Oracle Gece Arastirma - 2026-07-02

Curate eden: Mahsum Aktaş · Günlük otomatik AI sektör taraması

Oracle Gece Arastirma - 2026-07-02

Otomatik derleme | v3 pipeline | 6908 ham item | 1895 benzersiz item | 105 benzersiz kaynak | 5/5 kaynak ailesi kapsandi

Gunun Ozeti

Bugunun ana ekseni agentic web ekonomisi, agent guvenligi ve production AI altyapisiydi. Cloudflare, AI crawler trafiğini ucretlendirme ve yonetme hattini x402 tabanli Monetization Gateway ile sertlestirdi; bu, yayinlar ve AI sirketleri arasindaki "crawl et ama odeme" dengesini bozuyor. Kaynak: https://blog.cloudflare.com/monetization-gateway/ ve https://techcrunch.com/2026/07/01/cloudflares-new-policy-pushes-ai-companies-to-pay-for-publishers-content/

Ikinci hat coding-agent urunlesmesi: GitHub Copilot ayni gunde Kimi K2.7 Code, Copilot Vision, browser tools ve AI credit session limitlerini duyurdu; agent IDE'leri artik model secimi, gorme, browser ve harcama limitiyle tam calisma ortami oluyor. Kaynak: https://github.blog/changelog/2026-07-01-kimi-k2-7-is-now-available-in-github-copilot ve https://github.blog/changelog/2026-07-01-copilot-vision-is-generally-available ve https://github.blog/changelog/2026-07-01-set-ai-credit-session-limits-in-copilot-cli-and-sdk

Ucuncu hat guvenlik ve governance: Docker agent izolasyonunu, Godot AI-authored contribution yasağini, arXiv ise prompt injection, semantic cache collision, agent protocol governance ve confidential AI paper'larini one cikardi. Kaynak: https://www.docker.com/blog/why-ai-agents-need-isolation/ ve https://www.pcgamer.com/gaming-industry/open-source-game-engine-godot-will-no-longer-accept-ai-authored-code-contributions-we-cant-trust-heavy-users-of-ai-to-understand-their-code-enough-to-fix-it/ ve https://arxiv.org/abs/2606.30783

Trend Analizi

• SPIKE: AI Regulation ve Sonnet gorunurlugu yuksek; Sonnet tarafini tekrar haber yapmadim, cunku onceki raporlarda ana konu oldu. Regulasyon tarafinda bugunku yeni fark Cloudflare'in crawler ekonomisi ve state-level AI regulation tartismalari. Kaynak: https://blog.cloudflare.com/content-independence-day-ai-options/ ve https://mastodon.social/@byteseu/116846498044230577
• YUKSELIS: AI Agents, AI Safety, Amazon, Anthropic ve Autonomous etiketleri yukseliste; bugunku temiz sinyal AWS A2A gateway, AgentCore Memory ve Docker izolasyon bolumlerinde. Kaynak: https://aws.amazon.com/blogs/machine-learning/building-a-serverless-a2a-gateway-for-agent-discovery-routing-and-access-control/ ve https://aws.amazon.com/blogs/machine-learning/structured-memory-filtering-with-metadata-in-agentcore-memory/ ve https://www.docker.com/blog/why-ai-agents-need-isolation/
• AZALIS: DALL-E dususte; bugunku multimodal akisi daha cok voice, vision, video-world-model ve Copilot Vision tarafina kaydi. Kaynak: https://huggingface.co/blog/cerebras-gemma4-voice-ai ve https://github.blog/changelog/2026-07-01-copilot-vision-is-generally-available ve https://huggingface.co/papers/2606.31734

