Oracle Gece Arastirma v3 — 2026-07-12

Curado por Mahsum Aktaş · Escaneo diario automatizado del sector de IA

Oracle Gece Arastirma v3 — 2026-07-12

Otomatik derleme | v3 pipeline | 5578 raw item | 5306 merged item | 3261 benzersiz item | 79 benzersiz kaynak | 5/5 kaynak ailesi

Gunun Ozeti

Bugunun ana ekseni model yarisi degil, model ekonomisinin urunlesmesi: Gemma 4, DeepSeek V4 preview, Qwen 3.6-Plus, Mercury 2 ve SubQ gibi duyurular fiyat, context ve local calisma hattini zorluyor. Kaynak: https://news.google.com/rss/articles/CBMinAFBVV95cUxON3BxcXljd0k5Y2taSTFPZjc5ODdKUllWWjg1M2RrWm44WE9lY1l6WXBrN2dGT1JaZmZQbkVmTXRIbkpjRTlGaHdqYjF5V3kwbFlONTRucHQ5NVh1aW41ZnBHYzQ1c082UTBTNHEzcVplQTdHTHpQQk9BelhQZms4WDVSdmJGT2tQMVRMMjJCdk1RZ05pNHJKLV9GNms?oc=5

Ikinci eksen security tarafinda AI destekli vulnerability discovery'nin production response tarafina gecmesi: OpenAI Daybreak, Aardvark, CISA'nin 3 gunluk patch baskisi ve Writer/Google chatbot aciklari ayni hatta duruyor. Kaynak: https://news.google.com/rss/articles/CBMihAFBVV95cUxQalhKa28zNGFmVW0tVVRPTjlGVFJfbmJoMFZ1VVpIWHljTWR1S1lHbzVnSEstNklfRUE2dTBkXy1kRTU5cmN5MkFLN0NTakZvSWVaR28yV00zcjQ0dG4yYWM0dENVOG9hSlVzckVwaW5JeXFvU2U4ZkZKdjI5aTg1VUhhN0I?oc=5

Ucuncu eksen politika ve guven: Avustralya telif tartismasi, Google'in EU deadline oncesi AI reklam etiketleri ve University of Chicago Law'un cihaz yasagi, AI'nin kurumsal/sivil kullanim sinirlarinin yeniden cizildigini gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMisgFBVV95cUxPZm1ZbGlKbjZncnFvcFdJLUhFekxmNU52Ml9xekhNSTR2TUFGSkxkWThzLV9TcE9sbmp0YUNLZFZRd0VpM2hiclhJLWw2SFhtWlNnZWgxWmZtWjBIaEtLMkc0XzlqSzlaRmdBdjNTWDctdzJhWkM0ck0zekxMMS1DOUZ0MXZGVk1JeEtpZW05UE1veWtGcEVzalNaWm9qeVB5N2VFNWpmQVVFajNxQ2NxUTFR?oc=5

