Oracle Gece Arastirma v3 — 2026-06-02

Curate eden: Mahsum Aktaş · Günlük otomatik AI sektör taraması

Oracle Gece Arastirma v3 — 2026-06-02

Otomatik derleme | v3 pipeline | 7246 ham kaynak | 4634 benzersiz item | 5/5 kaynak ailesi kapsandi

Gunun Ozeti

Bugunun ana ekseni AI altyapisinin "model duyurusu"ndan "fiziksel compute, agent guvenligi ve regule operasyon" seviyesine kaymasi. Anthropic'in gizli S-1 basvurusu AI sermaye piyasasi icin esik olusturdu: https://www.anthropic.com/news/confidential-draft-s1-sec. NVIDIA, Cosmos 3, RTX Spark, Vera CPU ve AI factory yazilimlariyla hem cloud hem cihaz-ustu agent katmanina basti: https://huggingface.co/blog/nvidia/cosmos-3-for-physical-ai, https://developer.nvidia.com/blog/nvidia-vera-cpu-sets-a-new-standard-for-agentic-workloads-in-ai-factories. Guvenlik tarafinda BadHost, agent gateway'leri ve prompt-injection/persistent-control paper'lari agent runtime'inin yeni risk yuzeyi oldugunu gosterdi: https://www.infoq.com/news/2026/06/badhost-ai-systems-vulnerability/, https://arxiv.org/abs/2605.31042.

Top 7

1. Anthropic gizli S-1 ile halka arz surecini baslatti; bu, AI laboratuvarlarinin private funding doneminden public-market disiplinine gecis sinyali: https://www.anthropic.com/news/confidential-draft-s1-sec
2. NVIDIA Cosmos 3 fiziksel AI icin omni-model olarak duyuruldu; robotik, otonom arac ve dunya modeli hattinda teklesme hizlaniyor: https://huggingface.co/blog/nvidia/cosmos-3-for-physical-ai
3. Florida, OpenAI ve Sam Altman'a siddet olaylari uzerinden dava acti; chatbot safety artik urun riski degil hukuki sorumluluk basligi: https://techcrunch.com/2026/06/01/florida-sues-openai-sam-altman-in-first-of-its-kind-lawsuit-over-violent-incidents/
4. NVIDIA RTX Spark ile Grace Blackwell superchip'i PC sinifina tasidi; local agent calistirma cloud'a alternatif olmaya basliyor: https://www.theregister.com/systems/2026/06/01/nvidia-recasts-gb10-superchip-in-bid-for-high-end-pc-market/5249068
5. BadHost zafiyeti AI agent, evaluator ve LLM gateway mimarilerinin HTTP Host header gibi klasik web zaaflarindan etkilenebilecegini gosterdi: https://www.infoq.com/news/2026/06/badhost-ai-systems-vulnerability/
6. Claude Code Dynamic Workflows haberi, buyuk yazilim islerinde agent koordinasyonunun tek-agent modelinden cok-agent workflow'a kaydigini isaret ediyor: https://www.infoq.com/news/2026/06/dynamic-workflows-claude-code/
7. OpenAI Stargate Michigan veri merkeziyle 1GW sinifinda AI altyapisi kuruyor; enerji, su ve yerel altyapi AI stratejisinin merkezine oturdu: https://openai.com/index/stargate-michigan-data-center

TRENDS

• Spike: AI Regulation ve NVIDIA ayni anda yukseliyor; regulation tarafinda dava, youth-safety ve EU AI Act geri sayimi varken NVIDIA tarafinda Cosmos 3, RTX Spark, Vera ve AI Cloud ayni gun akti: https://artificialintelligenceact.substack.com/p/the-eu-ai-act-newsletter-103-the, https://blogs.nvidia.com/blog/ai-cloud-ecosystem/
• Yukselis: Databricks ve data/infra anlatimi, agent operasyonlarinin veri katmani olmadan calismayacagi temasina baglaniyor; "grounded intent data" tartismasi toplulukta da gorundu: https://mastodon.social/@ppcland/116674307072052195
• Azalis: AI Agents hacmi hala yuksek ama trend penceresinde ivme dusuyor; kalite sinyali artik "agent var"dan "agent nasil govern edilir" sorusuna kaydi: https://aws.amazon.com/blogs/machine-learning/secure-ai-agents-with-policy-and-lambda-interceptors-in-amazon-bedrock-agentcore-gateway/

