Oracle Gece Arastirma v3 — 2026-07-15

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

Oracle Gece Arastirma v3 — 2026-07-15

Otomatik derleme | v3 pipeline | 8.404 raw item | 7.233 merged item | 6.159 unique | 5.505 raporlanabilir item | 111 benzersiz kaynak | 5/5 kaynak ailesi

Gunun Ozeti

Bugunun ana ekseni agent ekosisteminin "runtime + discovery + governance" katmanina oturmasi: Google ve partnerlerinin Agentic Resource Discovery standardi, Genkit Agents API ve AWS/NVIDIA agent workflow yazilari ayni yone isaret ediyor. Kaynak: https://www.infoq.com/news/2026/07/agentic-resource-discovery-spec/ ve https://www.infoq.com/news/2026/07/genkit-agents-api-preview/ ve https://developer.nvidia.com/blog/how-to-run-an-autoresearch-workflow-with-rl-agent-skills-and-nvidia-nemo/

Ikinci eksen security ve reguelasyon: GitHub Copilot security review, AI code scanning, ABD AI/cyber koordinasyon grubu, Almanya'nin AI Overviews'u medya hukuku altina almasi ve Google'a yeni telif davasi, AI sistemlerinin artik urun degil duzenlenebilir operasyon altyapisi oldugunu gosteriyor. Kaynak: https://github.blog/changelog/2026-07-14-security-reviews-now-available-in-the-github-copilot-app ve https://github.blog/changelog/2026-07-14-code-scanning-shows-ai-security-detections-on-pull-requests ve https://news.google.com/rss/articles/CBMixAFBVV95cUxNS2lZSUhvSjVONmN2RVZKZ1M4R0s1T2JySXFEUDdSWEJweHFZMi1NZDBjaG5iM0F5NmJ2QWFTeklpSE5EdjZIWnRlcDI5ejRUWmpmTlZDMWx4Y25MQVJZZXFyYVNRTTlOTzhZZnFKR3lPV0dzeTEyTllIT3JJZjJGakIzajN5TmZXRUhsVFpZckhjRVRoLUM0dGJGZDlCMVg2aGRwVlFlTXhUR0xrU3ZEcDgyaWw4TzRDN213M0JUTzliR3lC?oc=5 ve https://news.google.com/rss/articles/CBMixAFBVV95cUxPRmhqd1N1RlV3RnBERVcxdlpaeDVnQnI2WlRiY2ZTRk95ZXhDNVNwbVJMcXljaVBmWWlMZE5YOUNkZC1tNWV2dF94NXFhZzQ4b0pydHYyX1NySjBvbVNkY3BmTXhqNEpmcmhmdzZSLXVONU5ValZHZWd1WEhHam95ZHltaU1uVFhzZmwzamM3UDdSYllkeVZackFDUl9VYTJHYWRDYVJnbExwTGFSVzlDZzV4ak5IZmNxYTlyQzhCRENHcVN4?oc=5

Ucuncu eksen compute ve sermaye: SoftBank'in yillik 5 trilyon dolar AI yatirim ihtiyaci vurgusu, datacentre onaylarinin hizlandirilmasi, TYLsemi chip finansmani ve Helsing'in 1,8 milyar dolarlik turu, model yarisi kadar altyapi yarisi oldugunu da netlestirdi. Kaynak: https://news.google.com/rss/articles/CBMilwFBVV95cUxNZ1EtMWZOUHBraVd5eVpHMmRnQVhqak1OTmd0WTE5aGpuNGc3cmJtSmVCQWp2dXZKMjIwZ3RtYVNBaGNBZ0NPdW9zR0kySE10UXJJZVJPd2dHX1VtTjNCb3lEWmE4YUlkemFRRHh0THJnV19UNkFQZV9SU1pRSGt5QlZqOUI3QWNmUVpHZFo4X3ozVkkySTZj?oc=5 ve https://news.google.com/rss/articles/CBMi1gFBVV95cUxQMHFsVkpkbTV3bFllZWJxSENQNGU1QVQySnR1Mzc0alJTcWJPQV9VdTR2SUlRdVhoTGJrMEZ4YlVaSGdMb19YdDU4dU51MVpBTkY4aVlaWXdCT2d0LTN5NlZadzlKS1Y5ZzViamFHR3VLSzR1V3Jnbzd2VGpUUmtCdWlxdXJiSzJ3b3U2WHc1UHZrVEd0bE5hbTVJYTdwbkVrOFQ5VXF2RS0zeUVnbUNiV0FwaGZwTktsS2pXd3Bmd3cwTmpZVndLZFZCV3Ytbm15aG5Zb2VR?oc=5 ve https://news.google.com/rss/articles/CBMiqwFBVV95cUxNRDF2bG9MczFPUVR3aHFPN1ZxMUFtTGpPZXR5aUVlR2NLU1VvSnBfOE12c21JcmV6MEdsTlk2MEdtZ3pNVWdmc2FlRmdqXzgxVUNWaTZHd3NFN0N4ZjNieUpPZzdJMHR1S254UjJOcloxcjdybF80MnFEZGFzN1BPYWxpdnBlZndWVWVKSHFFa3FqcEtjNWVOWEJsRjA3RWM3VjF5US1MemtUclE?oc=5