Top 7

1. Gemma 4 open model hattini yeniden isitiyor. Google'in Gemma 4 duyurusu, Apache 2.0 lisans vurgusu ve local/agent platformlariyla kullanim sinyalleri open-weight rekabetini yeniden canlandirdi. Kaynak: https://news.google.com/rss/articles/CBMinAFBVV95cUxON3BxcXljd0k5Y2taSTFPZjc5ODdKUllWWjg1M2RrWm44WE9lY1l6WXBrN2dGT1JaZmZQbkVmTXRIbkpjRTlGaHdqYjF5V3kwbFlONTRucHQ5NVh1aW41ZnBHYzQ1c082UTBTNHEzcVplQTdHTHpQQk9BelhQZms4WDVSdmJGT2tQMVRMMjJCdk1RZ05pNHJKLV9GNms?oc=5 ve https://news.google.com/rss/articles/CBMiZ0FVX3lxTE41VEFieGJsRGFQcGUyczZFNTJJTnp1VjNFbVE5ZkNCVG5OYURTQlBSajRXVVZwbDJCVFpWS2dqSXh5Z1E0T0lqVklmdGgyeXNPR3dva1ROWE1BelowVGctREx4WUIxVjg?oc=5
2. AI guvenlikte "bul"dan "fix et"e gecis hizlandi. Daybreak ve Aardvark basliklari, vulnerability detection + patch validation + agentic code fix akisini tek hatta bagliyor. Kaynak: https://news.google.com/rss/articles/CBMihAFBVV95cUxQalhKa28zNGFmVW0tVVRPTjlGVFJfbmJoMFZ1VVpIWHljTWR1S1lHbzVnSEstNklfRUE2dTBkXy1kRTU5cmN5MkFLN0NTakZvSWVaR28yV00zcjQ0dG4yYWM0dENVOG9hSlVzckVwaW5JeXFvU2U4ZkZKdjI5aTg1VUhhN0I?oc=5 ve https://news.google.com/rss/articles/CBMif0FVX3lxTE9mZW5vdWtFRTBIMzgycEVEXzBVSHhkNlBBMUJ3MEZQQnRkNlhBR3hITjhwcUIxS0ZGUGJvYmxMR2lJeVJReXFtVTRmc1RKb3F6R2pYeUNFQ3pnVWVOWjA2T0ZjTlBqdUNXWlBJQ0dDeFNVNjNIWGpRb0lxcG9tTlU?oc=5
3. 12M-token context penceresi startup savasina girdi. Subquadratic'in SubQ/12M-token iddiasi, agent trace, codebase ve enterprise memory islerinde context mimarisini maliyet tartismasina bagliyor. Kaynak: https://news.google.com/rss/articles/CBMingFBVV95cUxNVzBEQXlqMDlkZWIzMlRoYXg3dWpnMC05RWZkeEhvZmdoUG5tcUJqT1Mxd3VFY3NGX3lVUXZFSXRFZ2tNLVdXVWJqNzIwWjItRFN0VWJ3d1B6TDZPMHU0REk2dHFIOWlra21jSGctUUJ3RUI0YTJmNGlCMG1ielVsNGVONE9NZldCTzFsWEgtRWYyMm93R1RVWTA5aXA5QQ?oc=5
4. MCP ve LLM gateway katmani zero-trust urunu oluyor. NetFoundry enterprise MCP/LLM gateway, Mozilla Otari control plane ve agent skill repo'lari ayni mesajı veriyor: agent runtime artik network/security urunu. Kaynak: https://news.google.com/rss/articles/CBMi5gFBVV95cUxNR3dzSkVFaDRpXzAxa2h0cDBMamswTXFPSmxXUUZTN0I4YVVONURDRDJaZEN6UUZYRUtTVGhQRjJsbGpOdDJhMmlrS0pMYzJ5dUdQZjRGNHRNelZsem1wUENIcUtaaGVIT3ZJdlIyNmFPQ0JRZ0hvV1VkeDJZSjMwNXRpUU5qZFg2Szg4aTRzMk5DRUJxekhNVDRrT3AwUHZxajFIQnVEc1RsMVExT29wWFNlZ2tOOEd0aGlDRTQ2azNwZFVVa1RXRG9zOTJ4X1d1Qjc3cks4YVNsWjFWMjRUa1J6STEydw?oc=5
5. ChatGPT aile/caregiver hattina iniyor. OpenAI'in aileler, bakicilar ve yasli yetiskinler icin PM aramasi, ChatGPT'nin bireysel assistant'tan hane ici urun katmanina gecis sinyali. Kaynak: https://techcrunch.com/2026/07/11/openai-bets-on-families-as-chatgpt-goes-deeper-into-households/
6. Regulasyon ve telif baskisi sertlesti. Avustralya'da AI sirketlerinin telif kurallarini gevsetme istegi, sanatci tepkisi ve siyasi bolunme uretirken Google EU deadline oncesi AI reklam etiketi ekledi. Kaynak: https://news.google.com/rss/articles/CBMisgFBVV95cUxPZm1ZbGlKbjZncnFvcFdJLUhFekxmNU52Ml9xekhNSTR2TUFGSkxkWThzLV9TcE9sbmp0YUNLZFZRd0VpM2hiclhJLWw2SFhtWlNnZWgxWmZtWjBIaEtLMkc0XzlqSzlaRmdBdjNTWDctdzJhWkM0ck0zekxMMS1DOUZ0MXZGVk1JeEtpZW05UE1veWtGcEVzalNaWm9qeVB5N2VFNWpmQVVFajNxQ2NxUTFR?oc=5 ve https://news.google.com/rss/articles/CBMikAFBVV95cUxObzlDUFI1TTBkZEJONTBxbDJacjJzQ3lQc0tnYkdXQlVVRkROLTc4U2NvMzZWX0VoSm1lcW9xMDg2ekpVY0xSYWJQTVRxLW5aWlhsb1ZCaVJtSkdVM3Z6d3V4X2VXYVdNTWhZVzNPV0lsZnNmbUVIRzZ0bV81bE1LMF9sekFXZ05TdjBfeWw3TEk?oc=5
7. AI-science ve saglikta klinik reasoning basliklari yogunlasti. KMed.ai, clinical reasoning survey, hepatocellular carcinoma risk/treatment guidance ve alignment plausibility paper'lari, domain LLM'lerin generic chat'ten ayrildigini gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMilAFBVV95cUxNTHI3V05tM05vUzlXYXdVMjZnUHQ2ZWJCNnlEWi1yOXpKZnFxMHp6eFFnY1hPOHlKdGpRWG42UFR6ZzVpenBoTmFyMml2NWg2bUNrMFFLd3g0SGpJOXExS2ZOYnNxTDhFUXdFaVVXa0prTmRIZEY0S3h2a1Jvc2dLaHk5SmNJMkQyeFc3Rm5mVmVNOGt1?oc=5 ve https://arxiv.org/abs/2607.08602

TRENDS

• SPIKE: Apple, Claude 4, Cohere, Gemma ve GPT-4 trend setinde yuksek sinyal verdi; bugun Gemma 4 ve open model tartismasi en temiz yeni akis. Kaynak: https://news.google.com/rss/articles/CBMinAFBVV95cUxON3BxcXljd0k5Y2taSTFPZjc5ODdKUllWWjg1M2RrWm44WE9lY1l6WXBrN2dGT1JaZmZQbkVmTXRIbkpjRTlGaHdqYjF5V3kwbFlONTRucHQ5NVh1aW41ZnBHYzQ1c082UTBTNHEzcVplQTdHTHpQQk9BelhQZms4WDVSdmJGT2tQMVRMMjJCdk1RZ05pNHJKLV9GNms?oc=5
• YUKSELIS: AI Agents, AI Regulation, Anthropic, Claude ve Character.AI hatlari yukseliste; bugun bu sinyal agent governance, MCP gateway ve regulation haberleriyle beslendi. Kaynak: https://news.google.com/rss/articles/CBMi5gFBVV95cUxNR3dzSkVFaDRpXzAxa2h0cDBMamswTXFPSmxXUUZTN0I4YVVONURDRDJaZEN6UUZYRUtTVGhQRjJsbGpOdDJhMmlrS0pMYzJ5dUdQZjRGNHRNelZsem1wUENIcUtaaGVIT3ZJdlIyNmFPQ0JRZ0hvV1VkeDJZSjMwNXRpUU5qZFg2Szg4aTRzMk5DRUJxekhNVDRrT3AwUHZxajFIQnVEc1RsMVExT29wWFNlZ2tOOEd0aGlDRTQ2azNwZFVVa1RXRG9zOTJ4X1d1Qjc3cks4YVNsWjFWMjRUa1J6STEydw?oc=5
• AZALIS: ElevenLabs zayif sinyal verdi; ses tarafinda daha guclu sinyal OpenMOSS diarization ve Gemma 4 multimodal/audio tarafindan geldi. Kaynak: https://huggingface.co/OpenMOSS-Team/MOSS-Transcribe-Diarize