KAT-1 — LLM & Model Guncellemeleri

• NVIDIA Nemotron 3 Ultra acik ABD modeli olarak guclu benchmark iddiasiyla geldi; open-model rekabeti artik sadece Meta/Qwen hattinda degil: https://the-decoder.com/nvidias-nemotron-3-ultra-becomes-the-smartest-open-us-model-but-china-still-leads/
• Qwen3.6-35B-A3B-NVFP4 Hugging Face trendinde NVIDIA optimize varyanti olarak one cikti; quantized/open deployment pazari buyuyor: https://huggingface.co/nvidia/Qwen3.6-35B-A3B-NVFP4
• Google, I/O 2026 uretiminde Gemini kullanimini anlatti; model artik demo aracindan etkinlik-operasyon aracina kayiyor: https://blog.google/innovation-and-ai/technology/ai/io-2026-google-ai/

KAT-2 — Endustri & Sirketler

• Anthropic halka arz dosyasi ile AI degerlemelerinin kamu piyasasinda test edilecegi yeni faza girdi: https://www.wired.com/story/anthropic-files-s1-ipo-sec/
• Runway'in Londra genislemesi video AI tarafinda Avrupa talent ve enterprise satis savasini buyutuyor: https://www.cnbc.com/2026/06/01/nvidia-backed-runway-london-expansion.html
• DuckDuckGo "no-AI" arama deneyimini daha erisilebilir hale getirdi; anti-AI UX artik niş degil, urun konumlandirmasi: https://techcrunch.com/2026/06/01/duckduckgo-makes-its-no-ai-search-engine-easier-to-access-as-its-traffic-booms/

KAT-3 — Arastirma & Paper'lar

• SAGE, agentic LLM'lerde memory evolution icin novelty gate oneriyor; Oracle memory hijyeni icin izlenmeli: https://arxiv.org/abs/2605.30711
• ElasticMem, LLM agent hafizasini ogrenilebilir kaynak olarak ele aliyor; statik RAG yerine adaptif memory mimarisi gucleniyor: https://arxiv.org/abs/2605.30690
• Eywa, provenance-grounded long-term memory ile agent hafizasinda kaynak izlenebilirligini merkeze koyuyor: https://arxiv.org/abs/2605.30771
• LongTraceRL, search agent trajectory'lerinden long-context reasoning egitimi oneriyor; agent replay/trajectory pool icin dogrudan sinyal: https://arxiv.org/abs/2605.31584

KAT-4 — Araclar & Framework'ler

• Microsoft MarkItDown GitHub trendinde; dokumanlari Markdown'a ceviren arac agent ingestion hattinda pratik deger tasiyor: https://github.com/microsoft/markitdown
• Scrapling adaptif web scraping framework olarak one cikti; collector dayanıkliligi ve anti-brittle scraping icin izlenmeli: https://github.com/D4Vinci/Scrapling
• supermemory, AI memory API/engine olarak trendde; long-term memory pazari urunlesmeye devam ediyor: https://github.com/supermemoryai/supermemory
• GLIDE library, GenAI ve agentic system evaluation icin prediction-powered inference'i endustrilestirme iddiasinda: https://arxiv.org/abs/2605.31278

KAT-5 — Acik Kaynak

• OpenSkillEval, skill ekosistemini otomatik audit etmeyi hedefliyor; skill governance icin dogrudan kullanilabilir fikir: https://huggingface.co/papers/2605.23657
• COLLEAGUE.SKILL, expert knowledge distillation ile otomatik AI skill uretimi oneriyor; Oracle skill-creator hattina yakin sinyal: https://arxiv.org/abs/2605.31264
• VoxCPM2 tokenizer-free TTS modeli GitHub trendinde; multilingual speech generation ve voice cloning tarafinda acik agirlik hareketi suruyor: https://github.com/OpenBMB/VoxCPM
• Compound Engineering plugin, Codex/Claude/Cursor icin paylasilabilir engineering skill seti olarak trendde: https://github.com/EveryInc/compound-engineering-plugin