Top 7

1. Cloudflare, x402 tabanli Monetization Gateway ile web sayfasi, dataset, API ve MCP tool gibi kaynaklarin agentic web uzerinden ucretlendirilmesini aciyor; bu, AI crawler ekonomisinde dogrudan odeme katmani. Kaynak: https://blog.cloudflare.com/monetization-gateway/
2. Cloudflare, AI traffic options ile site sahiplerine search crawler, AI crawler ve training crawler ayrimi yapma imkani veriyor; robots.txt pasif sinyal olmaktan cikip policy surface haline geliyor. Kaynak: https://blog.cloudflare.com/content-independence-day-ai-options/
3. GitHub Copilot, Kimi K2.7 Code'u model picker'a ekledi; acik agirlikli bir coding modeli Copilot icinde resmi secenek oluyor. Kaynak: https://github.blog/changelog/2026-07-01-kimi-k2-7-is-now-available-in-github-copilot
4. Copilot Vision GA oldu; kod yardimcisi artik screenshot/PDF gibi gorsel baglami da dogrudan reasoning input'u yapabiliyor. Kaynak: https://github.blog/changelog/2026-07-01-copilot-vision-is-generally-available
5. AWS, serverless A2A gateway rehberiyle agent discovery, routing ve access-control'u production mimarisi olarak paketliyor. Kaynak: https://aws.amazon.com/blogs/machine-learning/building-a-serverless-a2a-gateway-for-agent-discovery-routing-and-access-control/
6. Docker, AI agent izolasyonunu ayri bir guvenlik meselesi olarak konumladi; agent calisma ortami icin sandbox ve ayrik runtime ihtiyaci netlesiyor. Kaynak: https://www.docker.com/blog/why-ai-agents-need-isolation/
7. Godot, AI-authored/vibe-coded contribution'lari kabul etmeyecegini duyurdu; open-source projelerde AI katkisi icin kalite, sahiplik ve bakim sorumlulugu tartismasi sertlesiyor. Kaynak: https://www.pcgamer.com/gaming-industry/open-source-game-engine-godot-will-no-longer-accept-ai-authored-code-contributions-we-cant-trust-heavy-users-of-ai-to-understand-their-code-enough-to-fix-it/

KAT-1 - LLM & Model Guncellemeleri

• Hugging Face ve Cerebras, Gemma 4'u real-time voice AI icin one cikardi; hizli inference + voice agent kombinasyonu kucuk/orta model sinifinda yeni rekabet alani. Kaynak: https://huggingface.co/blog/cerebras-gemma4-voice-ai
• OpenAI, ChatGPT adoption verilerini yayimladi; model haberi degil ama kullanici davranisi, bolge ve capability adoption haritasi icin degerli sinyal. Kaynak: https://openai.com/index/how-chatgpt-adoption-has-expanded
• BlockPilot, diffusion-based speculative decoding icin instance-adaptive policy oneriyor; inference hizlandirma artik sadece statik draft model degil, ornek bazli policy sorunu. Kaynak: https://arxiv.org/abs/2606.31315
• SeKV, long-context LLM inference icin resolution-adaptive KV cache ve semantic memory yaklasimi getiriyor; KV cache maliyeti ana darboğaz olmaya devam ediyor. Kaynak: https://arxiv.org/abs/2606.31145

KAT-2 - Arastirma & Paper'lar

• QVal, long-horizon LLM agents icin ucuz dense supervision sinyali uretmeyi hedefliyor; outcome-only reward'un yetersiz kaldigi agent egitimi icin pratik bir yol. Kaynak: https://arxiv.org/abs/2606.32034
• FinPersona-Bench, autonomous financial agents icin longitudinal psychometric stability olcuyor; finansal agent'larda karakter/mandate drift'i risk haline geliyor. Kaynak: https://arxiv.org/abs/2606.31522
• AutoTrainess, language model'lara language model iyilestirmeyi ogretme hattini arastiriyor; self-improvement trendi benchmark ve kontrol ihtiyaci doguruyor. Kaynak: https://arxiv.org/abs/2606.31551
• Moral Safety in LLMs, "performative compliance" sorununu puzzled cues ile test ediyor; modellerin etik davranisi gercek mi yoksa sinyal taklidi mi sorusu gucleniyor. Kaynak: https://arxiv.org/abs/2606.31644