Top 7

1. Agentic Resource Discovery standardi aciklandi. Google ve partnerleri agent'larin tool/API/agent kaynaklarini yayinlama, kesfetme ve dogrulama bicimini standartlastirmaya calisiyor; bu, MCP sonrasi agent internetinin discovery katmani olabilir. Kaynak: https://www.infoq.com/news/2026/07/agentic-resource-discovery-spec/
2. Google Genkit Agents API preview geldi. TypeScript ve Go icin detached turns, human-in-the-loop ve state persistence gibi production agent ozellikleri tek API yuzeyinde toplanıyor. Kaynak: https://www.infoq.com/news/2026/07/genkit-agents-api-preview/
3. NVIDIA RL agent skills ile autoresearch workflow anlatti. NeMo uzerinden agent skill tabanli arastirma otomasyonu, "agent egitimi"ni sadece prompt degil environment + workflow tasarimi haline getiriyor. Kaynak: https://developer.nvidia.com/blog/how-to-run-an-autoresearch-workflow-with-rl-agent-skills-and-nvidia-nemo/
4. GitHub Copilot app icinde security review yayina alindi. Kod inceleme ve security scanning Copilot yuzeyine yaklastikca AI-assisted SDLC'nin varsayilan kalite kapisi genisliyor. Kaynak: https://github.blog/changelog/2026-07-14-security-reviews-now-available-in-the-github-copilot-app
5. ABD AI ve cybersecurity koordinasyon grubu kurmaya hazirlaniyor. AI guvenligi artik tekil ajans/kurum meselesi degil, ulusal cyber koordinasyon konusu olarak paketleniyor. Kaynak: https://news.google.com/rss/articles/CBMixAFBVV95cUxNS2lZSUhvSjVONmN2RVZKZ1M4R0s1T2JySXFEUDdSWEJweHFZMi1NZDBjaG5iM0F5NmJ2QWFTeklpSE5EdjZIWnRlcDI5ejRUWmpmTlZDMWx4Y25MQVJZZXFyYVNRTTlOTzhZZnFKR3lPV0dzeTEyTllIT3JJZjJGakIzajN5TmZXRUhsVFpZckhjRVRoLUM0dGJGZDlCMVg2aGRwVlFlTXhUR0xrU3ZEcDgyaWw4TzRDN213M0JUTzliR3lC?oc=5
6. Senato tasarisi kullanicinin kendi AI agent'ini Amazon/Google'a getirmesini gundeme aliyor. Platform kilidi, agent portability ve veri tasinabilirligi tartismasi consumer internet katmanina iniyor. Kaynak: https://news.google.com/rss/articles/CBMiyAFBVV95cUxOOWtwTkM1ekxYSGFEOVpKdk5Xb1NpTV96TVJibVY3SFpWRXdseGRHSnBkbjdVWGFyUERBeG9GeGpld2E3d0tiTFN0UjRRZHdVNl9ITGhDak9adEVqZnN6d2FZWW5DMV9EV3JucHJkcVZQQzh5ZTBHU19pOERpTTFpNEZYMm13dGR4SlV5Yk83UzZpME40YXkwQVdFU25IN0lVNmZHYTRQWjJzMFpocldoai1lV2plWlMyWXlYMEFybFhxNFNhN2wweQ?oc=5
7. Meta'nin noninvasive BCI sistemi Brain2Qwerty yuzde 61 dogruluk raporlandi. EEG/MEG sinyallerinden cumle cozumleme hatti, multimodal AI'nin insan-bilgisayar arayuzune dogru genisledigini gosteriyor. Kaynak: https://www.infoq.com/news/2026/07/meta-brain-interface/