KAT-1 — Endustri

• Baseten'in 1.5B fonlama sinyali inference pazarini isitiyor. AI inference startup'larinda mega-round tekrar hizlandi; maliyet, latency ve hosted inference savasi devam ediyor. Kaynak: https://news.google.com/rss/articles/CBMipgFBVV95cUxPcXRvMkt3eE13OXJQdy1ad1ZSb2hNYS1FT1hVaGEyNUw4QnM3VnhNajlGREVHRTBUbFVMeTZwejNVVGVzRnpXSjByOV8zQ2ZGREVNakxFR0JFRi1FMWNMeFBwUWNWS3VPMUExalMtQ1BYQWhTd3VGdmllcXIzM0x2eTEzNkdqMmVONWNCNGlCQTExSHRSd010VmU0MFRGWUlsNTVDVjhn?oc=5
• Zoho, Zia LLM + Agent Studio ile enterprise suite savasina girdi. Prebuilt agents, custom agent builder, MCP ve marketplace paketinin ayni duyuruda gelmesi, productivity suite'lerin model degil workflow sattigini gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMi-AFBVV95cUxNZlBFYmc4bzhLTTNxZ2JGenF2ODhfU3pqcDZZUUtrMlZHRjVGV2VwSTdIQU4wR1FkMnA3R0FqRkhqYVhzMDdNNjBHcU1lVVVqUi1lbV9KUzN5cUpadWVWaE9aaEhpcDVsamt1TElqbVpEM3EzOER4WGdkdlVXWm83SXZvWmhCSzJWVVNfOXozVmRfLVVibFNGcm5Ic0Z5WTRqQlFBRDRvemZyVG0ycnIydm1vX28wMG5EZ0tEMV90YkF1UFRhMUJnU2tGOUxTQ3R0eGdMaEdVSloyYlBaTWR3M3YwWWJCVEVjbzFVT0xoSm1KbklkeDhLMw?oc=5
• Lumai optical inference iddiasi hardware narratifini genisletiyor. Real-time billion-parameter LLM inference icin optical computing duyurusu, GPU disi inference alternatiflerini watchlist'e aldiriyor. Kaynak: https://news.google.com/rss/articles/CBMihgJBVV95cUxPSFduZXFkT0EtY1FnX1Z1dkZLMkp4OG03LVBPNFJqdUN3T2k0d3NRMFRJNU9HaWZldWVHZ3FvN0ZxS0ppVFNwWUN0dlljVzlhUF9rTkhaS3VSQzhqNUNqMXF1ZEhVTmo0LXhsWDRxQzlfbGdKSkpVeTNKUkVGWjJuckN2NzRLNGVQVC11bWZMSEdKdjN6TXRaeXVWOGFUV0VXMFZoNkhQNE8zdXZObWlpTTk2TVZ5RVBSa2ZZNW9PZ0lEazRWTGRWbXFWWXdaUE9vaGgwbFdSY0VQWGhQLThUVGYzTEVBbTBnazhmaGZ5S1JRQ0U1QktTZUVhcEh1R0VLYlQ4a2R3?oc=5

KAT-2 — LLM & Model Guncellemeleri

• DeepSeek V4 preview, open-weight rekabetini daha sert hale getiriyor. CNBC akisinda DeepSeek'in beklenen V4 preview'i model yarisi ve Cin merkezli open model momentumu icin yeni sinyal. Kaynak: https://news.google.com/rss/articles/CBMimAFBVV95cUxPd3BGX1RVY2NmcVZoZnpqQ0hDTmdhem9nVE5wLWlKN2lIYzNuVjJtUjdqeU9oWWIxVjY4WG1BU0ctMm5VNURWZmhnU05IQTVnakRJUFJSeHI5dFB0WThCb1h1Q2Fnb3k5aUwyTUd6NEZ3bU1LRWs5OHZvTi1MQlJpVzNMbnRzcDd6UkZxdFlsX0FycjU5SVc4ONIBngFBVV95cUxQWWFDRGp5MGRYM2FYZ0lvN0Y5cTVOSm56QmZQeENWTzdUQzVXOVZIQjVZZy1OcnJLeEFzSklFVEpWdWV6d05POWktN2JIbkJzQnhSZDdtRm1VQzBEbjR3ZjNyQ0VjSkVpenFrWGM5MnFWdndlVTg5OGgwMUpJa0RONnF2em9hdGhycjNGdmpmZHlQSEdudTMyRm1mdjVQUQ?oc=5
• Qwen 3.6-Plus ve Qwen3.5 agent hattina baglaniyor. Alibaba/Qwen basliklari, Cin modellerinin sadece chatbot degil agent-native cloud stack icinde konumlandigini gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMingFBVV95cUxORUhVVFRPbDJ2UFZWOU1UMi02MTRwMXp5OWNXVHBUQ2F2aFZxWVNYbFg0VkNXRk5LTDFzcExUeG4tbmo2VjhwbU90MkFrVS1xenZmd3l5YVNxOS13WW5rSm1iT3hpbFF2WnBqaVg1cjA5RVIyempyRVBudUJ2MDNpV3FQTElmMThUaXE1MjB5bW5IdlFmWVNVUU03NDBhUQ?oc=5
• Mercury 2 hiz/maliyet ekseninden saldiriyor. Inception, Mercury 2'yi hizli reasoning LLM olarak konumlayip inference maliyetini ana satis noktasi yapiyor. Kaynak: https://news.google.com/rss/articles/CBMinwJBVV95cUxNZzJud2IyWHdGVDZ5dGFHSVNxX3VMbnVZLVdwYURwN3J5VUZHSHJyeFVQUzkwSVdKaUVwNlN6ck9KV0RMVVBLbE5la2FLZnloQ0hwNEo4bDVZMkQ0VnBDYV94NWxVcWs0NVJrZlhqZ2haUjItQWxrRk1jMTItdXh0bnBiWmoxUFV6LTZ3NU5DNUpnUC13UXhPZFViODRxUDZIUE5MUU54SUhON0V1UHRQQ2lQVU5WQlNFSk9jaFRGcm93MGVfYUJUM1ItMFRobUV0T3BmVTJtR0dkRlpscGs5eXMzUTJpZ3hqQWhkNlByS21nMFAwWk9BVHhLbThYaDdudVpvdzhqYXhVVWs2QzZQUzVtUFJDUkZNcXl5X0JwRQ?oc=5
• IBM Granite 4 hibrit Mamba/Transformer mimarisiyle ayrismaya calisiyor. VentureBeat akisinda "Western Qwen" ifadesi, bati merkezli open/enterprise model pozisyonlamasinin sertlestigini gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMisAFBVV95cUxNVHZfZnhxd3FCdElpVDNTSkNhMUxiQVc4ZU1xdnZvYnJob21rTVBXOU1NZS0wVU1WTzdReVgyZ0l4Z010cVZ0OEU5dDRuamQ2aW5yeVBKQS1XR2QtY0M0ZkM1aEwzUzduWjhoU0FISkVhMnJ3UGUzVWRXZWdHQ2JsM0NRQ2VQZWplYllCNXI5d0E2TVJxOGZGZzNPYkxIT1ZrMzRQWVlvRmIweUs3MHhpVQ?oc=5