KAT-3 - Araclar & Framework'ler

• GitHub Copilot CLI/SDK icin AI credit session limits geldi; agent harcamasini task/session bazinda sinirlama production ekipleri icin kritik kontrol. Kaynak: https://github.blog/changelog/2026-07-01-set-ai-credit-session-limits-in-copilot-cli-and-sdk
• Browser tools for GitHub Copilot in VS Code GA oldu; agent'in browser ile calismasi artik deneysel eklenti degil resmi tool surface. Kaynak: https://github.blog/changelog/2026-07-01-browser-tools-for-github-copilot-in-vs-code-are-generally-available
• Copilot CLI auto model selection, task'e gore model routing yapiyor; agent urunlerinde manuel model secimi yavas yavas policy/optimizer katmanina kayiyor. Kaynak: https://github.blog/changelog/2026-07-01-copilot-cli-auto-model-selection-routes-based-on-task
• New C++ language server config skill, Copilot CLI icin setup otomasyonunu genisletiyor; coding-agent skill ekosistemi IDE config'lerine iniyor. Kaynak: https://github.blog/changelog/2026-07-01-new-c-language-server-config-skill-for-copilot-cli

KAT-4 - Acik Kaynak

• Strix, open-source AI pentesting agent olarak GitHub trending'e girdi; security agent'lar sadece scanning degil, exploit/fix dongusu iddiasi tasiyor. Kaynak: https://github.com/usestrix/strix
• google/agents-cli, Google Cloud uzerinde agent yaratma, degerlendirme ve deploy etme skill seti sunuyor; cloud provider'lar agent toolchain'i CLI seviyesine cekiyor. Kaynak: https://github.com/google/agents-cli
• Astryx, agent-ready design system olarak one cikti; UI sistemleri insan developer kadar coding agent'in anlayacagi yapisal kurallara da odaklanmaya basliyor. Kaynak: https://github.com/facebook/astryx
• Herdr, terminal icinde agent multiplexer yaklasimi sunuyor; tek agent yerine coklu agent kontrolu developer workflow'una dogrudan giriyor. Kaynak: https://github.com/ogulcancelik/herdr

KAT-5 - Endustri & Sirketler

• Cloudflare'in crawler ucretlendirme hamlesi, publisher'larin AI sirketlerinden dogrudan deger talep etme cizgisini guclendiriyor. Kaynak: https://techcrunch.com/2026/07/01/cloudflares-new-policy-pushes-ai-companies-to-pay-for-publishers-content/
• Palantir CEO'su Alex Karp, OpenAI/Anthropic token ekonomisini elestirdi; enterprise musteri tarafinda "token bazli fiyatlama" sorgulanmaya basliyor. Kaynak: https://www.cnbc.com/2026/07/01/palantir-karp-open-ai-anthropic-tokens.html
• Meta'nin AI compute satmaya yonelik cloud is plani haberleri, hyperscaler olmayan buyuk model sirketlerinin excess capacity'yi pazara acabilecegini gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMiqwFBVV95cUxQYmw4T0JhR2IzMjFUR25WY041VktWR1ZKSWZ6V3FPUmszZUU4WGpVSjR6SDNNT2hINWlnbVdfYThRWUtwalF5Z0ZvamdmSTVVNVRNVy1iYVQ2OFFfS0ZQVFRmeVNUWEQ0YVo5VTBxa08xZWZZYjVISERPelZNaFROdkNlem1VcFRlVVJhM1Nodms5VVEzTDZfaUEyN2lUU2tCRW9tdVZ3RU5oZlE?oc=5
• Micron CEO'su memory shortage ve fiyat baskisini musteri pazarligi ile iliskilendirdi; AI altyapisi DRAM tarafinda son kullanici cihaz maliyetine kadar iniyor. Kaynak: https://www.cnbc.com/2026/06/30/micron-ceo-customers-driving-hard-bargain-on-price-contributed-to-memory-shortage.html