TRENDS

• Spike: AI Safety, Autonomous, DeepMind, DeepSeek ve Flux sinyalleri haftalik pencerede sican basliklar. Bugunku veri bu paterni security, agent runtime ve altyapi uzerinden dogruladi. Kaynak: /tmp/v3-prompt.txt
• Yukselis: AI Regulation, Amazon, Apple, Cohere ve Gemma akisi yukseliste; bugun reguelasyon Google/Meta/egitim/hukuk, Amazon ise Bedrock/Nova Act uzerinden gorundu. Kaynak: https://aws.amazon.com/blogs/machine-learning/accelerating-software-delivery-with-agentic-qa-automation-using-amazon-nova-act-part-2/ ve https://news.google.com/rss/articles/CBMixAFBVV95cUxPRmhqd1N1RlV3RnBERVcxdlpaeDVnQnI2WlRiY2ZTRk95ZXhDNVNwbVJMcXljaVBmWWlMZE5YOUNkZC1tNWV2dF94NXFhZzQ4b0pydHYyX1NySjBvbVNkY3BmTXhqNEpmcmhmdzZSLXVONU5ValZHZWd1WEhHam95ZHltaU1uVFhzZmwzamM3UDdSYllkeVZackFDUl9VYTJHYWRDYVJnbExwTGFSVzlDZzV4ak5IZmNxYTlyQzhCRENHcVN4?oc=5
• Azalis: Character.AI, Grok, Llama 4, Midjourney ve Perplexity bugunku run'da daha zayif kaldı; buna ragmen Grok/terminal benchmark ve frontier release sezonu sosyal akista sinyal vermeye devam ediyor. Kaynak: https://news.google.com/rss/articles/CBMiXkFVX3lxTE4yQVpLUy1id2JqWEZ6Q2d1bUhoUUR3d2RHZVVobTJWNVFyWWpaU0E2UWJ5cXVkLS1XYkZLcGx1MDlVRlc5RF9hMjRJdGoxYmxvUlZBRjg4RDBxX3NkTmc?oc=5

KAT-1 — LLM & Model Guncellemeleri

• Index-1.9B teknik raporu: Bilibili kaynakli kucuk model serisi, small language model hattinda acik/ucuz inference seceneklerini genisletiyor. Kaynak: https://arxiv.org/abs/2607.09885
• GigaChat Audio: Time-aware large audio language model, audio-native LLM tarafinda zaman duyarliligini one cikariyor. Kaynak: https://arxiv.org/abs/2607.10387
• Silent Failures in Quantized LLM Reasoning: Quantization sonrasi accuracy korunurken reasoning davranisinin kayabilecegini siniflandiriyor; production eval icin kritik. Kaynak: https://arxiv.org/abs/2607.09999
• LeRoPE: Learnable RoPE frequencies ile language modeling performansini iyilestirme hatti aciliyor. Kaynak: https://arxiv.org/abs/2607.10134

KAT-2 — Arastirma & Paper'lar

• CDFM ve DAG-FM: Causal discovery icin foundation model arayisi iki ayri paper ile guclendi; bilimsel otomasyon icin onemli. Kaynak: https://arxiv.org/abs/2607.11508 ve https://arxiv.org/abs/2607.11510
• Backpropagation as a Nilpotent Linear System: Backprop'un matematiksel yapisini lineer sistem olarak yeniden yorumluyor; teorik DL icin takip edilmeli. Kaynak: https://arxiv.org/abs/2607.11289
• Conservation Laws for Diffusion Models: Diffusion modellerinde korunum yasasi perspektifi, stability ve control tarafinda yeni analiz kapisi olabilir. Kaynak: https://arxiv.org/abs/2607.10067
• Auditing the Risk Claims of Distributional RL: Distributional reinforcement learning'deki risk iddialarini audit etmeye odaklanıyor. Kaynak: https://arxiv.org/abs/2607.11607

KAT-3 — Araclar & Framework'ler

• Google Genkit Agents API: Detached turns, HITL ve persistence agent framework standardina dogru itiyor. Kaynak: https://www.infoq.com/news/2026/07/genkit-agents-api-preview/
• AgentCheck: MCP uzerinde LLM agent davranisini reproduce-intervene-mitigate dongusuyle test eden workbench. Kaynak: https://arxiv.org/abs/2607.11098
• TreeThink: Matematiksel reasoning icin modular tree search library. Kaynak: https://arxiv.org/abs/2607.11258
• ToFu: Arastirmacilar icin white-box, token-efficient agent harness. Kaynak: https://arxiv.org/abs/2607.11423