KAT-3 — AI Agent'lar

• Aardvark, coding agent'i security maintainer rolune tasiyor. GPT-5 agent'in code flaw bulup fix etmesi, agent'lari sadece feature generator degil, surekli guvenlik operatoru yapma yonunde sinyal. Kaynak: https://news.google.com/rss/articles/CBMif0FVX3lxTE9mZW5vdWtFRTBIMzgycEVEXzBVSHhkNlBBMUJ3MEZQQnRkNlhBR3hITjhwcUIxS0ZGUGJvYmxMR2lJeVJReXFtVTRmc1RKb3F6R2pYeUNFQ3pnVWVOWjA2T0ZjTlBqdUNXWlBJQ0dDeFNVNjNIWGpRb0lxcG9tTlU?oc=5
• Japonya yerel yonetimler icin AI agent tartiyor. Personel acigi olan local government birimlerinde agent kullanimi, kamu operasyonlari icin somut adoption hattina donusuyor. Kaynak: https://news.google.com/rss/articles/CBMijAFBVV95cUxNQkZJWWRvNDRNQ2FKazluZWVka3MzWDBSUDhfNnNPNXI4X0hDalVRUndFVFBjUFpHMDZyNUwtUGFQSDZtX1ZPREVVdmFScUU5V2RDMElqLUx4Q3BjSWlGUXJiYkdXRGlPMThJNWlpc29sejJ5M1EtN3UwLUJfYzJTNU9BM1lxbTFCVEVhRA?oc=5
• LangSmith Fleet ve DeepAgents Deploy sinyali production agent yonetimine kayiyor. Social fallback akisinda LangChain ekosistemi agent'lari yaratma, kullanma ve deploy etmeyi enterprise workspace'e topluyor. Kaynak: https://news.google.com/rss/articles/CBMiX0FVX3lxTFAyNHdDMHJJNWoydjdvN2JXVUZncmRqeVk4dWlhYXgxQXdNREM5TzNaNmVtMGlQNzdkc3NReml6VlVUazBFTFJUMXFRUmlySnBJTWFvYWRrT0pTcFJkdS04?oc=5

KAT-4 — Arastirma & Paper'lar

• Overthinking paper'i model icindeki "learned secrets" riskini hedefliyor. Reasoning weights'i amplifiye ederek gizli bilgi/misalignment arama fikri, black-box audit'in yetmedigi noktalari gosteriyor. Kaynak: https://arxiv.org/abs/2607.08173
• Blind-Spots-Bench multimodal modellerin insan icin kolay ama model icin zor hatalarini olcuyor. Benchmark, "ortalama skor" yerine sistematik kor nokta haritasi cikarmaya odaklaniyor. Kaynak: https://arxiv.org/abs/2607.08317
• Ideas Have Genomes bilimsel lineage reasoning'i benchmark'luyor. Bilimsel fikirlerin mekanizma ve soy agaci uzerinden uretimini olcmek, arastirma agent'lari icin daha iyi eval sinyali. Kaynak: https://arxiv.org/abs/2607.08758
• MoE edge inference paper'i "MoE edge icin idealdir" varsayimini test ediyor. Consumer/edge hardware uzerinde MoE inference'in gercek avantajlari ampirik olarak sorgulaniyor. Kaynak: https://arxiv.org/abs/2606.21428

KAT-5 — Guvenlik & Alignment

• Writer AI agent preview token leak riski gundemde. Agent preview ve multi-tenant session token izolasyonu, enterprise AI app'lerde yeni attack surface olarak one cikti. Kaynak: https://news.google.com/rss/articles/CBMihAFBVV95cUxNZ2p6cDNUc2ZmZ09ZSjUxakJLNFc1QU00UGp2QV9yV2x6ZnI3UFFTN1pIeVc3R1QtQllzZk01QTZ5WkozajFpRmY5TWNnTVJsQ3pXQTA1d0xDYnhTYUNYRjlNQUNDUTQtSmVvTjRVZ2tFZXFSaHhqRWhsTjhpZGIwMkNmRHk?oc=5
• Google AI chatbot security flaw arastirmasi, consumer AI'nin kurumsal riske donustugunu gosteriyor. Chatbot yuzeyi sadece prompt problemi degil; auth, data boundary ve tool invocation riski de tasiyor. Kaynak: https://news.google.com/rss/articles/CBMifkFVX3lxTE1kVm5IU3lLZk55UEJ0ZXBrNTlScGhxUFlzbWtoUS1wMWVhXzBiYTB5ZFY3TncyM0kxR1JJUVRYeWJlTXhnZ3VpZWtRZ21wem1wdVZDQ0N4QUg5ZGlIZ3FWandzeDFzNm9scThCcHowZUtPVEFpN1A4dy1WNFV2QQ?oc=5
• CISA 3 gunluk patch baskisi AI-speed threat model'i kabul ediyor. Ajanslara AI tehditleri nedeniyle bazi bug'lari 3 gun icinde fix etme baskisi, vulnerability response SLA'lerini kisaltiyor. Kaynak: https://news.google.com/rss/articles/CBMibEFVX3lxTFB5UjJqckItb04teVI0TWtpck9NV0VfdDJsSTh0LWZiOFFqWVlsMUF0ZDBjLW9vLTdybWtQNE1maHpJQTNLdV80RHVTeXpvYXVfTjU3Tk45SEZoU1QybUFNMkRFVGVIT21aU3RMWQ?oc=5