KAT-6 - AI Agent'lar

• AWS A2A gateway, path-based routing ile coklu agent'i tek domain arkasinda yayinlama modeli sunuyor; agent discovery ve access control artik mimari primitive. Kaynak: https://aws.amazon.com/blogs/machine-learning/building-a-serverless-a2a-gateway-for-agent-discovery-routing-and-access-control/
• AgentCore Memory metadata filtering, multi-agent ve multi-tenant bellek sistemlerinde retrieval kontrolu icin pratik pattern veriyor. Kaynak: https://aws.amazon.com/blogs/machine-learning/structured-memory-filtering-with-metadata-in-agentcore-memory/
• LUMOS, accessibility-grounded semantic OS layer oneriyor; GUI agent'larin pixel yerine anlamli UI state'e ihtiyaci var. Kaynak: https://arxiv.org/abs/2606.30697
• What Memory Do GUI Agents Really Need? calismasi passive record'dan active task-driving state'e gecisi savunuyor; agent memory sadece log degil, karar durumu. Kaynak: https://arxiv.org/abs/2606.31612

KAT-7 - Multimodal

• Copilot Vision GA, developer workflow'unda screenshot/PDF gibi gorsel kanitlari prompt'a bagliyor; debug ve UI review akislari genisler. Kaynak: https://github.blog/changelog/2026-07-01-copilot-vision-is-generally-available
• AVTok, audio-video generation icin 1D unified tokenization oneriyor; senkron ses-goruntu uretiminde ortak token uzayi arayisi guclu. Kaynak: https://arxiv.org/abs/2606.30811
• MemLearner, video world model'larinda context memory sorgulamayi ogretiyor; uzun video state'i icin secici hafiza gerekecek. Kaynak: https://arxiv.org/abs/2606.31734
• RedVox, speech modellerinde dil bazli safety ve fairness gap'lerini ortaya koyuyor; voice agent'lar icin English disi guvenlik hala zayif alan. Kaynak: https://huggingface.co/papers/2606.26968

KAT-8 - Robotik & Embodied AI

• Prompting Robot Teams with Natural Language, robot ekiplerini dogal dille yonlendirme hattini arastiriyor; multi-robot orchestration icin LLM interface trendi suruyor. Kaynak: https://arxiv.org/abs/2509.24575
• Robustness of Robotic Manipulation: Foundations and Frontiers, manipülasyon dayanıklılıgını temel sorun olarak topluyor; embodied AI'da policy kalitesi kadar perturbation toleransi onemli. Kaynak: https://arxiv.org/abs/2606.31494
• ELASTIC, generative control policies icin test-time compute'u adaptif olceklendirmeyi hedefliyor; robotik policy'lerde her input'a ayni hesap ayirmak verimsiz. Kaynak: https://arxiv.org/abs/2606.31132
• Wavelet Policy, world-prior memory ile scale-domain imitation learning oneriyor; robot hareketinde temporal scale ayrimi daha fazla calisilacak. Kaynak: https://arxiv.org/abs/2504.04991

KAT-9 - Edge & Cihaz

• Gemma 4 + Cerebras real-time voice AI, edge'e yakin dusuk latency voice assistant hattini destekliyor; "konusan agent" icin inference gecikmesi urun siniri. Kaynak: https://huggingface.co/blog/cerebras-gemma4-voice-ai
• MAX modellerinin Apple silicon GPU uzerinde calismasi toplulukta yankilandi; local inference ve Apple cihazlari tekrar ciddi gelistirme hedefi. Kaynak: https://forum.modular.com/t/max-models-can-now-run-on-apple-silicon-gpus/3283
• FluidVoice, macOS icin on-device STT ve AI enhancement iddiasiyla trending'e girdi; local dictation pazari Wispr benzeri urunlere acik. Kaynak: https://github.com/altic-dev/FluidVoice
• Little Brains, Big Feats calismasi compact language model'lari inceliyor; frontier disi kucuk model optimizasyonu edge ve maliyet acisindan kritik. Kaynak: https://huggingface.co/papers/2606.30062

SECURITY

• Docker, AI agent isolation ihtiyacini ayri bir urun/guvenlik problemi olarak anlatti; agent runtime'lari dosya sistemi, network ve secret sinirlarini default kapatmali. Kaynak: https://www.docker.com/blog/why-ai-agents-need-isolation/
• Security-Fidelity Tradeoffs paper'i, prompt injection savunmalarinin fayda kadar capability kaybi getirdigini gosteriyor; savunma sadece "daha cok filtre" olamaz. Kaynak: https://arxiv.org/abs/2606.30783
• Semantic cache key collision attack, LLM semantic caching katmaninda benzerlik tabanli cache'in guvenlik yuzeyi olabilecegini gosteriyor. Kaynak: https://arxiv.org/abs/2601.23088
• EnclaveX, CPU/GPU TEE ile confidential AI dagitimini hedefliyor; model ve veri gizliligi inference katmaninda daha fazla talep gorecek. Kaynak: https://arxiv.org/abs/2606.31408