KAT-6 — AI Agent'lar

• Amazon Bedrock AgentCore gateway icin policy ve Lambda interceptor modeli yayimlandi; agent izinleri runtime'da kesilebilir hale geliyor: https://aws.amazon.com/blogs/machine-learning/secure-ai-agents-with-policy-and-lambda-interceptors-in-amazon-bedrock-agentcore-gateway/
• MAVEN tool-calling agent generalization icin yeni benchmark/iyilestirme hattina isaret ediyor: https://arxiv.org/abs/2605.30738
• Harness Updating Is Not Harness Benefit paper'i self-evolving agent'larda "harness degisti diye performans artti" varsayimini ayiriyor: https://arxiv.org/abs/2605.30621
• Stateful Online Monitoring, dagitik agent saldirilarini online izleme ile yakalamayi oneriyor: https://arxiv.org/abs/2605.31593

KAT-7 — Multimodal

• SANA-Streaming, real-time video-to-video editing icin hybrid diffusion transformer oneriyor; interaktif video uretimi latency sinirina yaklasiyor: https://arxiv.org/abs/2605.30409
• VLM3, vision-language modellerinin 3D ogrenme kabiliyetini one cikariyor; embodied reasoning icin temel sinyal: https://arxiv.org/abs/2605.30561
• UniAudio-Token, semantic speech tokenizer'i genel audio perception ile genisletiyor; speech agent altyapisi icin onemli: https://arxiv.org/abs/2605.31521
• ImmersiveTTS, ortam farkindalikli TTS icin multimodal diffusion transformer kullaniyor; sesli agent deneyimi daha baglamsal hale geliyor: https://arxiv.org/abs/2605.30965

KAT-8 — Robotik & Embodied AI

• NVIDIA Cosmos 3 ve Alpamayo, physical AI reasoning/action hattini robotaksi ve embodied agent'lara bagliyor: https://developer.nvidia.com/blog/develop-physical-ai-reasoning-world-and-action-models-with-nvidia-cosmos-3
• NVIDIA Unitree secimi humanoid platform tarafinda acik robot ekosistemini hizlandirabilir: https://www.cnbc.com/2026/06/01/nvidia-unitree-humanoid-robotics-system-researchers.html
• DriveMA, driving VLA modellerini dogrulanabilir meta-action'larla ele aliyor; safety-critical driving agent'lari icin izlenmeli: https://arxiv.org/abs/2605.31271
• EMBGuard, embodied agent planlama icin hazard-aware guardrail oneriyor; fiziksel dunya agent'larinda safety katmani sart: https://arxiv.org/abs/2605.30924

KAT-9 — Edge & Cihaz

• RTX Spark destekli Windows laptoplar, local AI agent calistirma isini PC ureticilerinin ana hikayesine soktu: https://www.cnbc.com/2026/05/31/nvidias-new-chip-to-power-fresh-line-of-windows-laptops-by-dell-hp.html
• Sixfab AI HAT+ Raspberry Pi 5 icin DEEPX DX-M1 accelerator entegre ediyor; edge AI maker ekosistemi ucuzluyor: https://web.brid.gy/r/https://www.cnx-software.com/2026/06/01/sixfab-ai-hat-for-raspberry-pi-5-integrates-deepx-dx-m1-ai-accelerator/
• China'nin invasive brain-computer chip onayi, AI/BCI kesismesinin policy ve medtech tarafinda hizlandigini gosteriyor: https://www.technologyreview.com/2026/06/01/1138133/china-world-first-brain-chip/

SECURITY

• BadHost, Starlette kaynakli auth bypass zafiyetinin AI agent, evaluator ve gateway servislerine siçrayabilecegini gosterdi: https://www.infoq.com/news/2026/06/badhost-ai-systems-vulnerability/
• Prompt Injection to Persistent Control paper'i agentic harness'larda backdoor kaliciligina odaklaniyor; "tek prompt" degil, uzun yasayan kontrol problemi: https://arxiv.org/abs/2605.31042
• Depth-Dependent Indirect Prompt Injection, ReAct tool-calling agent'larda injection derinligi ve turn budget hassasiyetini inceliyor: https://arxiv.org/abs/2605.30686
• Docker sandbox security yazisi, coding agent'lari izole calistirma tartismasini pratik guvenlik katmanina cekiyor: https://www.docker.com/blog/what-is-sandbox-security/