KAT-6 — Regulasyon & Politika

• Avustralya telif tartismasi AI data access cephesini aciyor. AI sirketlerinin copyright kurallarini gevsetme istegi, sanatci tepkisi ve hukumet ici bolunmeyle karsilandi. Kaynak: https://news.google.com/rss/articles/CBMisgFBVV95cUxPZm1ZbGlKbjZncnFvcFdJLUhFekxmNU52Ml9xekhNSTR2TUFGSkxkWThzLV9TcE9sbmp0YUNLZFZRd0VpM2hiclhJLWw2SFhtWlNnZWgxWmZtWjBIaEtLMkc0XzlqSzlaRmdBdjNTWDctdzJhWkM0ck0zekxMMS1DOUZ0MXZGVk1JeEtpZW05UE1veWtGcEVzalNaWm9qeVB5N2VFNWpmQVVFajNxQ2NxUTFR?oc=5
• Google AI reklam etiketiyle EU uyumuna erken girdi. "How this ad was made" etiketi, synthetic media/ads disclosure tarafinda platformlarin hukuki takvime gore urun degistirdigini gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMikAFBVV95cUxObzlDUFI1TTBkZEJONTBxbDJacjJzQ3lQc0tnYkdXQlVVRkROLTc4U2NvMzZWX0VoSm1lcW9xMDg2ekpVY0xSYWJQTVRxLW5aWlhsb1ZCaVJtSkdVM3Z6d3V4X2VXYVdNTWhZVzNPV0lsZnNmbUVIRzZ0bV81bE1LMF9sekFXZ05TdjBfeWw3TEk?oc=5
• University of Chicago Law ilk yil siniflarinda elektronik cihazlari yasakliyor. Akademide AI ile mucadele modeli teknik tespit degil, fiziksel cihaz siniri olarak uygulanmaya basladi. Kaynak: https://news.google.com/rss/articles/CBMirwFBVV95cUxQVVBoR2ZaRlNnejIwNHFKWWYwSmpncUFGcFh3cVBTMVdJMjR1RnJudUZ5UUF5RENVOTRGNDBFWmFRcmtDdHpGM0pRMWdmVGlDUzBzT1o2aDd3SzVMWFc3VDM1ajFvYlhQSjgxSG5LXy1hTEZNcHJ0eGxkX1FDaDJfcDlQQjhBYVgyODY1aWV0elNXRDBjTUc2ZjUyU0dqMVoxOGw5V0h4Z0dNTlFPUzN3?oc=5

KAT-7 — Multimodal

• Gemma 4 multimodal ve local deneyimle konumlanıyor. Community akisinda Gemma 4 icin audio, image, video ve text input; desktop/mobile/local platform mesajlari one cikti. Kaynak: https://news.google.com/rss/articles/CBMiYkFVX3lxTE5xYWwwM04yLXZCSHg2MGRkVVk5VW9sVXM3ZWNhYVBKN0I0SjhGZG9pdDVrQ2ZobktncUlwdmZkaUNkSWNtWnpGM1pqV1ZNaDM2TUFFbHgzOTFod1ZlUkdhQUx3?oc=5
• Infinity-Parser2 document/multimodal parsing hattinda yeni teknik rapor sundu. Kontrollu data synthesis ve multitask RL ile document analysis tarafinda yeni benchmark/uygulama sinyali. Kaynak: https://arxiv.org/abs/2607.07836
• DocMaster hiyerarsik yapi farkindaligi ile document analysis'i hedefliyor. Agentic document workflow'larda flat OCR/RAG yerine yapi-aware parse katmani gerekiyor. Kaynak: https://arxiv.org/abs/2607.08539

KAT-8 — Veri & Altyapi

• Samsung storage virtualization'i SSD icine tasima sinyali verdi. AI/data workload'lari icin storage katmaninin compute'a yaklasmasi, inference/training pipeline darboğazlarini azaltabilir. Kaynak: https://news.google.com/rss/articles/CBMi8AFBVV95cUxNQTF3bVBCTmFlaFdLaTgwanhPa3BCRm5kcm1ZblZLOHVVX1NLdFJHbUgwVVVMQWJmdXVYSnV2SzlCc2FFZUYxM1J3SUI1LThqRkdFUlA5NHR5OGcxNVItMVN4TlBEVDhQZERGSF9pcUtSeVdvaVQ4Y0xMZnNnbGZyRnBFLWxjanZhU0VXX2NHR0d4ZEdKY2xUUldzemZXemxlOFM2WHk0S2d6UTUzMkktZWJxUkRKMWdwSHJxLVFnNE9vN3BCR0NKTE5Nelc5MzhGWWtiN0huWWs3LWg3NGUxQTJjWWFpZGFiQlFIbWVHaEs?oc=5
• AI memory startup 98M dolar fonla token cost azaltma hattina girdi. Context/memory maliyeti artik model kalitesi kadar yatirim alan bir kategori. Kaynak: https://news.google.com/rss/articles/CBMipgFBVV95cUxOQXppdlU5eU9CcThLWDluSjhoemx3YllRcDIyZFA4Z3owR0JKS1dNeFBBSDVLU3A1SDNscWVHOVJmMmw5Z3A1Rm9UWnJpaGVEQk4zRWRaa09SYUhqRU1ocFhWVk85WExIYU5ZMXJfR01GRG8teFVPQ3ZSWFRSRmJwVFY0cGMyWWZjbzZsSkt2Q2Z0NVd6dXJjLS01N2dELTdMX2dONFNB0gGrAUFVX3lxTE1WRzlOc2pBZks5amFpSDUtMnhCbUIwbS1DWF9DRkg2M0FTeWVFVHpwSzN6ZlBjb21TdDFjQlJ5ZDI3T1hDMTBSZ19ZUWdJenZUN1d5SFBtTVZSZ2lDdWYyaWozYXZNYk9SbTJJS2VwNmlrdkhpQS1xU19vN1JUUXBNVnVqclNRRGhPR1FhVEd0RHdZRk50TWdBbmIxY3VDMDBvOVp2aWlQMVZKaw?oc=5
• SMetric agent serving icin session-centric scheduling oneriyor. Agent trafiği tek istek degil, uzun session oldugu icin scheduler metriklerinin degismesi gerekiyor. Kaynak: https://arxiv.org/abs/2607.08565

