아카이브
시간순으로 보기
AI와 개발 분야에서 어떤 주제가 반복됐고, 어느 시점에 방향이 바뀌었는지 순서대로 확인합니다.
2026
290개Moonshot AI, Kimi K3 공개 weights 배포: 2.8T MoE의 실제 serving 경로 확인NVIDIA·업계, Open Secure AI Alliance 출범: 공개 방어 스택·NOOA 추진OpenAI, Health in ChatGPT 출시: 개인 건강 데이터 연결과 별도 보호 계층OpenAI Presence: 정책·평가·승인 루프로 Enterprise Agent 운영 제품화Anthropic, Claude Sonnet 5 출시: 에이전트 성능과 비용 효율 강화Google Cloud, AlphaEvolve GA: 채점 함수 기반 코드 최적화 에이전트Google Cloud, k8s-aibom 오픈소스화: 런타임 ML-BOM 자동 생성NVIDIA, Cosmos 3 Edge 공개: 4B 온디바이스 Physical AI 모델Vercel Ship 2026: AI SDK 7·Connect·eve로 에이전트 스택 확장Google, Gemini 3.6 Flash·3.5 Flash-Lite·3.5 Flash Cyber 공개Moonshot AI, Kimi K3 공개: 2.8T open 3T-class 모델OpenAI·Hugging Face, 모델 평가 중 AI 에이전트 보안 사고 공개UN AI Scientific Panel, Preliminary Report로 글로벌 거버넌스 논의 지원GitHub Models, 7월 30일 전면 종료…API·BYOK 모두 중단Moonshot, 2.8T Kimi K3 공개…7월 27일 full weights 배포Thinking Machines, 975B MoE 오픈 웨이트 Inkling 공개Fireworks, USD 1.505B Series D로 custom-model 플랫폼 확장GitHub Copilot code review, 하네스·도구 지침 최적화로 평균 비용 약 20% 절감Mistral, 단일 RGB 카메라 로봇 내비게이션 모델 Robostral Navigate 공개Chai Discovery·argenx, de novo antibody 설계를 위한 AI 플랫폼 협업Google Ads, 생성형 AI 사용을 표시하는 ‘How this ad was made’ 공개Prime Intellect, 오픈형 agentic RL 스택 확장에 USD 130M Series AGoogle, Qwen 3.5-397B MoE를 Ironwood TPU에서 최적화: prefill 최대 4.7배 향상SambaNova, Series F 첫 클로징으로 USD 1B 유치…AI inference 인프라에 USD 11B 가치UN 독립 AI 과학 패널, 기회·위험·영향을 다룬 첫 예비 보고서 공개Google, Gemini 3.5 Flash computer use와 Gemini Omni Flash API 공개Helsing, USD 1.8B Series E로 USD 18B 가치 평가…유럽 AI 국방 기술 투자 가속Meta, Muse Image·Muse Video 공개: 툴 사용·자기개선형 미디어 생성 모델GitLost: We Tricked GitHub's AI Agent into Leaking Private ReposClaude Code sends 33k tokens before reading the prompt; OpenCode sends 7kI think I have LLM burnoutGPT-5.6 Sol Ultra produces proof of the Cycle Double Cover Conjecture [pdf]Show HN: Getting GLM 5.2 running on my slow computerApple sues OpenAI, accuses ex-employees of stealing trade secretsGPT‑LiveGPT-5.6Jamesob's guide to running SOTA LLMs locallyZuckerberg says AI agent development going slower than expectedLeanstral 1.5 - open proof-engineering model moves formal verification toward practical code reviewOpenAI GeneBench-Pro - computational biology agents get a harder benchmark for research judgmentVercel AI Gateway voice support - realtime multimodal agents move into the same routing plane as text modelsAlibaba to ban Claude Code in workplace over alleged backdoor risks, source saysMicrosoft's open source tools were hacked to steal passwords of AI developersAI for Good Global Commission - UN/ITU moves AI governance toward executive coordinationClaude Science - Anthropic turns scientific workflows into auditable agent workbenchesEpoch CVE spike - frontier cyber models reshape vulnerability disclosure volumeWafer GLM5.2 on AMD MI355X - ROCm inference narrows the CUDA deployment