REGULATION

• Cloudflare, AI traffic options ile site sahiplerine crawler policy kontrolu veriyor; regule edilmeden once piyasa kendi teknik sozlesmesini yazmaya basladi. Kaynak: https://blog.cloudflare.com/content-independence-day-ai-options/
• Incentive Aware AI Regulations, regülasyonu incentive yapilariyla modellemeye calisiyor; kural koymak kadar aktorlerin kacis davranisini tahmin etmek onemli. Kaynak: https://arxiv.org/abs/2603.05175
• RCTs for Frontier AI Governance, human uplift studies icin randomize deney tasarimi sorunlarini tartisiyor; frontier AI policy'si deneysel kanit istiyor. Kaynak: https://arxiv.org/abs/2603.11001
• Probing Stylistic Appropriation under EU Law, LLM copyright ihlalini stil sahiplenmesi uzerinden teknik olarak test etmeyi oneriyor. Kaynak: https://arxiv.org/abs/2606.31250

AI-SCIENCE

• AWS BoltzGen on SageMaker AI, protein design deneyini cloud workflow'a tasiyor; AI-science urunlesmesi modelden cok repeatable lab pipeline'a bagli. Kaynak: https://aws.amazon.com/blogs/machine-learning/accelerate-protein-design-with-boltzgen-on-amazon-sagemaker-ai/
• BrainJanus, brain/vision/language arasinda unified understanding-generation modeli oneriyor; neuroscience ve multimodal AI siniri daraliyor. Kaynak: https://huggingface.co/papers/2606.30319
• Materials science literature icin multimodal dataset calismasi, bilimsel figurleri AI icin erisilebilir bilgi kaynagina cevirmeyi hedefliyor. Kaynak: https://huggingface.co/papers/2606.29667
• ReactionAtlas, chemical reaction networks icin ML destekli ab origine exploration sunuyor; discovery pipeline'lari otomatiklesiyor. Kaynak: https://arxiv.org/abs/2606.30778

INFRA

• Cloudflare Attribution Business Insights, crawler davranisini ve potansiyel degeri raporlayan dashboard sunuyor; AI bot observability, ticari pazarlik verisine donusuyor. Kaynak: https://blog.cloudflare.com/attribution-business-insights/
• HippoRAG, Bedrock + Neptune + personalized PageRank ile graph tabanli RAG mimarisini AWS stack uzerinde uyguluyor. Kaynak: https://aws.amazon.com/blogs/machine-learning/hipporag-neurobiologically-inspired-rag-using-amazon-bedrock-amazon-neptune-and-personalized-pagerank/
• MOA, codebase scale memory optimization automation icin profiling-guided LLM framework oneriyor; LLM'ler runtime/memory optimizasyonuna daha fazla girecek. Kaynak: https://arxiv.org/abs/2606.31368
• TIRx, frontier ML kernels icin open compiler stack olarak one cikiyor; kernel optimizasyonu model performansinin gizli carpani. Kaynak: https://tvm.apache.org/2026/06/22/tirx

SAFETY

• AWS "Safely Releasing Frontier Models to Customers" yazisi, frontier model release surecini guvenlik ve musteri deployment acisindan belgeliyor. Kaynak: https://aws.amazon.com/blogs/machine-learning/safely-releasing-frontier-models-to-customers/
• Moral Safety paper'i, modellerin ahlaki uyumunun sadece prompt compliance mi oldugunu test ediyor; high-stakes domain'lerde yuzeysel uyum yeterli degil. Kaynak: https://arxiv.org/abs/2606.31644
• RedVox, speech modellerinde fairness ve safety'nin dil genisliginde esitsiz oldugunu gosteriyor; voice agent'larda non-English risk yuksek. Kaynak: https://huggingface.co/papers/2606.26968
• When the Database Fails, task-oriented dialogue agent'larinda backend hata/empty-result durumunda safe recovery prompt'larini ele aliyor. Kaynak: https://arxiv.org/abs/2606.31307