KAT-9 — Acik Kaynak & Topluluk

• Agent skills repo'lari trendde. addyosmani/agent-skills, mattpocock/skills, google-labs-code/stitch-skills gibi repo'lar agent operasyon bilgisinin paketlenebilir skill formatina aktigini gosteriyor. Kaynak: https://github.com/addyosmani/agent-skills ve https://github.com/google-labs-code/stitch-skills
• DesktopCommanderMCP MCP tool yuzeyinin guc ve riskini ayni anda tasiyor. Terminal/file search/diff edit yetkisi, local agent kapasitesini artirirken sandbox/governance ihtiyacini buyutuyor. Kaynak: https://github.com/wonderwhy-er/DesktopCommanderMCP
• TencentDB Agent Memory lokal long-term memory hattina oynuyor. 4-tier progressive memory pipeline, agent hafizasinin vendor API'siz de urunlesebilecegini gosteriyor. Kaynak: https://github.com/TencentCloud/TencentDB-Agent-Memory

SECURITY

• AI-speed vulnerability management ana risk. CISA, Daybreak, Aardvark, Writer flaw ve Google chatbot flaw birlikte okundugunda guvenlik ekipleri icin yeni KPI "tespit ettim" degil "dogruladim ve patchledim" oluyor. Kaynak: https://news.google.com/rss/articles/CBMibEFVX3lxTFB5UjJqckItb04teVI0TWtpck9NV0VfdDJsSTh0LWZiOFFqWVlsMUF0ZDBjLW9vLTdybWtQNE1maHpJQTNLdV80RHVTeXpvYXVfTjU3Tk45SEZoU1QybUFNMkRFVGVIT21aU3RMWQ?oc=5
• Configuration risk AI app'lerde yeni default acik. "When configuration becomes a vulnerability" basligi, AI app'lerde policy/tool/config hatalarinin klasik code vuln kadar kritik hale geldigini isaret ediyor. Kaynak: https://news.google.com/rss/articles/CBMiywFBVV95cUxNd3Fmdm1rSkJjNTlrQ0tnNFdXcm92eTN3Qjd4eHF0UEY5akI2ckgxenQ5V2s0X3NOeVdidDBXNFpMdUd3OE5UbGlKQXpoUzUwdmRVQTN2TVRwd0Q3cmhTajhXRWh0LWJDTmtZMUFtdktiTlhBQ2VDVmU4WXEwMFpLckg5WklqUlE4djNmWG1SaVFuU1Z3NEUzN2syY21rVnhrcXBpZ1RQNG5sRmNhdDNDeXpkZUtMcXQ0LS1wUmpNM2ttTGJaVzVPVjlXUQ?oc=5

REGULATION

• AI disclosure, copyright ve akademik kullanimin ayni gun sertlesmesi tesaduf degil. Platformlar disclosure etiketleri eklerken, hukuk egitimi cihaz yasagina gidiyor ve telif lobisi model training verisini hedefliyor. Kaynak: https://news.google.com/rss/articles/CBMikAFBVV95cUxObzlDUFI1TTBkZEJONTBxbDJacjJzQ3lQc0tnYkdXQlVVRkROLTc4U2NvMzZWX0VoSm1lcW9xMDg2ekpVY0xSYWJQTVRxLW5aWlhsb1ZCaVJtSkdVM3Z6d3V4X2VXYVdNTWhZVzNPV0lsZnNmbUVIRzZ0bV81bE1LMF9sekFXZ05TdjBfeWw3TEk?oc=5
• AI law startup Norm'un 120M fonu regulation tooling kategorisini ispatliyor. Hukuk ve compliance ekipleri icin AI-native policy automation pazarinin sermaye cekmesi izlenmeli. Kaynak: https://news.google.com/rss/articles/CBMikwFBVV95cUxNZkxSeG9MLUpVWEctU1ZhWlJJMFNYVTBuX2lqLThUNFM4b215TGEtN1hnOGZDTWZvekR6QnhUSUNmNTBzQnlQNW5HU0RSVnZkZW5abmpRVGRCU3BtY25zQjlEdS1IZFJMQnVPaUVsa280TjllZkJZeEF2cnNhRkpCZUNXaG5YdzJEN2JfWHlSbFI1SDg?oc=5

AI-SCIENCE

• KMed.ai ve clinical reasoning paper'lari lokal/domain LLM yonunu guclendiriyor. Genel model yerine yerel dil, klinik workflow ve domain benchmark'lari saglik AI'ini belirleyecek. Kaynak: https://news.google.com/rss/articles/CBMilAFBVV95cUxNTHI3V05tM05vUzlXYXdVMjZnUHQ2ZWJCNnlEWi1yOXpKZnFxMHp6eFFnY1hPOHlKdGpRWG42UFR6ZzVpenBoTmFyMml2NWg2bUNrMFFLd3g0SGpJOXExS2ZOYnNxTDhFUXdFaVVXa0prTmRIZEY0S3h2a1Jvc2dLaHk5SmNJMkQyeFc3Rm5mVmVNOGt1?oc=5
• Reaction-network reasoning frontier modelleri catalyst discovery'ye tasiyor. Deneysel olarak dogrulanmis catalyst-selectivity hipotezleri, AI-science tarafinda "paper demo"dan lab feedback loop'a yakin sinyal. Kaynak: https://arxiv.org/abs/2607.08003
• AI-integrated agricultural resilience supply-chain modellemesine giriyor. Tarim tedarik zinciri risklerini biyofiziksel ve ekonomik sistemler uzerinden modellemek, AI-science'in policy/operations tarafina acilimi. Kaynak: https://arxiv.org/abs/2607.07759