gapNoam Shazeer Joins OpenAIZCode – Harness for GLM-5.2GitHub Copilot Browser Tools GA - coding agents get controlled live-browser execution in VS CodeMistral OCR 4 - document parsing shifts from text extraction to structured AI ingestionScarfBench - enterprise Java migration exposes coding-agent behavior gapsApple reveals new AI architecture built around Google Gemini modelsOpenAI unveils its first custom chip, built by BroadcomEtched frontier inference clusters - transformer ASIC startup moves toward rack-scale productionF-Droid ADV critique - Android app distribution control becomes developer policy riskGitHub Models retirement - standalone model API gives way to Copilot and Azure AI FoundryMixedbread asymmetric quantization - late-interaction retrieval cuts vector storage 97 percentGemma 4 12B: A unified, encoder-free multimodal modelDepartment of Commerce has lifted export controls on Claude Fable 5 and Mythos 5Godot AI code policy - OSS maintainers draw a hard review boundaryMeta AI compute cloud - hyperscalers turn excess capacity into a marketVenice AI Series A - privacy-first AI platforms become infrastructure businessesClaude Sonnet 5Claude Code is steganographically marking requestsAWS FDE - agentic AI deployment moves from tooling to embedded engineeringBase44 Base1 - vibe-coding platforms chase vertical model ownershipEvery Eval Ever - fragmented AI benchmarks get a shared metadata layerX hosted MCP - social data APIs become agent-native infrastructureGLM 5.2 beats Claude in our benchmarksIdentity verification on ClaudeAnthropic says Alibaba illicitly extracted Claude AI model capabilitiesU.S. government will decide who gets to use GPT-5.6GitHub Advisory Database surge - vulnerability disclosure pipelines hit AI-era scaleOlmoLogic - Prolog-verifiable RLVR improves open logical reasoningVLX-Go - lightweight vision-language waypoints for embodied navigationAI RFIC inverse design - reinforcement learning reaches radio-chip layout bottlenecksAzure Copilot Observability Agent - cloud ops moves from alerts to real-time reasoningBIS AI investment warning - AI capex becomes a macro-financial stability riskGemini 3.5 Flash Computer Use - screen-driving agents move into the main modelAWS Lambda MicroVMs - serverless sandboxes target AI-generated code executionDeepSpec - speculative decoding becomes an open production optimization stackGPT-5.6 Sol preview - frontier model releases become policy-gated infrastructure decisionsGeneral Intuition Series A - gameplay data becomes the next action-model training substrateHF Jobs vLLM server - throwaway OpenAI-compatible endpoints get pay-per-second GPUsQHexRT - Qualcomm Hexagon NPU inference moves small LLMs fully on-deviceClaude Tag - Slack-native team agents move from private assistants to shared workspacesGLM-5.2 - open