WATCHLIST

• Agentic web payments: Cloudflare x402, MCP tool monetization ve crawler policy birlikte izlenmeli. Kaynak: https://blog.cloudflare.com/monetization-gateway/
• Coding agent cost controls: Copilot session limits, Palantir token elestirisi ve enterprise AI spend ayni yere isaret ediyor. Kaynak: https://github.blog/changelog/2026-07-01-set-ai-credit-session-limits-in-copilot-cli-and-sdk ve https://www.cnbc.com/2026/07/01/palantir-karp-open-ai-anthropic-tokens.html
• Open-source AI contribution policy: Godot karari diger projelere yayilirsa AI-authored PR'lar icin disclosure/provenance zorunlu olabilir. Kaynak: https://www.pcgamer.com/gaming-industry/open-source-game-engine-godot-will-no-longer-accept-ai-authored-code-contributions-we-cant-trust-heavy-users-of-ai-to-understand-their-code-enough-to-fix-it/
• Agent protocol governance: MCP, A2A ve ACP'nin ifade edemedigi governance gap'leri teknik standart riskine donusuyor. Kaynak: https://arxiv.org/abs/2606.31498

TRENDS

• Launches, Regulation ve Models en yuksek topic yogunlugunu verdi; ancak gercek karar sinyali "agent control plane" tarafinda: routing, budget, memory, sandbox, crawler policy. Kaynak: https://github.blog/changelog/2026-07-01-copilot-cli-auto-model-selection-routes-based-on-task ve https://aws.amazon.com/blogs/machine-learning/structured-memory-filtering-with-metadata-in-agentcore-memory/
• AI Agents topic'i 1113 toplam gorunurluge cikti; bugun bunu destekleyen yeni maddeler A2A gateway, LUMOS, GUI memory ve Docker isolation. Kaynak: https://aws.amazon.com/blogs/machine-learning/building-a-serverless-a2a-gateway-for-agent-discovery-routing-and-access-control/ ve https://arxiv.org/abs/2606.30697 ve https://www.docker.com/blog/why-ai-agents-need-isolation/
• AI Regulation spike'i sadece devlet kurali degil; Cloudflare, Godot ve Copilot credit limits gibi urun-policy kararlarinda da belirgin. Kaynak: https://blog.cloudflare.com/content-independence-day-ai-options/ ve https://github.blog/changelog/2026-07-01-set-ai-credit-session-limits-in-copilot-cli-and-sdk