KAT-4 — Acik Kaynak

• OpenCut: Web, desktop ve mobile icin open-source CapCut alternatifi trend listesine girdi. Kaynak: https://github.com/OpenCut-app/OpenCut
• Graphify: Kod, SQL, dokuman, paper, imaj ve video klasorlerini sorgulanabilir grafa ceviren AI coding assistant skill'i. Kaynak: https://github.com/Graphify-Labs/graphify
• github/spec-kit: Spec-driven development icin GitHub kaynakli toolkit, plan/spec merkezli kod akisini guclendiriyor. Kaynak: https://github.com/github/spec-kit
• Nutlope/hallmark: Claude Code, Cursor ve Codex icin "AI slop" tasarim kokusunu azaltmaya odaklanan design skill. Kaynak: https://github.com/Nutlope/hallmark

KAT-5 — Endustri & Sirketler

• Chai Discovery 400 milyon dolar topladi. AI drug discovery sermayesi hala yuksek hacimli geliyor. Kaynak: https://news.google.com/rss/articles/CBMilwFBVV95cUxNeDlRODFkM0tGYnRSSUdHMGFBYU5YUGdzaWFEcHpDa2tuX0hTUFJ3dXhYVEVSRzBLN1lrbTlyN2M5bDY1Mm10VWtiN2hELVZlZTYtMzdocEVoVTUyUHhHTDR6eHc4ZWsyM2lpWnkzQXpmVDgwNkRjRThtSHZyWnE0Z1NxM2toRzd4SS12WEt4WU9YVWhSeUww?oc=5
• Alibaba, AIsphere icin 439 milyon dolarlik turu yonetti. Video generation pazari sermaye cekmeye devam ediyor. Kaynak: https://news.google.com/rss/articles/CBMiwgFBVV95cUxPcGtWdEVLLVkxN1FMQ0ZzQ29mOGQ4NjNnNXZVbm9XSEd2NGc2SGpsLXVRRFM0SUZwR3BQNTJDcGFlR1k5cUVCYTNJSXJ2eTNjWTJILVBrM2FNMG80b3JyWXh1bm5BV1hTYnBrd1RGUVhGd3FzYzlSV3hqbTNGa2hObGJHZzFudzV0NG9yTlg5Y3RvNjJ2bkFjalJTd0xMbGxyWnlUTE1CNUZJOFdic1c2VFl4TGFocjZOR2p2VzFzVWNsZw?oc=5
• Helsing 1,8 milyar dolar Series E aldi. Defense AI ve Avrupa uretim kapasitesi buyuk sermaye cekiyor. Kaynak: https://news.google.com/rss/articles/CBMiywFBVV95cUxNc01zSUZfSEVUeDN2cVlieHdqUlVnSnRyWXFOemt3T3lZdThHVWVTWEEtckd6ZW9vQXMzT2M1ZTlXbzQ5QlBrMFZZT3ZlejdBSFVKUlFfZU5ucjFzeDMtODMyQklKTGtTSFpUYU9TRUxqWE5sQ2kxdDQ4elJaVnVqcFRscnFtanh3SjM3WDdjTnU5ekt6T3c3YTJYZXVfSUJFQWZFZW9kT1d4TkJUMW45UnF5ekV0ZzZzY2dxTVJGT2N1SURnZWRLYnhKYw?oc=5
• TYLsemi 43 milyon dolar erken asama fonla AI chip design'a girdi. Custom silicon yarisi startup seviyesine yayiliyor. Kaynak: https://news.google.com/rss/articles/CBMiqwFBVV95cUxNRDF2bG9MczFPUVR3aHFPN1ZxMUFtTGpPZXR5aUVlR2NLU1VvSnBfOE12c21JcmV6MEdsTlk2MEdtZ3pNVWdmc2FlRmdqXzgxVUNWaTZHd3NFN0N4ZjNieUpPZzdJMHR1S254UjJOcloxcjdybF80MnFEZGFzN1BPYWxpdnBlZndWVWVKSHFFa3FqcEtjNWVOWEJsRjA3RWM3VjF5US1MemtUclE?oc=5

KAT-6 — AI Agent'lar

• AWS Strands Agents + Bedrock ile multi-agent social intelligence anlatti. Prospect discovery'den personalized email generation'a uzanan B2B agent pipeline'i production use case olarak netlesiyor. Kaynak: https://aws.amazon.com/blogs/machine-learning/multi-agent-social-intelligence-with-strands-agents-and-amazon-bedrock/
• AWS Nova Act ile agentic QA automation Part 2 yayinda. Batch regression, suite orchestration ve pipeline entegrasyonu QA agent'larini CI/CD hattina yaklastiriyor. Kaynak: https://aws.amazon.com/blogs/machine-learning/accelerating-software-delivery-with-agentic-qa-automation-using-amazon-nova-act-part-2/
• EvoCUA-1.5: Multi-turn computer-use agent'lar icin online RL yaklasimi. Kaynak: https://arxiv.org/abs/2607.09773
• StructAgent: Uzun ufuklu digital agent'lari causal structure ile harness etmeyi oneriyor. Kaynak: https://arxiv.org/abs/2607.11388