REGULATION

• EU AI Act Newsletter #103, Agustos geri sayimini vurguluyor; compliance takvimi artik operasyonel risk: https://artificialintelligenceact.substack.com/p/the-eu-ai-act-newsletter-103-the
• Singapur, 150.000 kamu gorevlisi icin AI agent registry planliyor; agent kimligi ve faaliyet takibi kamu standardi haline geliyor: http://www.bing.com/news/apiclick.aspx?ref=FexRss&aid=&tid=6a1dfa1de6584377b479065a1beea629&url=https%3a%2f%2fwww.straitstimes.com%2ftech%2fspore-to-create-a-registry-of-ai-agents-for-150000-public-officers-amid-ai-push%3fref%3dlatest-headlines&c=9209766996917815269&mkt=tr-tr
• Colorado youth safety protections haberi, AI safety'nin cocuk/ergen riski uzerinden eyalet seviyesinde sertlestigini gosteriyor: https://news.google.com/rss/articles/CBMiugFBVV95cUxQYkxUQnJGMUJ4aDZfbGtHX29SU25JbXpWeVRreG41SEcyUkltWEwzMFdtNWMxQkxOenpITEg1TGJRaVgzeTNGelNIZ2RfOHFMbENfemM1QnhmcV95M2VFS0xJMW1tbVpQTUNIZ0V0Q29jNnlxN2hrTDFhYl9oR0VGVG9LUElHeFBDeV9EN1QyZHlLVkVhUWIzWUR6UnczVE00U2dfMWpFa0phMVVrOWZZRklNT253NEVIZ3c?oc=5
• Environmental AI Regulation paper'i "green AI hakkı" ve reasoning maliyeti uzerinden yeni regule alan aciyor: https://arxiv.org/abs/2603.00068

AI-SCIENCE

• Kernel Foundry, multi-expert diagnosis-driven kernel optimizer ile GPU kernel uretimini otomatiklestirmeye calisiyor: https://arxiv.org/abs/2605.30359
• GPU Forecasters, language modellerini kernel runtime optimizasyonunda selective surrogate olarak kullaniyor: https://arxiv.org/abs/2605.31464
• AutoSci, bilimsel arastirma yasam dongusu icin memory-centric agentic system oneriyor: https://arxiv.org/abs/2605.31468
• Extending AI for Research to the Humanities, evidence-grounded multi-agent framework ile humanitelerde arastirma ajanlarini inceliyor: https://arxiv.org/abs/2605.30947

INFRA

• NVIDIA AI Cloud ecosystem genislemesi, AI factory kapasitesinin ulusal/kurumsal rekabete donustugunu gosteriyor: https://blogs.nvidia.com/blog/ai-cloud-ecosystem/
• NVIDIA DSX OS, AI factory operasyonu icin acik/moduler yazilim katmani sunuyor: https://developer.nvidia.com/blog/nvidia-dsx-os-delivers-open-modular-software-for-operating-ai-factories-at-scale/
• NVIDIA DOCA In-Silicon Security, agentic AI altyapisinda guvenligi network/silicon seviyesine tasiyor: https://developer.nvidia.com/blog/advancing-ai-infrastructure-for-agentic-ai-with-nvidia-doca-in-silicon-security/
• Cloudflare core unit boot time optimizasyonu, hyperscale altyapida firmware/UEFI/iPXE darboğazlarinin AI kapasite operasyonunu etkileyebilecegini hatirlatiyor: https://blog.cloudflare.com/optimizing-core-unit-boot-time/

SAFETY

• Configurable Reward Model for Balanced Safety Alignment, farkli safety gereksinimlerine ayarlanabilir reward modeli oneriyor: https://arxiv.org/abs/2605.30487
• ConsisGuard, safety deliberation ile policy enforcement arasindaki tutarliligi guardrail seviyesinde hizalamayi hedefliyor: https://arxiv.org/abs/2605.31073
• Reinforcement Learning Amplifies Emergent Misalignment paper'i zararsiz reward'lardan bile misalignment buyuyebilecegini gosteriyor: https://arxiv.org/abs/2605.31328
• LLM Judges Inconsistently Disagree paper'i, safety kriterlerinde judge modellerin tutarsizligini olcuyor; evaluator governance sart: https://arxiv.org/abs/2605.31381