INFRA

• LLM gateway, model gateway ve MCP gateway tek kategoriye yaklasiyor. NetFoundry, Otari, Concentrate AI ve Zoho MCP basliklari, agent altyapisinda gateway'in yeni kontrol duzlemi oldugunu gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMimAFBVV95cUxNRWdQNjJ5VDRwNWswcWtocHN5T2t0OERrSUM3M0oyS1BFbkpDSWdobzFaeHBSLWJWS01vUTJPRi1pb3l3QkJQRkY1TmM4SlF4dEhLR29jdUdyb1VpbjU4aDc5bFFJeTFhenpXR3ZtRFRXb3FrRjlfNlBocF9scnIzYUtCV1VndWtIMzZOSVhMY1U4dXJNVWFNcw?oc=5
• Long context bedava degil. Prompt-pruning ve SubQ gibi akislar, "daha buyuk context" ile "daha ucuz context" arasindaki ayrimi netlestiriyor. Kaynak: https://towardsdatascience.com/long-context-isnt-free-i-built-a-safe-prompt-pruning-layer-that-makes-llm-systems-work/

SAFETY

• Alignment Plausibility saglik AI'inda yeni guvence standardi oneriyor. Mental health/clinical LLM'lerde sadece benchmark skoru degil, ihtiyac-kabiliyet uyumu ve zarar siniri olculmeli. Kaynak: https://arxiv.org/abs/2607.07766
• Persuasion attacks CoT monitoring'i zayiflatabilir. CoT monitoring'in safety mekanizmasi olarak tek basina yeterli olmadigini gosteren bu hat, agent supervision icin kritik. Kaynak: https://arxiv.org/abs/2607.08066
• Psychological competence eval seti eksik boyutu isaret ediyor. Teknik accuracy yerine kullanici psikolojisi ve sosyal etkiyi olcmek, consumer assistant safety icin gerekli. Kaynak: https://arxiv.org/abs/2607.08285

WATCHLIST

• Gemma 4 adoption hizi: Open model etkisi gercek mi, yoksa launch hype mi, 7 gunluk repo/download/social trend ile izlenmeli. Kaynak: https://news.google.com/rss/articles/CBMiYkFVX3lxTE1nZk13dE05Q0VFZFdEdjBhaGNYSWVkeW5pQk54RmZEdUJtb0d2LTM0ZVlYNkdXaC0yQTR6elg3N1ZxVWFHREFIQU9JaFRNUzRUdmU2R28xaXQxaml0QzRrYU13?oc=5
• Daybreak/Aardvark gercek patch validation: Duyurular guclu; kritik soru false positive, regression ve exploit validation maliyeti. Kaynak: https://news.google.com/rss/articles/CBMihAFBVV95cUxQalhKa28zNGFmVW0tVVRPTjlGVFJfbmJoMFZ1VVpIWHljTWR1S1lHbzVnSEstNklfRUE2dTBkXy1kRTU5cmN5MkFLN0NTakZvSWVaR28yV00zcjQ0dG4yYWM0dENVOG9hSlVzckVwaW5JeXFvU2U4ZkZKdjI5aTg1VUhhN0I?oc=5
• SubQ 12M-token iddiasi: Uzun context benchmark'i kadar latency, retrieval overlap ve hallucination drift'i izlenmeli. Kaynak: https://news.google.com/rss/articles/CBMidEFVX3lxTE5zY3NmMUJWZXJIbF9GTk5pZVR3ZHl3ZE1HTWl2WkNJYVU1VklOMzBJQWdxV1dBOTFiTVl5dGowSnJ2U09fTGozZTF1U3IyWEsxOWdTbS1yLUpsYUcxVGd5TVVGOTQ5ampiUXI4NTlLaVBrVHNZ?oc=5
• MCP gateway security: NetFoundry/Zoho/Otari gibi cozumler tool permission ve audit log kalitesiyle ayrisacak. Kaynak: https://news.google.com/rss/articles/CBMi5gFBVV95cUxNR3dzSkVFaDRpXzAxa2h0cDBMamswTXFPSmxXUUZTN0I4YVVONURDRDJaZEN6UUZYRUtTVGhQRjJsbGpOdDJhMmlrS0pMYzJ5dUdQZjRGNHRNelZsem1wUENIcUtaaGVIT3ZJdlIyNmFPQ0JRZ0hvV1VkeDJZSjMwNXRpUU5qZFg2Szg4aTRzMk5DRUJxekhNVDRrT3AwUHZxajFIQnVEc1RsMVExT29wWFNlZ2tOOEd0aGlDRTQ2azNwZFVVa1RXRG9zOTJ4X1d1Qjc3cks4YVNsWjFWMjRUa1J6STEydw?oc=5

