk for CI/CD pipelines https://24-ai.news/en/news/2026-04-27/github-copilot-gpt-5-5-ga/ 24 AI en 2026-04-27T00:00:00Z GitHub Copilot receives GPT-5.5 GA: available on all major IDEs with 7.5× premium multiplier https://24-ai.news/en/news/2026-04-27/openai-our-principles-agi-deklaracija/ 24 AI en 2026-04-27T00:00:00Z OpenAI publishes 'Our principles' document: five foundational principles guiding the path toward AGI https://24-ai.news/de/news/2026-04-27/arxiv-agentic-ai-science-automation/ 24 AI de 2026-04-27T00:00:00Z ArXiv 2604.21910: Agentic AI automatisiert wissenschaftliche Workflows mit 83 % Genauigkeit, 92 % weniger Datentransfer und $0,001 pro Anfrage https://24-ai.news/de/news/2026-04-27/arxiv-agentic-world-modeling-survey/ 24 AI de 2026-04-27T00:00:00Z arXiv 2604.22748: Survey von 42 Autoren führt Taxonomie ‚levels × laws' für World Models in AI-Agenten ein — Synthese aus über 400 Arbeiten https://24-ai.news/de/news/2026-04-27/arxiv-biominer-protein-ligand-drug-discovery/ 24 AI de 2026-04-27T00:00:00Z ArXiv 2604.21508 BioMiner: multimodales KI-System extrahiert Protein-Ligand-Bioaktivität aus der Literatur, 5,59× schneller als manuelle Arbeit https://24-ai.news/de/news/2026-04-27/arxiv-ocp-distributed-inference-failures/ 24 AI de 2026-04-27T00:00:00Z ArXiv 2604.21361: Open Compute Project kartiert Zeit-/Kausalitätsfehler in verteilten KI-Inferenzsystemen — 5 ms Clock-Skew bricht Observability https://24-ai.news/de/news/2026-04-27/arxiv-separable-expert-gdpr-llm-personalization/ 24 AI de 2026-04-27T00:00:00Z ArXiv 2604.21571 'Separable Expert': Architektur für LLM-Personalisierung, die DSGVO-Recht auf Vergessenwerden ohne Retraining ermöglicht https://24-ai.news/de/news/2026-04-27/arxiv-superminds-test-collective-intelligence/ 24 AI de 2026-04-27T00:00:00Z arXiv 2604.22452: Superminds Test zeigt, dass kollektive Intelligenz in einer Gesellschaft von 2 Millionen AI-Agenten nicht spontan entsteht https://24-ai.news/de/news/2026-04-27/arxiv-thinking-reasoning-skills-fewer-tokens/ 24 AI de 2026-04-27T00:00:00Z ArXiv 2604.21764: 'Thinking with Reasoning Skills' reduziert Reasoning-Token bei höherer Genauigkeit — ACL 2026 Industry Track https://24-ai.news/de/news/2026-04-27/deepmind-korea-partnership-ai-campus/ 24 AI de 2026-04-27T00:00:00Z Google DeepMind und Südkorea gründen AI Campus in Seoul — zehn Jahre nach dem AlphaGo-Match https://24-ai.news/de/news/2026-04-27/github-app-installation-tokens-novi-format/ 24 AI de 2026-04-27T00:00:00Z GitHub ändert Format der App-Installationstokens: von 40 auf ~520 Zeichen, Bruchrisiko für CI/CD-Pipelines https://24-ai.news/de/news/2026-04-27/github-copilot-gpt-5-5-ga/ 24 AI de 2026-04-27T00:00:00Z GitHub Copilot erhält GPT-5.5 GA: verfügbar in allen wichtigen IDEs mit 7,5× Premium-Multiplikator https://24-ai.news/de/news/2026-04-27/openai-our-principles-agi-deklaracija/ 24 AI de 2026-04-27T00:00:00Z OpenAI veröffentlicht Dokument 'Our