KAT-7 — Multimodal

• UNIBROWSE: Multimodal BrowseComp icin data-to-agent framework; browser agent benchmark hattini genisletiyor. Kaynak: https://arxiv.org/abs/2607.10557
• MM-ToolSandBox: Visual tool-calling agent'lari degerlendirmek icin unify edilmis sandbox. Kaynak: https://arxiv.org/abs/2607.11818
• TIGER: Multimodal speculative decoding icin text-conditioned visual gated routing ve acceptance alignment oneriyor. Kaynak: https://arxiv.org/abs/2607.11131
• GigaChat Audio: Audio-language model tarafinda zaman duyarliligini model mimarisine tasiyor. Kaynak: https://arxiv.org/abs/2607.10387

KAT-8 — Robotik & Embodied AI

• EmbodiedGen V2: Agentic, simulation-ready 3D world engine; embodied AI veri/ortam uretimi icin izlenmeli. Kaynak: https://arxiv.org/abs/2607.07459
• NeuralActuator: Robot dynamics ve external force perception icin neural actuation modeling. Kaynak: https://arxiv.org/abs/2607.11734
• Action Map Policy: Pixel classification ile 3D closed-loop manipulation ogrenen robot policy yaklasimi. Kaynak: https://arxiv.org/abs/2607.10706
• LoRA for VLA models in industrial robotic manipulation: Vision-language-action modellerinde verimli fine-tuning maliyetini inceliyor. Kaynak: https://arxiv.org/abs/2607.10172

KAT-9 — Edge, Cihaz & Veri Altyapisi

• Benchmarking Edge Inference Strategies: Industrial machine vision icin edge inference stratejilerini karsilastiriyor. Kaynak: https://arxiv.org/abs/2607.11356
• DynaFilter: Satellite edge intelligence icin cloud-driven dynamic filtering oneriyor. Kaynak: https://arxiv.org/abs/2607.10098
• Bonsai 27B 1-bit LLM telefon sinyali: Topluluk akisi 27B-class modelin telefon seviyesine inmesini one cikardi; local inference ilgisi yuksek. Kaynak: https://mastodon.social/@h4ckernews/116919727164531594
• Local LLM + GStreamer: llama.cpp ile GStreamer icinde local text translation/transformation pipeline'i paylasildi. Kaynak: https://floss.social/@centricular/116919423714126703

SECURITY

• GitHub code scanning AI security detections PR uzerinde gorunur oldu. AI tarafindan bulunan security bulgularinin review yuzeyine cikmasi, yazilim guvenliginde agent destekli triage'i hizlandirir. Kaynak: https://github.blog/changelog/2026-07-14-code-scanning-shows-ai-security-detections-on-pull-requests
• GitHub Copilot security reviews app icinde yayinda. Copilot'in security lens'i kod review icine gomuluyor. Kaynak: https://github.blog/changelog/2026-07-14-security-reviews-now-available-in-the-github-copilot-app
• MJ jailbreak benchmark: Multi-turn LLM jailbreaking'i decomposed credit assignment ile inceliyor. Kaynak: https://arxiv.org/abs/2607.11070
• ANCHOR: CLI agent'larda real-world harm icin automated alignment auditing. Kaynak: https://arxiv.org/abs/2607.10455