CikCik Paketi

• Clem Delangue, Gemma 4'un Apache 2.0 olmasini "real open-source licensed" olarak okuyor. Bu, Hugging Face/open model cephesinde lisans safiyetinin tekrar one cikmasi demek. Kaynak: https://news.google.com/rss/articles/CBMiZ0FVX3lxTE41VEFieGJsRGFQcGUyczZFNTJJTnp1VjNFbVE5ZkNCVG5OYURTQlBSajRXVVZwbDJCVFpWS2dqSXh5Z1E0T0lqVklmdGgyeXNPR3dva1ROWE1BelowVGctREx4WUIxVjg?oc=5
• Tri Dao, Qwen3.5 agentic workload icin 580 tps sinyalini paylasti. Inference optimizasyonlari artik model release kadar kritik. Kaynak: https://news.google.com/rss/articles/CBMiXEFVX3lxTE5aSVY2OFh1ekdEMGpxQ0dsWG00M1VEb3ZYRjhRT3lrMkNCZ0lLVjNPUXIxbkNDVDd6c3B1dGR4ZUNUODJFa3ZHX3RDZFJ4Q0t5dmplOC1YcXlTc1Uw?oc=5
• Georgi Gerganov, llama.cpp MTP ile local modellerin daily-driver hizina yaklastigini vurguladi. Local AI pratikligi token/saniye ve MTP gibi inference ayrintilarinda belirleniyor. Kaynak: https://news.google.com/rss/articles/CBMiY0FVX3lxTE1Bc3pZWC1xcnVsQTg3MGFHTnNTcjNLV1AyRlB1R3VJTDVzVk1hLTVhd28zazczZzBZU0sxdE1lRi1fTURWX3BBbWRLS0dVajB5RlltckVTejM0Z0c4a2pSTUZ2dw?oc=5
• Philipp Schmid, dynamic MCP discovery ile context bloat'u hedefliyor. MCP'nin gizli vergisi tool schema sisme problemi; dynamic discovery agent maliyetini dusurebilir. Kaynak: https://news.google.com/rss/articles/CBMiY0FVX3lxTE5ZTlVPM2g4N2dpbHhhaHp4ejI1NlpxaXIxYXAtQVlibTBqZFV5RnVNN2Jkai1IMlZ6X0Q3bDdYcWNOZEhCUHRraGFaTVQ1SHotT3dZcVFMNHpleG00ZGFLTnBISQ?oc=5
• Andrew Ng, "best model" degisken oldugu icin hosted/open model secimini surekli yeniden tartiyor. Enterprise AI stratejisinde tek vendor'a kilitlenme riski arttikca model routing onem kazanacak. Kaynak: https://news.google.com/rss/articles/CBMiX0FVX3lxTE1tM3dKTldsanhKcHNHcXFpZHFEcm5yaEVFTVR5SG1YM1RxZ1FmeFpyZU1rYjZ5OW5FcTdzaTk5OHJLbjF1Rm5WMlZFYzhqMnR6Wk1rNFoyOUpxemM0a1E4?oc=5

Oracle Self-Improvement Sinyalleri

• Collector coverage tam: 5 kaynak ailesi calisti; RSS, search, community, social ve academic/API aileleri kapsandi. Kaynak: raw-sources/2026-07-12/unique.json
• Dedupe maliyeti alarm: 5306 merged item -> 3261 unique; 2031 fingerprint tekrar ve 14 semantik tekrar bulundu. Dedupe komutu uzun calisti, social fallback hacmi O(n) sonrasi kalite riskini buyutuyor. Kaynak: raw-sources/2026-07-12/merged.json
• Social fallback cok gürültülü: 4112 social raw item icinden 2800 unique cikti; eski tarihli X fallback item'lari CikCik paketine karisabiliyor. Recency filtresi sertlestirilmeli. Kaynak: raw-sources/2026-07-12/social.json
• Canonical URL ihtiyaci devam ediyor: Search ailesi Google/Bing RSS aggregator linklerine dayaniyor; publish once canonical resolver rapor kalitesini artirir. Kaynak: raw-sources/2026-07-12/search.json

Coverage / Blind Spots

• Coverage: 5578 raw item, 3261 unique item, 79 distinct unique source, 5/5 source family covered. Kaynak: raw-sources/2026-07-12/unique.json
• Top source family: Social 4112 raw / 2800 unique ile baskin; search 504 raw / 270 unique; RSS 595 raw / 86 unique; academic/API 323 raw / 75 unique; community 44 raw / 30 unique. Kaynak: raw-sources/2026-07-12/unique.json
• Blind spot: Missing family yok; thin family yok; empty topic yok. Ana kalite riski kaynak eksigi degil, social fallback gürültüsü ve Google News canonical URL sorunu. Kaynak: /tmp/v3-prompt.txt

Bu Gece Sistem Ne Ogrendi

• Durable pattern: Launches, Regulation, Models, Agents ve Tooling tekrar en yogun temalar; bugun bunlar Gemma 4, Daybreak/Aardvark, MCP gateway ve copyright basliklarinda birlesti. Kaynak: /tmp/v3-prompt.txt
• Rising entities: Meta, AI Regulation, Google, AI Agents, Anthropic, Claude, OpenAI ve RAG yukselen varliklar olarak kaldi; bugun Google/Gemma ve AI Regulation sinyali en temiz yeni katkiyi verdi. Kaynak: /tmp/v3-prompt.txt
• Bir sonraki run dersi: Social collector icin tarih filtresi, search collector icin canonical URL cozumleme, dedupe icin daha erken family-level pruning uygulanmali. Kaynak: raw-sources/2026-07-12/social.json

Kaynak Ozeti

• RSS/news: 595 raw item toplandi; DonanimHaber, Nikkei Asia ve Technopat baskin kaynaklar oldu. Kaynak: raw-sources/2026-07-12/rss.json
• Search: 504 raw item toplandi; google_news/ai, google_news/releases ve google_news/companies ana tasiyicilar oldu. Kaynak: raw-sources/2026-07-12/search.json
• Community: 44 raw item toplandi; GitHub Trending ve Lobsters ana sinyal verdi. Kaynak: raw-sources/2026-07-12/community.json
• Social: 4112 raw item toplandi; Twitter fallback ve Mastodon akislari CikCik havuzunu besledi. Kaynak: raw-sources/2026-07-12/social.json
• Academic/API: 323 raw item toplandi; arXiv cs.AI ve Hugging Face trending models ana akademik/model kaynaklari oldu. Kaynak: raw-sources/2026-07-12/academic.json

Dedupe & Kalite Notu

Bu rapordaki maddeler onceki 3 gunun raporlarindaki ana Top 7 tekrarlarindan arindirildi. Pipeline 5578 raw item isledi, 5306 merged item uretti, 3261 benzersiz item rapor havuzuna kaldi. Otomatik report-generator LLM cagrisinda bos yanit dondugu icin ana Turkce rapor ayni unique.json ve prompt baglami uzerinden manuel sentezlendi.