long-context models push agentic coding toward 1M-token workspacesMicrosoft AutoJack - browsing agents expose local MCP control planes to RCEFFASR Leaderboard - voice AI benchmarks move from clean speech to far-field realityKog Laneformer 2B - latency-first coding models move architecture into the serving layerKrea 2 technical report - open image models compete on creative control, not only fidelityNVIDIA NeMo AutoModel - MoE fine-tuning gets a drop-in performance path for TransformersFika Jobs - AI interview agents expose the product-risk tradeoff in hiring automationGoogle Jules evals - coding agents need insight-policy benchmarks, not just SWE-bench taskshuggingface_hub weekly release CI - open-weight agents make release automation auditableOpenAI Patch the Planet - AI-assisted security needs maintainer-controlled remediation loopsIntel XPU Kernel Skill - coding agents optimize Triton kernels beyond CUDA-first defaultsMosaicLeaks - deep research agents can leak private facts through harmless-looking searchesPP-OCRv6 on Hugging Face - document AI stays specialized, small, and multilingualReflection-SpaceX compute deal - open-source frontier AI hits a capacity wallArcade Series A — enterprise agents need an authorization layer, not just MCP gatewaysCloudflare Temporary Accounts — coding agents can deploy Workers without human signup flowGitHub Code Quality GA — code governance becomes subscription plus AI meteringNVIDIA Cannes AI marketing stack — agentic workflows move into campaign operationsAdani-Jabil AI infra alliance — AI 경쟁이 모델에서 전력·랙·제조 공급망으로 확장된다JEP 401 Value Classes — Java object model이 identity-free domain value로 이동한다Norway school AI restrictions — 초등 AI 금지가 교육용 AI 확산의 반작용을 보여준다Salesforce-Fin acquisition — customer service agents가 CRM suite의 핵심 실행 계층으로 편입된다Anthropic Public Record — 미국 대중은 AI 효용보다 책임성과 규제를 먼저 요구한다ChatGPT Enterprise spend controls — AI 도입의 병목이 모델 접근에서 비용 거버넌스로 이동MAI-Code-1-Flash 확장 — coding model 경쟁이 Copilot surface coverage로 이동OpenAI AI chemist — GPT-5.4가 자동화 실험실과 결합해 Chan-Lam 수율을 개선Google UCP open rails — agentic commerce가 쇼핑 UI에서 표준 프로토콜 경쟁으로 이동OpenAI June 2026 Threat Report — AI 논쟁 자체가 영향공작 표적이 됐다Probably $9M seed — AI 신뢰성 경쟁이 더 큰 모델에서 deterministic harness engineering으로 이동Google Colab CLI — agent-ready compute가 로컬 터미널에서 즉시 GPU·TPU orchestration으로 이동OpenEnv committee launch — open agent training이 harness별 튜닝에서 공유 environment protocol로 이동Prometheus $12B Series B — industrial AI가 chatbot에서 physical engineering cycle compression으로 이동AI brands as bait — AI 열풍이 모델 출시 경쟁에서 social engineering 공격면 확대로 번지다GitHub Agentic Workflows public preview — 에이전트 자동화가 YAML 작성에서 policy-aware SDLC 실행 계층으로 이동GitHub Copilot CLI + language servers — AI 코딩이 text grep에서 semantic code intelligence 단계로 이동AI in the Enterprise: How People Use M365 Copilot Chat — enterprise AI 채택이 검색 보조에서 문서·커뮤니케이션 작업으로 이동Cloudflare acquires VoidZero — AI 코딩 시대의 배포 스택이 framework 선택에서 execution path 통합으로 이동OpenRouter·Concentrate AI 부상 — LLM 경쟁이 모델 성능에서 routing economics 계층으로 이동Claude Fable 5 — frontier model 공개가 capability race에서 guardrailed deployment 경쟁으로 이동OpenAI S-1 confidential filing — AI 경쟁이 모델·제품 전쟁에서 자본시장 체력전으로 이동Salesforce Agentforce layoffs — enterprise AI가 성장 서사에서 조직 재편과 제품 현실성 검증 단계로 이동Dreaming: Better memory for a more helpful ChatGPT — AI personal memory가 saved note에서 지속적 user model로 전환ECB AI risk letter — 금융권 AI 도입이 pilot enthusiasm에서 board-level defensive posture로 이동Introducing Mellum2 — software engineering용 small expert model 경쟁이 giant general model에서 low-latency control layer로 이동US House AI draft bill — 미국 AI 규제 경쟁이 state patchwork에서 federal model-development preemption으로 이동IBM-Google Cloud Practice — enterprise agent 도입이 PoC에서 서비스 채널과 산업별 delivery asset 경쟁으로 이동Ollama 0.30 — local AI 배포 경쟁이 모델 자체에서 runtime 호환성과 GPU 보편성으로 이동WWDC26 Apple Intelligence APIs — on-device model access가 앱 기능에서 workflow substrate로 확장IBM and Red Hat Project Lightwell — open source AI 시대의 공급망 보안이 clearinghouse 모델로 재편Introducing Gemma 4 12B — local multimodal agent 실행이 16GB급 엣지 하드웨어로 내려오다Meta Business Agent — customer support agent가 CRM 플러그인에서 메시징-native 운영 계층으로 확장Protecting against token theft — AI endpoint 보안이 인증에서 per-request 경제성 방어로 이동GitHub Copilot usage-based billing — AI 코딩 도구 경쟁이 모델 품질에서 token economics와 admin control로 이동Intel Xeon 6+ — agentic AI 인프라 병목이 GPU 단일 경쟁에서 orchestration CPU·memory·network 균형으로 이동Supabase Series F — vibe coding이 backend를 demo layer에서 agentic production substrate로 밀어올리다Palo Alto Frontier AI Defense — AI 보안이 모델 평가에서 machine-speed 대응 체계로 이동Redis Iris — agent stack이 prompt tuning에서 context engine 아키텍처로 이동SAP sustainability AI agents — enterprise AI가 챗봇에서 규제 워크플로 자동화로 이동Snowflake acquires Natoma — MCP가 실험적 연결 규약에서 enterprise governance layer로 이동Coralogix 200M Series F: AI agent observability가 독립 인프라 카테고리로 부상Postman AI Engineer: API 조직이 context debt를 관리하는 agentic engineering 계층Workday DevCon 2026: enterprise agent가 HR·Finance system of record로 진입하는 검증 스택 공개Build 2026: Microsoft가 Windows를 local agent runtime으로 전환Cisco Cloud Control: IT 운영이 dashboard에서 agentic control plane으로 이동Codex for every role, tool, and workflow — 코딩 에이전트가 팀 업무 플랫폼으로 확장미국 AI 행정명령: frontier model 정책이 보안 운영 체계로 구체화Introducing Command A+ — sovereign enterprise AI가 폐쇄형 API 의존에서 배포 가능한 open model stack으로 이동NVIDIA Alpamayo 2 Super — autonomous driving이 perception stack에서 reasoning-first physical AI stack으로 이동Salesforce acquires Contentful — enterprise AI가 CRM assistant에서 content orchestration layer 통합으로 이동AWS launches Amazon Quick desktop AI assistant that works across your applications, tools, and dataEnhanced AI Management and Analytics for OrganizationsIntroducing