WATCHLIST

• Anthropic IPO sureci: valuation, governance disclosure ve model-risk anlatimi takip edilmeli: https://www.anthropic.com/news/confidential-draft-s1-sec
• OpenAI Florida davasi: chatbot harm/safety sorumlulugu icin emsal olma potansiyeli var: https://www.cnbc.com/2026/06/01/florida-ag-open-ai-altman-lawsuit.html
• NVIDIA local-agent PC hatti: RTX Spark gercek geliştirici adoption'i ve battery/thermal performansi izlenmeli: https://webrazzi.com/2026/06/01/nvidia-dan-windows-bilgisayarlari-destekleyecek-yapay-zeka-super-cipi-rtx-spark/
• Agent registry/compliance: Singapur modeli diger kamu kurumlarina kopyalanabilir: http://www.bing.com/news/apiclick.aspx?ref=FexRss&aid=&tid=6a1dfa1de6584377b479065a1beea629&url=https%3a%2f%2fwww.straitstimes.com%2ftech%2fspore-to-create-a-registry-of-ai-agents-for-150000-public-officers-amid-ai-push%3fref%3dlatest-headlines&c=9209766996917815269&mkt=tr-tr

CikCik Paketi

• Simon Willison, Hacker News'te AI slop toleransinin arttigini ve linkli yazilarda slop'un yorumlardan daha yaygin oldugunu yazdi; CikCik icin "slop detection" thread'i degerli: https://bsky.app/profile/simonwillison.net/post/3mna77rupi22d
• Willison'in "inhuman consistency" heuristigi, AI-generated content tespiti icin kisa ve guclu bir post acisi veriyor: https://bsky.app/profile/simonwillison.net/post/3mnadcj4i422d
• Harrison Chase'in "Build -> Test -> Deploy -> Monitor" agent development notu, agent operasyonlarini DevOps dongusune oturtan net bir mini-framework: https://news.google.com/rss/articles/CBMiX0FVX3lxTE84NkdHN3NSVUVxc1g4dm1WZGlTN1BQSVBzQk5XdHd3SzhPXzNwaVBnYmZ2SHBJcS0wRk1lRGlMdnZja09VV0dlZ3Q1Y3lBTFBYT3FZWkVNR1ZfVDJoTEhB?oc=5
• Omar Sanseviero'nun Gemma skills sinyali, model-spesifik skill ekosisteminin basladigini gosteriyor: https://news.google.com/rss/articles/CBMiYkFVX3lxTE9RZTVLeVRMVF9IeTNZN2ZhZ2NzZGN3ZWlOTWRHeDFadHBpdWxNMmI1T3FiWjhoSlQ1OGNpX1QzSEZRVzhlaVdWVWoxQTNHdTNnN1hkU3ZTQy1fTXc4cFJOTDF3?oc=5
• Nathan Lambert, NVIDIA'i ABD open model buyumesinin spearhead'i olarak konumlandirdi; Nemotron/Cosmos anlatimina sosyal kanit: https://news.google.com/rss/articles/CBMiYkFVX3lxTE1QSTZNOGx5WFVXdFg0amJvdXhiTmFfbzllX0dkbktzbEtWRGJIb3RUcmZwSHVZY3NNajFPdFdPOC0zNTk0VlZwa0MySXFrYjZoWng5bFJsem1lRnIwbDR3OFh3?oc=5
• HN Launch: Expanse GPU kapasite israfini unlock etmeyi hedefliyor; infra ve GPU marketplace takibi icin izlenmeli: https://news.ycombinator.com/item?id=48356312

Oracle Self-Improvement Sinyalleri

• Memory: SAGE, ElasticMem ve Eywa ayni gece geldi; Oracle memory pipeline'i novelty gate + provenance + learnable memory uclusuyle yeniden degerlendirilmeli: https://arxiv.org/abs/2605.30711
• Agent guvenligi: BadHost, prompt injection depth ve persistent-control paper'lari agent gateway'lerinde host/header, tool-depth ve long-lived state audit'i gerektiriyor: https://www.infoq.com/news/2026/06/badhost-ai-systems-vulnerability/
• Skill sistemi: OpenSkillEval ve COLLEAGUE.SKILL, mevcut skill catalog icin otomatik kalite auditi ve expert-distillation yolu oneriyor: https://huggingface.co/papers/2605.23657
• Evaluation: GLIDE ve LLM Judges inconsistency paper'i, reviewer/verifier agent'larin tek judge'a yaslanmamasi gerektigini guclendiriyor: https://arxiv.org/abs/2605.31278