CikCik Paketi

• Percy Liang tarafinda "open frontier" ve acik arastirma vurgusu devam ediyor; frontier AI liderligi sadece kapali model yarisi olarak anlatilmamali. Kaynak: https://news.google.com/rss/articles/CBMiZEFVX3lxTE40NnZfZ1ZxenRCWjBuUFM2Q2x5NzQ0WHluaTJ6VDZzZXM3VHRhTzFqNkhzOGlvblM4N216Z256eVFZWC1XdS16XzdWc3YxS1I1TG4zRUNYcklTaUhMYU5uMldnRHc?oc=5
• Harrison Chase tarafindan wiki memory pattern'i one cikti; agent memory tasariminda "kalici wiki + episodic state" ayrimi gucleniyor. Kaynak: https://news.google.com/rss/articles/CBMiX0FVX3lxTE1HNDdsbFZSMUo3Ujl4Y3F3U3pZWE1vSktNSE1nV1psV1ZBdGlfd2RIN0hKLW5tUU9VZHI2YlBodGVReUM5dHNJX0Riamk4T0ZzVENiRWxzekFMbVNEVUt3?oc=5
• Andrew Ng "loop engineering" tartismasini isaret etti; prompt engineering'den farkli olarak dongu, feedback ve eval tasarimi one cikiyor. Kaynak: https://news.google.com/rss/articles/CBMiX0FVX3lxTE53RjJ4ZlBrZEVnSjEwT051NlZKUHYtSUZaaWlBWnJPNGgxenlpSFNRU3B3cm10cXY5RGtKVGJVc2ttNVBjUDNEMGlJZzlnUGIwVTl4NmZzZW1UQkREZFZV?oc=5
• Clement Delangue, frontier model query'lerinin buyuk bolumunun daha kucuk/ucuz modellerle calisabilecegi fikrini one cikardi; cost routing daha kritik hale geliyor. Kaynak: https://news.google.com/rss/articles/CBMiXEFVX3lxTE9DSjhTSGpDVkktNE9qelFsR1d2YmhBdkJMaFJzc0MzWEpIVlJqWDFtQ1ppc08tWFQxY2c5NDdKTEpScVZ4cVA3T1M4YXUtVHVIQUVocVpBZ0lBaVE1?oc=5
• The New Oil / ResearchBuzz, AI browser attack riskini yaydi; browser-using agent'lar icin isolated browsing ve origin-aware guardrail zorunlu. Kaynak: https://mastodon.thenewoil.org/@thenewoil/116846375927610640 ve https://rbfirehose.com/2026/07/01/ars-technica-new-attack-provides-one-more-reason-why-ai-browsers-are-a-bad-idea/
• Mastodon'da Godot AI contribution yasağı topluluk sinyali verdi; AI katkisi kullanmakla maintainability sorumlulugu arasindaki gerilim buyuyor. Kaynak: https://mastodon.social/@beyondthecode/116843612919047556

Rehberler & Kaynaklar

• AWS A2A gateway post'u, agent discovery/routing/access control icin dogrudan uygulanabilir mimari rehber. Kaynak: https://aws.amazon.com/blogs/machine-learning/building-a-serverless-a2a-gateway-for-agent-discovery-routing-and-access-control/
• AWS AgentCore Memory metadata filtering, multi-tenant memory retrieval icin okunmali. Kaynak: https://aws.amazon.com/blogs/machine-learning/structured-memory-filtering-with-metadata-in-agentcore-memory/
• Docker AI agent isolation yazisi, local/remote agent calistirmada sandbox checklist'i olarak kullanilabilir. Kaynak: https://www.docker.com/blog/why-ai-agents-need-isolation/
• PyMuPDF 1.28 Markdown destegi, PDF/Markdown rapor otomasyonlari icin pratik guncelleme. Kaynak: https://www.reddit.com/r/MachineLearning/comments/1ukyciw/new_pymupdf_release_supports_markdown_n/

Oracle Self-Improvement Sinyalleri

• Agent memory icin wiki + metadata + active task state uclusu artik net pattern: AgentCore metadata filtering, GUI active memory ve Harrison Chase wiki memory ayni yonde. Kaynak: https://aws.amazon.com/blogs/machine-learning/structured-memory-filtering-with-metadata-in-agentcore-memory/ ve https://arxiv.org/abs/2606.31612 ve https://news.google.com/rss/articles/CBMiX0FVX3lxTE1HNDdsbFZSMUo3Ujl4Y3F3U3pZWE1vSktNSE1nV1psV1ZBdGlfd2RIN0hKLW5tUU9VZHI2YlBodGVReUM5dHNJX0Riamk4T0ZzVENiRWxzekFMbVNEVUt3?oc=5
• Browser kullanan agent'larda izolasyon kuralini sertlestirmek gerekiyor: Docker isolation, AI browser attack ve Copilot browser tools ayni risk yuzeyini buyutuyor. Kaynak: https://www.docker.com/blog/why-ai-agents-need-isolation/ ve https://rbfirehose.com/2026/07/01/ars-technica-new-attack-provides-one-more-reason-why-ai-browsers-are-a-bad-idea/ ve https://github.blog/changelog/2026-07-01-browser-tools-for-github-copilot-in-vs-code-are-generally-available
• Cost guard icin session-level cap standardi izlenmeli; Copilot AI credit session limits Oracle cron/agent harcama limitleri icin benzer pattern veriyor. Kaynak: https://github.blog/changelog/2026-07-01-set-ai-credit-session-limits-in-copilot-cli-and-sdk
• Self-improvement tarafinda AutoTrainess, SkillHone ve QVal ayni sinyali veriyor: kalici karar gecmisi + ucuz ara reward + kontrollu skill evrimi. Kaynak: https://arxiv.org/abs/2606.31551 ve https://huggingface.co/papers/2606.08671 ve https://arxiv.org/abs/2606.32034