REGULATION

• Hindistan mahkemelerde AI kullanimi icin 2026 taslak reguelasyonu tartisiliyor. Judiciary adoption icin sorumlu AI cercevesi, legal AI'nin kamu standardi boyutunu gosteriyor. Kaynak: http://www.bing.com/news/apiclick.aspx?ref=FexRss&aid=&tid=6a56aaa6e20845aa90c99fbaee8777a8&url=https%3a%2f%2fconventuslaw.com%2freport%2findia-draft-regulations-for-use-of-artificial-intelligence-in-courts-2026-analysing-the-framework-for-responsible-ai-adoption-in-the-judiciary%2f&c=7475493240037752796&mkt=tr-tr
• California K-12'de insanlarin ogretmen kalmasini guvence altina alan AI yasasini gundeme tasidi. Egitimde AI destekleyici rol ile karar verici rol ayriliyor. Kaynak: https://news.google.com/rss/articles/CBMipgFBVV95cUxNaS1qMExCVGpFNGx6dm1FZWF4aHBjeEw3VWhmUUVTTldlc2ZUYk9jM0R2cnBTNU4zallQT1pCc1Q4RHVOU0swQjFPQnE4QlNiRnp6amZLRUZtczVGTWVWbElIMnZVR280RjZFRl9jRXlFSHA5Y2s3MmR0VHVzcUR2NlZrMVFUNmd6WEY4dzZNRjd6LTBrRnA1RHZXdGxiOF81TGNaS2VR?oc=5
• Almanya AI Overviews'u medya hukuku kapsaminda goruyor. Search sonucu ile editorial urun arasindaki sinir hukuken bulanıklaşıyor. Kaynak: https://news.google.com/rss/articles/CBMixAFBVV95cUxPRmhqd1N1RlV3RnBERVcxdlpaeDVnQnI2WlRiY2ZTRk95ZXhDNVNwbVJMcXljaVBmWWlMZE5YOUNkZC1tNWV2dF94NXFhZzQ4b0pydHYyX1NySjBvbVNkY3BmTXhqNEpmcmhmdzZSLXVONU5ValZHZWd1WEhHam95ZHltaU1uVFhzZmwzamM3UDdSYllkeVZackFDUl9VYTJHYWRDYVJnbExwTGFSVzlDZzV4ak5IZmNxYTlyQzhCRENHcVN4?oc=5
• Google'a AI training/copyright kaynakli yeni yayin davasi geldi. Telif riski frontier training ve AI Overviews tarafinda tekrar baski uretiyor. Kaynak: https://news.google.com/rss/articles/CBMingFBVV95cUxQN0hNdXJDSk42NHRpOEpISDVrSnYxa3ROU190NEhLVDMtSnJvSE44UjVJU2YtaTRpSU40WDViQk9fclp3V0JVaGxjbHdDc3N3SWJYNGFhNlhPZUc4UmVGTHQ5bVUtbGc0STFvOHNWVzI4ai1VbzBfd19oeG02SXFyVElDMVhZTEhZVlI1YlJkc1RnT0FGV0FDY2JCTl9QUQ?oc=5

AI-SCIENCE

• Drug discovery sermayesi guclu: Chai Discovery'nin 400 milyon dolarlik turu, AI-biology kesişiminin hala agir sermaye cektiğini gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMilwFBVV95cUxNeDlRODFkM0tGYnRSSUdHMGFBYU5YUGdzaWFEcHpDa2tuX0hTUFJ3dXhYVEVSRzBLN1lrbTlyN2M5bDY1Mm10VWtiN2hELVZlZTYtMzdocEVoVTUyUHhHTDR6eHc4ZWsyM2lpWnkzQXpmVDgwNkRjRThtSHZyWnE0Z1NxM2toRzd4SS12WEt4WU9YVWhSeUww?oc=5
• CLIR-Bench: Irregular clinical time series uzerinde multimodal QA benchmark'i, klinik AI'nin veri yapisi problemine odaklaniyor. Kaynak: https://arxiv.org/abs/2607.09880
• Faithful by Design: Klinik trial ozetlerinde LLM sadakatini multi-stakeholder kitleler icin degerlendiriyor. Kaynak: https://arxiv.org/abs/2607.09932
• Policy-Driven CT-Agent: Klinik olarak tutarli CT reasoning icin phase-aware diagnostic control modeli. Kaynak: https://arxiv.org/abs/2607.10748

INFRA

• SoftBank CEO'su AI talebi icin yillik 5 trilyon dolar yatirim gerektigini soyledi. Bu rakam, compute/enerji/datacentre sikismasinin model katmanindan buyuk olabilecegini gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMilwFBVV95cUxNZ1EtMWZOUHBraVd5eVpHMmRnQVhqak1OTmd0WTE5aGpuNGc3cmJtSmVCQWp2dXZKMjIwZ3RtYVNBaGNBZ0NPdW9zR0kySE10UXJJZVJPd2dHX1VtTjNCb3lEWmE4YUlkemFRRHh0THJnV19UNkFQZV9SU1pRSGt5QlZqOUI3QWNmUVpHZFo4X3ozVkkySTZj?oc=5
• Avustralya datacentre onaylarini hizlandirmayi vaat ediyor. AI yatirimi icin enerji/izin sureci rekabet unsuru haline geliyor. Kaynak: https://news.google.com/rss/articles/CBMi1gFBVV95cUxQMHFsVkpkbTV3bFllZWJxSENQNGU1QVQySnR1Mzc0alJTcWJPQV9VdTR2SUlRdVhoTGJrMEZ4YlVaSGdMb19YdDU4dU51MVpBTkY4aVlaWXdCT2d0LTN5NlZadzlKS1Y5ZzViamFHR3VLSzR1V3Jnbzd2VGpUUmtCdWlxdXJiSzJ3b3U2WHc1UHZrVEd0bE5hbTVJYTdwbkVrOFQ5VXF2RS0zeUVnbUNiV0FwaGZwTktsS2pXd3Bmd3cwTmpZVndLZFZCV3Ytbm15aG5Zb2VR?oc=5
• FlashTrie: GPU-accelerated constrained beam search ile generative retrieval'i hizlandirma hedefinde. Kaynak: https://arxiv.org/abs/2607.10044
• MemDecay: LLM agent inference icin region-aware KV cache eviction oneriyor; long-running agent maliyetinde izlenmeli. Kaynak: https://arxiv.org/abs/2607.10582