principles': fünf grundlegende Prinzipien auf dem Weg zur AGI https://24-ai.news/zh/news/2026-04-27/arxiv-agentic-ai-science-automation/ 24 AI zh 2026-04-27T00:00:00Z ArXiv 2604.21910：智能体AI以83%准确率自动化科学工作流，数据传输减少92%，每次查询费用$0.001 https://24-ai.news/zh/news/2026-04-27/arxiv-agentic-world-modeling-survey/ 24 AI zh 2026-04-27T00:00:00Z ArXiv 2604.22748：42位作者的综述引入「层级×定律」分类框架——整合逾400篇论文，构建AI智能体世界模型分类体系 https://24-ai.news/zh/news/2026-04-27/arxiv-biominer-protein-ligand-drug-discovery/ 24 AI zh 2026-04-27T00:00:00Z ArXiv 2604.21508 BioMiner：多模态AI从文献中提取蛋白质-配体生物活性，比人工快5.59倍 https://24-ai.news/zh/news/2026-04-27/arxiv-ocp-distributed-inference-failures/ 24 AI zh 2026-04-27T00:00:00Z ArXiv 2604.21361：Open Compute Project映射分布式AI推理系统中的时间/因果故障——5毫秒时钟偏差破坏可观测性 https://24-ai.news/zh/news/2026-04-27/arxiv-separable-expert-gdpr-llm-personalization/ 24 AI zh 2026-04-27T00:00:00Z ArXiv 2604.21571 'Separable Expert'：无需重新训练即可实现GDPR被遗忘权的LLM个性化架构 https://24-ai.news/zh/news/2026-04-27/arxiv-superminds-test-collective-intelligence/ 24 AI zh 2026-04-27T00:00:00Z ArXiv 2604.22452：超级思维测试揭示——200万AI智能体构成的社会中，集体智能并不会自然涌现 https://24-ai.news/zh/news/2026-04-27/arxiv-thinking-reasoning-skills-fewer-tokens/ 24 AI zh 2026-04-27T00:00:00Z ArXiv 2604.21764：'Thinking with Reasoning Skills'在ACL 2026工业赛道减少推理token数的同时提升准确率 https://24-ai.news/zh/news/2026-04-27/deepmind-korea-partnership-ai-campus/ 24 AI zh 2026-04-27T00:00:00Z Google DeepMind与韩国在AlphaGo对局十周年之际，于首尔共建AI园区 https://24-ai.news/zh/news/2026-04-27/github-app-installation-tokens-novi-format/ 24 AI zh 2026-04-27T00:00:00Z GitHub更改App安装令牌格式：从40字符增至约520字符，CI/CD管道存在故障风险 https://24-ai.news/zh/news/2026-04-27/github-copilot-gpt-5-5-ga/ 24 AI zh 2026-04-27T00:00:00Z GitHub Copilot获得GPT-5.5 GA：在所有主要IDE上可用，7.5倍高级积分倍数 https://24-ai.news/zh/news/2026-04-27/openai-our-principles-agi-deklaracija/ 24 AI zh 2026-04-27T00:00:00Z OpenAI发布文件'Our principles'：通往AGI之路的五项基本原则 https://24-ai.news/ja/news/2026-04-27/arxiv-agentic-ai-science-automation/ 24 AI ja 2026-04-27T00:00:00Z ArXiv 2604.21910：Agentic AIが科学ワークフローを83%の精度で自動化、データ転送92%削減、クエリあたり$0.001 https://24-ai.news/ja/news/2026-04-27/arxiv-agentic-world-modeling-survey/ 24 AI ja 2026-04-27T00:00:00Z arXiv 2604.22748：42名著者によるサーベイが「レベル×法則」分類体系を導入——400本超の論文を統合したAIエージェントの世界モデル研究 https://24-ai.news/ja/news/2026-04-27/arxiv-biominer-protein-ligand-drug-discovery/ 24 AI ja 2026-04-27T00:00:00Z ArXiv 2604.21508 BioMiner：マルチモーダルAIが文献からタンパク質-リガンド生物活性を抽出、人工作業比5.59倍高速 https://24-ai.news/ja/news/2026-04-27/arxiv-ocp-distributed-inference-failures/ 24 AI ja 2026-04-27T00:00:00Z ArXiv 2604.21361：Open Compute Projectが分散AI推論システムの時間/因果障害をマッピング——5msのクロックスキューが可観測性を破壊 https://24-ai.news/ja/news/2026-04-27/arxiv-separable-expert-gdpr-llm-personalization/ 24 AI ja 2026-04-27T00:00:00Z ArXiv 2604.21571 'Separable Expert'：再トレーニング不要でGDPR忘れられる権利を実現するLLMパーソナライゼーションアーキテクチャ https://24-ai.news/ja/news/2026-04-27/arxiv-superminds-test-collective-intelligence/ 24 AI ja 