Trusted Remote Execution: Policy-Enforced Scripts for AI Agents and HumansTeamCity 2026.1: CLI, MCP for AI Agents, Pipelines Enhancements, and MoreAI coding startup Cognition raises $1B at $25B pre-money valuation — 코딩 에이전트 경쟁이 데모 품질에서 revenue proof와 orchestration economics로 이동AI Now Summit 2026 — 산업용 AI 경쟁이 범용 assistant에서 domain-specific engineering stack으로 이동Introducing Search Toolkit — agent retrieval 경쟁이 RAG 데모에서 검색 파이프라인 운영력으로 이동NVIDIA and IREN Announce Strategic Partnership to Accelerate Deployment of up to 5 Gigawatts of AI Infrastructure — AI infra 경쟁이 GPU 조달에서 전력·부지 결합형 factory rollout으로 이동Arm open-sources Metis — AI 보안 검증이 규칙 기반 스캐너에서 repo-context reasoning으로 이동Intel introduces SuperClaw — agent 인프라 경쟁이 cloud-only에서 hybrid on-device routing으로 이동Linux Foundation launches DNS-AID — agent discovery 경쟁이 중앙 레지스트리에서 DNS 기반 개방 표준으로 이동OpenAI launches Rosalind Biodefense — frontier model 배치가 범용 챗봇에서 방어형 life-science workflow로 확장CoreWeave Closes the Training-to-Inference Gap for Autonomous Agent Improvement — 에이전트 운영 경쟁이 모델 선택에서 closed-loop 학습 인프라로 이동Introducing Claude Opus 4.8 — 모델 경쟁이 지능 향상에서 장시간 agent workflow 신뢰성 경쟁으로 이동Warp is now open-source — agentic development 도구 경쟁이 폐쇄형 제품에서 공개형 orchestrated workflow로 이동Building the agentic future: Developer highlights from I/O 2026 — Google이 agent 개발 스택을 managed runtime으로 끌어올렸다EY and Microsoft launch a $1B enterprise AI initiative — enterprise AI 경쟁이 PoC에서 field engineering 운영으로 이동Vera Arrives: NVIDIA’s First CPU Built for Agents Lands at Top AI Labs — agent 인프라 경쟁이 GPU 단독에서 CPU 설계로 확장GitHub Copilot usage metrics update — 코드 리뷰 자동화의 '실사용'과 '자동 노출'을 분리 측정this is how an AI generated cow looked 12 years agoGitHub Copilot in VS Code 3월 릴리즈 — Autopilot과 integrated browser debugging으로 에이전트 실행 범위 확대GitHub Mobile 업데이트 — Copilot cloud agent를 PR 전 단계부터 모바일에서 운영Meta, AI 기반 Risk Review 고도화 — 규제 준수를 '사후 검토'에서 '항상 켜진 개발 단계 탐지'로 전환Meta, Muse Spark 공개 — Meta AI를 'social-context aware' personal superintelligence로 재정의GLM-5.1: Towards Long-Horizon TasksCursor, warp decode 공개 — Blackwell 기반 MoE 추론을 1.84x 가속하며 정확도도 개선GitHub Advanced Security, Dynatrace 런타임 컨텍스트 연동 — 배포된 취약점부터 우선순위화GitHub Copilot CLI, BYOK·로컬 모델 지원 — 코딩 에이전트가 SaaS에서 사내 런타임으로 확장OpenAI, 'Industrial Policy for the Intelligence Age' 발표 — AI 경제의 분배·세제·전력까지 정책 의제로 끌어올리다Show HN: I built a tiny LLM to demystify how language models workawesome-design-md — AI 에이전트를 위한 디자인 시스템 컬렉션System Card: Claude Mythos Preview [pdf]Compound Engineering — AI 네이티브 개발 철학, Plan→Work→Review→Compound 루프로 지식을 누적하는 방식Anthropic, Google·Broadcom과 차세대 TPU 수 GW 계약 — 컴퓨트 병목이 곧 전략 그 자체가 된 AI 산업GitHub, Dependabot 경고를 Copilot·Claude·Codex에 직접 할당 — 보안 패치가 에이전트 워크플로우로 편입Project Glasswing 공개 — Anthropic, Mythos Preview로 핵심 소프트웨어 공급망 방어 연합 출범HappyHorse-1.0Issue: Claude Code is unusable for complex engineering tasks