Kaynak Ozeti

• RSS/news ailesi 1512 ham item verdi; NVIDIA, TechCrunch, CNBC ve The Register baskindi: https://blogs.nvidia.com/blog/ai-cloud-ecosystem/
• Search ailesi 510 ham item verdi; Google News aggregator agirligi yuksek, canonical URL cozumleme hala iyilestirme alani: https://news.google.com/rss/articles/CBMif0FVX3lxTE4xSjA1OERrenh2U2ZHdDRlZUZxRkxDVXFQYkpaODU2WmxDaTB0bklTdW1CTk5uc3dIUXVJQVoyVm93cndTMGVFNElRaVdVbzlTLU12NU4xTWdtWFExR3FjTG9xU0JFOU1JRnFCNVE4Wm5uSllzUXV6d0JsQXFLbDA?oc=5
• Academic/API ailesi 1098 ham item verdi; arXiv cs.LG, cs.CL ve cs.CV agirligi cok yuksek: https://arxiv.org/abs/2605.30711
• Community ailesi 41 ham item verdi; GitHub trending ve HN sinyalleri pratik arac/infra kesfinde daha kaliteli: https://github.com/microsoft/markitdown
• Social ailesi 4085 ham item verdi; Bluesky/Mastodon net sinyal uretirken X fallback eski ve gurultulu veri tasiyor: https://bsky.app/profile/simonwillison.net/post/3mna77rupi22d

Coverage / Blind Spots

• Coverage tam: rss/news, search, community, social ve academic/api ailelerinin tamami calisti; kaynak ailesi eksigi yok: https://blogs.nvidia.com/blog/ai-cloud-ecosystem/
• Thin spot yok ama social ailesinde 4085 item icinde X fallback stale/aggregator linkleri fazla; CikCik icin Bluesky/Mastodon sinyali daha guvenilir: https://bsky.app/profile/simonwillison.net/post/3mnab3bqa7k2d
• Search tarafinda Google News ve Bing redirect URL'leri canonical kaliteyi dusuruyor; publish oncesi canonical resolver faydali olacak: https://news.google.com/rss/articles/CBMiWkFVX3lxTE5IUFN4T3VUZFhlSnZ4eVVWaEJqVE1mUFNJaC1vU01jcmU5VFplNE81WkhKYnRKclUyV0Vtb2J5Ykh6LXVJcHJySUZ0UWQteC1OSkg4NmJEeWxjQQ?oc=5

Bu Gece Sistem Ne Ogrendi

• Rising entity seti Meta/AI Regulation/Google/AI Agents'tan bugun NVIDIA/Anthropic/regulation eksenine kaydi; AI anlatimi sermaye + altyapi + governance uclusune sıkisti: https://www.anthropic.com/news/confidential-draft-s1-sec
• Tekil model haberleri yerine agent runtime, memory, evaluator ve security paper'lari ayni gece kümelendi; self-improvement icin "agent lifecycle observability" one alindi: https://arxiv.org/abs/2605.31593
• Collector kalitesi acisindan social hacim buyuk ama sinyal seyrek; sonraki run'da stale X fallback icin tarih filtresi sertlestirilmeli: https://mastodon.social/@h4ckernews/116675897732022888

Dedupe & Kalite Notu

Bu rapordaki maddeler onceki 3 gun raporlarina gore tekrar kontrolunden gecirildi. Pipeline: 6208 merge cikisi -> 4634 yeni item | 1565 fingerprint tekrar | 9 semantik tekrar | 0 materyal guncelleme. Not: Codex CLI rapor modu bos yanit dondurdugu icin rapor, v3 prompt + unique.json fallback senteziyle olusturuldu; publish/feed adimlari standart v3 scriptleriyle devam ettirilecek.