Kaynak Ozeti

• Toplam 6908 ham item toplandi, merge sonrasi 6059 cikis, dedupe sonrasi 1895 benzersiz yeni item rapor havuzuna girdi. Kaynak: raw-sources/2026-07-02/merged.json ve raw-sources/2026-07-02/unique.json
• Kaynak aileleri 5/5 kapsandi: rss/news, search, community, social, academic/api. Kaynak: raw-sources/2026-07-02/unique.json
• En verimli haber kaynaklari: DonanimHaber, Planet AI, The Register; search tarafinda google_news/ai, google_news/companies, google_news/security baskin. Kaynak: raw-sources/2026-07-02/rss.json ve raw-sources/2026-07-02/search.json
• Academic/API tarafinda arxiv/cs.CV, arxiv/cs.LG ve arxiv/cs.AI en yogun kaynaklar oldu. Kaynak: raw-sources/2026-07-02/academic.json

Coverage / Blind Spots

• Eksik family yok: rss/news, search, community, social, academic/api tamamlandi. Kaynak: raw-sources/2026-07-02/unique.json
• Bos topic yok; 10/10 topic kapsandi. Kaynak: raw-sources/2026-07-02/unique.json
• Thin family yok; community daha kucuk ama 69 ham / 64 unique item ile yeterli sinyal verdi. Kaynak: raw-sources/2026-07-02/community.json
• Zayif nokta: Search akisi hala Google News/Bing aggregator URL'lerine dayaniyor; canonical URL cozumleme sonraki iyilestirme olmali. Kaynak: raw-sources/2026-07-02/search.json
• Social collector 4095 ham item verdi ama 385 unique'e dustu; fallback X/Google News akisi gürültülü, kalite filtresi daha agresif olmali. Kaynak: raw-sources/2026-07-02/social.json

Bu Gece Sistem Ne Ogrendi

• Kalici ders: Launches, Regulation ve Models yuksek hacimli; ancak bugunku karar degeri Agents + Security + Infra kesişiminden geldi. Kaynak: raw-sources/2026-07-02/unique.json
• Rising entities baglami bugun de dogrulandi: Google/Anthropic/OpenAI gorunurlugu devam ediyor, fakat yeni operasyonel sinyal Cloudflare, GitHub Copilot ve AWS AgentCore tarafindan geldi. Kaynak: https://blog.cloudflare.com/monetization-gateway/ ve https://github.blog/changelog/2026-07-01-kimi-k2-7-is-now-available-in-github-copilot ve https://aws.amazon.com/blogs/machine-learning/structured-memory-filtering-with-metadata-in-agentcore-memory/
• Dedupe dersi: Mythos/Fable, Claude Sonnet 5, GeneBench, Etched ve Amazon FDE gibi onceki 3 gun ana maddeleri tekrar edilmedi; bugunku rapor yeni sinyallere agirlik verdi. Kaynak: reports/2026-06-29.md ve reports/2026-06-30.md ve reports/2026-07-01.md
• Bir sonraki run icin ders: social fallback'te "@@" handle ve Google News X proxy kaynaklari kaliteyi dusuruyor; canonical tweet/post cozumleme ve recency filtresi guclendirilmeli. Kaynak: raw-sources/2026-07-02/social.json

Dedupe & Kalite Notu

Bu rapordaki haber maddeleri onceki 3 gunun ana ozet/top maddeleriyle karsilastirildi; belirgin tekrarlar ana haber yapilmadi. Toplam 6908 ham item islendi; merge 6059 cikis verdi; dedupe 1895 yeni item birakti; 4152 fingerprint tekrar ve 12 semantik tekrar elendi. Kapsam: 5/5 kaynak ailesi, 10/10 topic, 105 benzersiz kaynak.