with Feb updatesAnthropic RSP 3.1 업데이트 — Frontier Safety Roadmap를 실험 약속에서 운영 거버넌스로 세분화Cursor 3 공개 — AI 코딩 IDE가 단일 에이전트 채팅에서 멀티워크스페이스 운영 체계로 전환Hugging Face State of Open Source Spring 2026 — 오픈 모델 경쟁의 무게중심이 미국 중심에서 다극·주권형 생태계로 이동Qwen3.6-Plus: Towards real world agentsAirLLM — 4GB GPU에서 70B LLM 돌리는 초경량 추론 라이브러리Claw Code, Claude Code 소스 유출 계기로 등장한 오픈소스 AI 코딩 에이전트 — 출시 1주일 만에 GitHub 100K starsElgato Stream Deck 7.4, MCP 지원 추가 — AI 에이전트 프로토콜이 처음으로 소비자 하드웨어로 진입Google Gemma 4 공개 — Apache 2.0·256K 컨텍스트·멀티모달, 오픈 에이전틱 모델의 새 기준PrismML, Bonsai 1-bit LLM 출시 — 1GB 메모리로 8B 추론, 엣지 AI의 현실화Anthropic, 8만508명 인터뷰 공개 — AI 수요가 '더 강한 모델'보다 '더 나은 삶'에 가깝다는 데이터GitHub, Copilot cloud agent 조직 러너 제어 공개 — 에이전트 실행 환경을 저장소별 설정에서 조직 정책으로 승격JetBrains Central 공개 — Claude Agent·Codex·Gemini CLI를 묶는 에이전트 제어 평면How Microsoft Vaporized a Trillion DollarsDomo AI Agent Builder + MCP Server 공개 — 엔터프라이즈 BI가 멀티-LLM AI 에이전트 인프라로 전환하는 첫 사례Google TurboQuant — KV Cache 6배 압축·H100 어텐션 8배 가속, 정확도 손실 제로로 LLM 서빙 비용 구조 재정의JetBrains AI Pulse 서베이 — Claude Code, 시장 최고 로열티 지표(CSAT 91%·NPS 54)로 agentic coding 패러다임 전환 입증테네시주 SB 1580 서명 — AI의 정신건강 전문가 사칭 금지, 미국 AI 규제 초당적 확산 신호Agency Swarm — 조직형 멀티에이전트 오케스트레이션 프레임워크Agno — 프레임워크·런타임·컨트롤 플레인을 묶은 에이전트 스택AutoGen — Microsoft의 멀티에이전트 프로그래밍 프레임워크AutoGPT — 지속 실행형 AI 에이전트 플랫폼browser-use — 웹사이트를 AI 에이전트용 인터페이스로 바꾸는 브라우저 자동화CrewAI — 역할 기반 멀티에이전트 협업 프레임워크Dify — 워크플로·RAG·에이전트를 묶은 프로덕션 플랫폼LangGraph — 상태를 가진 에이전트를 그래프로 설계하는 프레임워크Model Context Protocol Servers — MCP 레퍼런스 서버 모음Semantic Kernel — 엔터프라이즈 지향 에이전트 오케스트레이션 SDKCLI-Anything — 기존 소프트웨어를 에이전트용 CLI로 바꾸는 프레임워크Karpathy LLM Wiki — RAG 대신 누적형 지식 위키 패턴AutoAgent — 벤치마크로 에이전트 하네스를 스스로 개선하는 메타 에이전트AutoClaw — OpenClaw 원클릭 온보딩/제품화 레이어NanoClaw — 경량·격리형 OpenClaw 대안Open WebUI — 셀프호스트 AI 웹 인터페이스ZooClaw — OpenClaw 기반 hosted 멀티 에이전트 팀Chrome Zero-Day CVE-2026-5281 긴급 패치 — Dawn WebGPU UAF 취약점 실사용 공격 확인, 2026년 네 번째Meta MTIA 칩 4세대 로드맵 공개 — 6개월 주기 출시, GenAI 추론 전담 아키텍처로 Nvidia 의존 분산Microsoft, 일본에 $100억 AI 인프라 투자 — SoftBank·Sakura Internet과 협력, 데이터 주권 전면 보장Sarvam AI, $300M 펀딩 완료 / $1.5B 밸류에이션 — 인도 주권 AI 유니콘 탄생, Bessemer·Nvidia·Amazon 참여GitHub Copilot in Visual Studio 업데이트 — custom agents·agent skills·MCP 거버넌스 도입Google Gemini 3.1 Flash Live 공개 — 실시간 음성 에이전트용 오디오 모델, ComplexFuncBench Audio 90.8%Meta BOxCrete 공개 — 데이터센터 콘크리트 배합을 AI로 최적화, 강도 도달 43% 단축Qodo, $70M Series B 유치 — AI 코딩 시대의 병목이 생성에서 검증으로 이동OpenAI closes funding round at an $852B valuationGoogle Gemini API, Flex & Priority 인퍼런스 티어 도입 — 비용-신뢰성 트레이드오프를 개발자가 제어Meta KernelEvolve 공개 — AI 에이전트가 GPU 커널 최적화, 수주 작업을 수 시간으로Anthropic, Claude 구독 제3자 도구 지원 중단 — OpenClaw 포함 외부 에이전트 하네스 차단Google Veo 3.1 Lite 출시 — AI 비디오 생성 비용 50% 절감, 개발자용 고용량 API 제공Meta, MTIA 4세대 AI 칩 6개월 주기 로드맵 공개 — GenAI 인퍼런스 전용 실리콘 전략Flowith Canvas / FlowithOS — 캔버스형 AI 워크스페이스Anthropic, Claude 내 171개 '기능적 감정' 벡터 발견 — 행동 인과관계 최초 규명NVIDIA Blackwell