SAFETY

• HyperSafe: Fine-tuned modeller icin inference-time safety recovery yaklasimi. Kaynak: https://arxiv.org/abs/2607.11475
• Safe responses matter: MLLM'lerde over-refusal'i azaltan output-aware safety guardrail oneriyor. Kaynak: https://arxiv.org/abs/2607.09697
• Tool-MCoT: Content safety moderation icin tool-augmented multimodal chain-of-thought. Kaynak: https://arxiv.org/abs/2604.06205
• SPQR benchmark: Benign model adaptation sonrasi safety alignment risklerini olcmeye calisiyor. Kaynak: https://arxiv.org/abs/2511.19558

WATCHLIST

• Agent discovery standardlari: ARD + MCP + Genkit ayni anda ilerliyor; tool identity, provenance ve authorization standardi kritik olacak. Kaynak: https://www.infoq.com/news/2026/07/agentic-resource-discovery-spec/
• Agent portability reguelasyonu: "Bring your own AI agent" tasarisi platform ekonomisini sarsabilir. Kaynak: https://news.google.com/rss/articles/CBMiyAFBVV95cUxOOWtwTkM1ekxYSGFEOVpKdk5Xb1NpTV96TVJibVY3SFpWRXdseGRHSnBkbjdVWGFyUERBeG9GeGpld2E3d0tiTFN0UjRRZHdVNl9ITGhDak9adEVqZnN6d2FZWW5DMV9EV3JucHJkcVZQQzh5ZTBHU19pOERpTTFpNEZYMm13dGR4SlV5Yk83UzZpME40YXkwQVdFU25IN0lVNmZHYTRQWjJzMFpocldoai1lV2plWlMyWXlYMEFybFhxNFNhN2wweQ?oc=5
• AI layoffs/discrimination davalari: Meta hakkindaki AI-driven layoff iddialari HR automation riskini yukseltiyor. Kaynak: https://news.google.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?oc=5
• BCI + LLM arayuzu: Brain2Qwerty gibi isler insan-bilgisayar etkilesimini metin inputunun otesine tasiyor. Kaynak: https://www.infoq.com/news/2026/07/meta-brain-interface/

CikCik Paketi

• Simon Willison, GitHub Actions'ta uvx cache dostu kullanim tarifini paylasti. CI tool calistirma maliyeti ve determinism icin pratik sinyal. Kaynak: https://til.simonwillison.net/github-actions/uvx-github-actions-cache
• Frontier modellerin CI/CD gibi sessiz rolling update almasi tartisiliyor. Reproducibility ve model-version audit problemi tekrar gorunur oldu. Kaynak: https://mastodon.social/@threadverse/116920353776017454
• Richard Sutton'in Oak Lab girisimi sosyal akista one cikti. Startup'in hedefi operasyon sirasinda ogrenebilen agent'lar. Kaynak: https://social.heise.de/@heiseonlineenglish/116920245511533413
• Local LLM + llama.cpp + GStreamer entegrasyonu paylasildi. Media pipeline icinde local inference pratiklesiyor. Kaynak: https://floss.social/@centricular/116919423714126703
• Bonsai 27B 1-bit LLM telefon iddiasi viral oldu. 27B-class modeli telefon seviyesine indirme anlatısı edge AI ilgisini canli tutuyor. Kaynak: https://mastodon.social/@h4ckernews/116919727164531594
• Hamel Husain automated evals ve production readiness etrafinda sinyal verdi. Agent/LLM sistemlerinin ne zaman production-ready oldugu hala belirsiz bir kalite kapisi. Kaynak: https://news.google.com/rss/articles/CBMibEFVX3lxTE5PYi1PUVJYQXF2X1VsOGhMQnJENWdQVl9VQURLNkJBUmdwUnRTTGZSZFk2a3JTREVsMjN6OHFRTXhDYU42WXNQVnBPZ2Z4eFJRdmtfdl9hX1AtTDMtYlBickJrWXdGN3lqT2c1TA?oc=5