2026-04-27T00:00:00Z arXiv 2604.22452：スーパーマインドテストが示す——200万体のAIエージェント社会でも集合知は自然発生しない https://24-ai.news/ja/news/2026-04-27/arxiv-thinking-reasoning-skills-fewer-tokens/ 24 AI ja 2026-04-27T00:00:00Z ArXiv 2604.21764：'Thinking with Reasoning Skills'がACL 2026 Industry Trackでトークン数削減と精度向上を同時に実現 https://24-ai.news/ja/news/2026-04-27/deepmind-korea-partnership-ai-campus/ 24 AI ja 2026-04-27T00:00:00Z Google DeepMindと韓国、AlphaGo対局10周年を機にソウルにAIキャンパスを設立 https://24-ai.news/ja/news/2026-04-27/github-app-installation-tokens-novi-format/ 24 AI ja 2026-04-27T00:00:00Z GitHubがApp installationトークン形式を変更：40文字から約520文字へ、CI/CDパイプラインに障害リスク https://24-ai.news/ja/news/2026-04-27/github-copilot-gpt-5-5-ga/ 24 AI ja 2026-04-27T00:00:00Z GitHub CopilotがGPT-5.5 GAを取得：全主要IDEで利用可能、7.5倍プレミアム倍率 https://24-ai.news/ja/news/2026-04-27/openai-our-principles-agi-deklaracija/ 24 AI ja 2026-04-27T00:00:00Z OpenAIが「Our principles」を発表：AGIへの道を導く5つの基本原則 https://24-ai.news/ko/news/2026-04-27/arxiv-agentic-ai-science-automation/ 24 AI ko 2026-04-27T00:00:00Z ArXiv 2604.21910: Agentic AI가 83% 정확도로 과학 워크플로를 자동화, 데이터 전송 92% 감소, 쿼리당 비용 $0.001 https://24-ai.news/ko/news/2026-04-27/arxiv-agentic-world-modeling-survey/ 24 AI ko 2026-04-27T00:00:00Z arXiv 2604.22748: 42명 저자의 서베이, AI 에이전트 세계 모델을 위한 '레벨×법칙' 분류 체계 제시 — 400편 이상 논문 종합 https://24-ai.news/ko/news/2026-04-27/arxiv-biominer-protein-ligand-drug-discovery/ 24 AI ko 2026-04-27T00:00:00Z ArXiv 2604.21508 BioMiner: 멀티모달 AI가 문헌에서 단백질-리간드 생물활성 추출, 수동 작업 대비 5.59배 빠름 https://24-ai.news/ko/news/2026-04-27/arxiv-ocp-distributed-inference-failures/ 24 AI ko 2026-04-27T00:00:00Z ArXiv 2604.21361: Open Compute Project가 분산 AI 추론 시스템의 시간/인과 장애 매핑——5ms 클록 스큐가 관측 가능성을 파괴 https://24-ai.news/ko/news/2026-04-27/arxiv-separable-expert-gdpr-llm-personalization/ 24 AI ko 2026-04-27T00:00:00Z ArXiv 2604.21571 'Separable Expert': 재학습 없이 GDPR 잊혀질 권리를 구현하는 LLM 개인화 아키텍처 https://24-ai.news/ko/news/2026-04-27/arxiv-superminds-test-collective-intelligence/ 24 AI ko 2026-04-27T00:00:00Z arXiv 2604.22452: 슈퍼마인드 테스트가 밝힌다 — 200만 AI 에이전트 사회에서도 집단 지능은 자연 발생하지 않는다 https://24-ai.news/ko/news/2026-04-27/arxiv-thinking-reasoning-skills-fewer-tokens/ 24 AI ko 2026-04-27T00:00:00Z ArXiv 2604.21764: 'Thinking with Reasoning Skills'가 ACL 2026 Industry Track에서 추론 토큰 감소와 정확도 향상 동시 달성 https://24-ai.news/ko/news/2026-04-27/deepmind-korea-partnership-ai-campus/ 24 AI ko 2026-04-27T00:00:00Z Google DeepMind와 한국, AlphaGo 대국 10주년을 기념해 서울에 AI 캠퍼스 설립 https://24-ai.news/ko/news/2026-04-27/github-app-installation-tokens-novi-format/ 24 AI ko 2026-04-27T00:00:00Z GitHub App 설치 토큰 형식 변경: 40자에서 약 520자로, CI/CD 파이프라인 장애 위험 https://24-ai.news/ko/news/2026-04-27/github-copilot-gpt-5-5-ga/ 24 AI ko 2026-04-27T00:00:00Z GitHub Copilot에 GPT-5.5 GA 출시: 모든 주요 IDE에서 사용 가능, 7.5배 프리미엄 배율 https://24-ai.news/ko/news/2026-04-27/openai-our-principles-agi-deklaracija/ 24 AI ko 2026-04-27T00:00:00Z OpenAI, 'Our principles' 문서 발표: AGI를 향한 길을 이끄는 다섯 가지 기본 원칙