Ultra, MLPerf Inference v6.0 신기록 — 288 GPU로 DeepSeek-R1 초당 249만 토큰 처리Pinterest, 도메인별 MCP 에코시스템 프로덕션 배포 — 중앙 레지스트리·인간 승인으로 월 수천 시간 절감Anthropic Institute 출범 — frontier lab 내부 데이터를 정책·경제 연구 인프라로 전환Arcee Trinity-Large-Thinking 출시 — 미국계 오픈 에이전트 모델이 가격 대비 frontier 경쟁력 제시Gemma 4 공개 — Apache 2.0 오픈 모델을 agentic workflow 중심으로 재정의GitHub Copilot SDK 공개 프리뷰 — agent runtime이 제품 기능에서 플랫폼 계층으로 확장Claude Code Unpacked : A visual guideChrome 제로데이 CVE-2026-5281 — WebGPU use-after-free 실제 악용, CISA 긴급 패치 요구OpenAI, Codex Pay-As-You-Go 좌석 도입 + ChatGPT Business $20로 가격 인하OpenAI, 테크 토크쇼 TBPN 인수 — AI 기업 최초 미디어 직접 소유Anthropic-호주 MOU 체결 — AI Safety Institute와 정식 안전 평가 협력GitHub Copilot CLI /fleet 공개 — 병렬 서브에이전트로 코드 작업 동시 실행Microsoft, MAI 모델 3종 출시 — Foundry를 독자 멀티모달 모델 유통 채널로 본격 전환Mistral, $8.3억 부채 조달 — 유럽 독자 AI 컴퓨트 확보에 본격 베팅NVIDIA Mission Control 3.0 공개 — AI 팩토리 운영 KPI를 ‘GPU 활용률’에서 ‘token per watt’로 전환Alibaba Qwen3.6-Plus 공개 — 1M 컨텍스트·에이전트 코딩, Claude Opus 4.5 수준 달성Cisco, RSA 2026서 에이전트 AI 보안 프레임워크 DefenseClaw 공개 — Zero Trust를 AI 에이전트로 확장Google Gemini 3.1 Flash-Lite 출시 — Pro 대비 1/8 가격에 Gemini 2.5 Flash 동등 성능vLLM Model Runner V2 출시 — Prefill-Decode 분리 스케줄링으로 오픈소스 LLM 추론 아키텍처 혁신The Claude Code Source Leak: fake tools, frustration regexes, undercover modeGoogle Gemini Code Assist, 개인 개발자 무료 전환 — Gemini 2.5 기반 일 6,000회 코딩 요청 제공PrismML Bonsai — 세계 최초 상용 가능 1-bit LLM, iPhone에서 44 tok/s 달성Q1 2026 글로벌 VC $3,000억 사상 최고치 — AI가 전체 81% 독식, 단 4개 딜이 전체의 65% 차지캘리포니아 Newsom, 미국 최초 주정부 AI 안전 행정명령 서명 — 주계약 AI 기업에 안전·프라이버시 가이드라인 의무화Claude Code 내부 작동 원리 — 에이전트 루프, 컨텍스트 조립, 도구 실행 구조 해설Claw Code — Claude Code 소스 기반 Python/Rust 클린룸 재구현 프로젝트 (130k★)NVIDIA Nemotron 3 Super — 120B MoE 오픈소스 에이전트 모델, SWE-Bench 60.5% 달성Linux Foundation, MCP 기부 및 AAIF 출범 — AI 에이전트 표준화의 중립 거버넌스 시대 개막Google TurboQuant — LLM KV 캐시 메모리 6배 압축, H100에서 8배 속도 향상OpenAI, $122B 펀딩 완료 — $852B 밸류에이션으로 IPO 전 최대 사모 투자 기록Claude Code's source code has been leaked via a map file in their NPM registryAnthropic Claude Code npm 패키징 오류로 51만 줄 소스코드 유출 — KAIROS 자율 데몬 모드·미공개 모델 코드네임 노출Google TurboQuant, LLM KV 캐시 메모리 6배 압축·H100 속도 8배 향상 달성GPT-5.4 출시 — 추론·코딩·에이전트 통합 모델, OSWorld-V 인간 기준선 AI 최초 초과OpenAI, 최초 오픈웨이트 모델 gpt-oss-120b 공개 — Apache 2.0, o4-mini 수준 추론 성능gstack — Garry Tan(YC 회장)이 만든 AI 소프트웨어 팩토리Paperclip — AI 에이전트 팀을 회사처럼 운영하는 오케스트레이션 플랫폼Google TurboQuant — KV 캐시를 3비트로 6배 압축, 재학습 없이 H100에서 8배 처리량GPT-5.4 출시 — 컴퓨터 사용 에이전트로 인간 기준선(OSWorld 72%) 돌파Mistral Voxtral TTS — 4B 오픈소스 음성 합성 모델, ElevenLabs 대비 7~9배 저렴NVIDIA Nemotron 3 Super — 120B Mamba-Transformer MoE 오픈 에이전트 모델, 이전 대비 5배 처리량Cursor, 유료 개발자 100만 명 돌파 — 병렬 서브에이전트 & BugBot으로 AI 코딩 '5가데일' 재정의Google TurboQuant — LLM KV 캐시 3.5비트 압쳙으로 메모리 6배 절감, 오픈소스 공개OpenAI, $1,200억 역대 최대 평더링 완료 — Amazon $500억 주도, 기업가치 $7,300억 돌파Anthropic Mythos 유출 — 코딩·사이버보안 SOTA, "역량의 단계적 도약" 확인MCP 9,700만 설치 돌파 — AI 에이전트 인프라 표준으로 안착, 그러나 보안 위협도 급부상