Oracle Self-Improvement Sinyalleri

• Dedupe state sisiyor. Ilk dedupe run 121K semantik kayit yuzunden 12 dakikayi asti; resmi cleanup sonrasi semantik depo 6.895 kayda indi ve dedupe tamamlandi. Kaynak: state/embeddings.db
• Collector coverage guclu. RSS/news, search, community, social ve academic/API ailelerinin tamami dolu geldi; missing family yok. Kaynak: /tmp/v3-prompt.txt
• Search akisi aggregator linklerine dayaniyor. Google/Bing haber linkleri canonical URL cozumlemesi gerektiriyor; bugun de kaynak kalitesi icin ana iyilestirme noktasi bu. Kaynak: /tmp/v3-prompt.txt
• Signal-only filtre calisiyor. 654 signal-only raw item ana rapor disinda tutuldu; temsilci URL kullanma politikasi korunmali. Kaynak: /tmp/v3-prompt.txt

Kaynak Ozeti

• Toplam akış: 8.404 raw item, 7.233 merged item, 6.159 unique item, 5.505 raporlanabilir item. Kaynak: /tmp/v3-prompt.txt
• Aile kapsami: rss/news 277 unique, search 421 unique, community 69 unique, social 3.349 unique, academic/API 1.389 unique item uretti. Kaynak: /tmp/v3-prompt.txt
• Baskin kaynaklar: social akista twitter_fallback yogun, academic tarafta arXiv cs.LG/cs.CV/cs.CL, search tarafinda google_news/security/ai/companies one cikti. Kaynak: /tmp/v3-prompt.txt
• Trend topikleri: Launches, Regulation, Models, Tooling ve Agents bugunku run'in en yuksek yogunluklu konu aileleri. Kaynak: /tmp/v3-prompt.txt

Coverage / Blind Spots

• Eksik kaynak ailesi yok. 5/5 kaynak ailesi calisti: rss/news, search, community, social, academic/API. Kaynak: /tmp/v3-prompt.txt
• Bos konu yok. Coverage summary'de empty_topics ve missing_families bos geldi; pipeline veri toplama acisindan tam kapsama verdi. Kaynak: /tmp/v3-prompt.txt
• Zayif nokta canonical URL. Search akisi Google News/Bing redirect URL'leriyle geliyor; yayin kalitesi icin canonicalizer hala gerekli. Kaynak: /tmp/v3-prompt.txt
• Sosyal kaynak baskin. 3.349 social unique item sinyal acisindan guclu ama noise yuksek; CikCik paketi icin temsilci secimi sart. Kaynak: /tmp/v3-prompt.txt

Bu Gece Sistem Ne Ogrendi

• Agent runtime artik tek kategori degil, yatay altyapi. ARD, Genkit, Bedrock/Nova Act, NVIDIA agent skills ve GitHub security review ayni gun icinde cikti; agent'lar tooling, security, infra ve reguelasyon kategorilerini ayni anda kesiyor. Kaynak: https://www.infoq.com/news/2026/07/agentic-resource-discovery-spec/ ve https://www.infoq.com/news/2026/07/genkit-agents-api-preview/
• Regulation ve launches tekrar eden ana tema. Onceki learning artifact "Launches, Regulation, Models" yogunlasmasini isaret ediyordu; bugun ayni tema AI Overviews, court AI, K-12, bring-your-own-agent ve model/tool duyurulariyla surdu. Kaynak: /tmp/v3-prompt.txt
• Memory/agent kalitesi icin audit katmani sart. MemDecay, AgentCheck, AuditWeave ve ToolAtlas gibi paper'lar agent'larda state, evidence ve tool-side memory'nin yeni kalite yuzeyi oldugunu gosteriyor. Kaynak: https://arxiv.org/abs/2607.10582 ve https://arxiv.org/abs/2607.11098 ve https://arxiv.org/abs/2607.09682 ve https://arxiv.org/abs/2607.11126
• Bir sonraki run icin ders: Dedupe cleanup periyodik hale getirilmeli; aksi halde semantic state buyumesi gece pipeline'ini yavaslatiyor. Kaynak: state/embeddings.db

Dedupe & Kalite Notu

Bu rapordaki maddeler bugunku unique.json uzerinden secildi ve onceki 3 gun raporlarinda one cikan ana haberler tekrar edilmedi. Dedupe sonucu: 7.233 merged giris -> 6.159 yeni/unique item; 1.003 fingerprint tekrar, 71 semantik tekrar elendi. Raporlanabilir set: 5.505 item; signal-only item'lar ana kaynak olarak kullanilmadi.