Oracle Gece Arastirma — 2026-07-01

Curate eden: Mahsum Aktaş · Günlük otomatik AI sektör taraması

Oracle Gece Arastirma — 2026-07-01

Otomatik derleme | v3 pipeline | 8.371 ham kaynak | 2.758 benzersiz item | 106 benzersiz kaynak | 5/5 kaynak ailesi kapsandi

Gunun Ozeti

Bugunun ekseni uc hatta toplandi: daha ucuz/agent odakli model katmani, bilim ve enterprise icin paketlenmis AI is akislari, agent guvenligi. Anthropic Claude Sonnet 5'i daha ucuz agent calistirma iddiasiyla cikardi ve AWS Bedrock tarafinda da duyurdu; ayni gun Claude Science, bilim insanlarina tek workbench sunan ayri bir urun olarak one cikti. Kaynak: https://techcrunch.com/2026/06/30/anthropic-launches-claude-sonnet-5-as-a-cheaper-way-to-run-agents/ ve https://aws.amazon.com/blogs/machine-learning/introducing-claude-sonnet-5-on-aws-anthropics-most-capable-sonnet-model/ ve https://www.anthropic.com/news/claude-science-ai-workbench

Ikinci ana hat, AI altyapisinin model duyurusundan "saha ekibi + chip + maliyet" modeline kaymasi: Amazon 1 milyar dolarlik forward-deployed engineering organizasyonu kurarken, Etched $5B degerleme ve $1B satis kontrati sinyaliyle inference donaniminda Nvidia disi ciddi bir cephe acti. Kaynak: https://techcrunch.com/2026/06/30/amazon-launches-new-1-billion-fde-org-following-openai-and-anthropic/ ve https://techcrunch.com/2026/06/30/nvidia-competitor-etched-hits-5b-valuation-1b-in-sales-for-ai-chip/

Ucuncu hat guvenlik: Microsoft poisoned MCP tool descriptions riskini, HN ise Claude Code request steganography tartismasini one cikardi; agent ekosistemi artik sadece capability degil, tool metadata, iz surme, policy ve bellek sızıntısı problemi. Kaynak: https://news.google.com/rss/articles/CBMifEFVX3lxTFA5LS1ZMExFM2I1R2FwWW1scGh4MGJNWGhvMG9yQ0diSUotZzEzTE4yTXZtQ0VvcWVEU1dod3NLQ2FaeVFhU2lhTC0tdVhhczlQRVFQakZ3X2NhY3lKVEJxdVFncW05MXYzNDhSZWRySW9jV29WZ1lBTDdFdm0?oc=5 ve https://thereallo.dev/blog/claude-code-prompt-steganography

Top 7

1. Claude Sonnet 5, agent calistirma maliyetini dusurme ve Opus seviyesine yaklasma iddiasiyla cikti; AWS Bedrock duyurusu enterprise dagitimini de ayni anda guclendirdi. Kaynak: https://techcrunch.com/2026/06/30/anthropic-launches-claude-sonnet-5-as-a-cheaper-way-to-run-agents/ ve https://aws.amazon.com/blogs/machine-learning/introducing-claude-sonnet-5-on-aws-anthropics-most-capable-sonnet-model/
2. Claude Science, bilimsel arastirma icin hazir skill setleri, computational workflow ve verification agent yaklasimi ile "model" degil "research operating system" iddiasi tasiyor. Kaynak: https://www.anthropic.com/news/claude-science-ai-workbench ve https://techcrunch.com/2026/06/30/anthropics-claude-science-bets-on-workflow-not-a-new-model-to-win-over-scientists/
3. OpenAI GeneBench-Pro, genomik ve biyoloji icin gercek dunya datasetleriyle bilimsel model performansini olcmeye odaklaniyor. Kaynak: https://openai.com/index/introducing-genebench-pro
4. Etched, inference chip pazarinda $5B degerleme ve $1B kontratla cikti; custom silicon savasi sadece Nvidia/Google/AMD degil, vertical startup katmanina indi. Kaynak: https://techcrunch.com/2026/06/30/nvidia-competitor-etched-hits-5b-valuation-1b-in-sales-for-ai-chip/
5. Amazon, $1B FDE organizasyonu ile musteri icine gomulen AI deployment modelini OpenAI/Anthropic cizgisine tasiyor. Kaynak: https://techcrunch.com/2026/06/30/amazon-launches-new-1-billion-fde-org-following-openai-and-anthropic/
6. Microsoft'un poisoned MCP tool description uyarisı, agent tool metadata'sinin yeni prompt injection yuzeyi oldugunu netlestirdi. Kaynak: https://news.google.com/rss/articles/CBMifEFVX3lxTFA5LS1ZMExFM2I1R2FwWW1scGh4MGJNWGhvMG9yQ0diSUotZzEzTE4yTXZtQ0VvcWVEU1dod3NLQ2FaeVFhU2lhTC0tdVhhczlQRVFQakZ3X2NhY3lKVEJxdVFncW05MXYzNDhSZWRySW9jV29WZ1lBTDdFdm0?oc=5
7. Agent memory, policy verifier ve context management paper yığını ayni sinyali verdi: long-horizon agent icin memory sadece recall degil, leak, confound ve policy compliance meselesi. Kaynak: https://arxiv.org/abs/2606.29788 ve https://arxiv.org/abs/2606.29225 ve https://arxiv.org/abs/2606.30005

Trend Analizi

• Spike: Sonnet, RAG, Perplexity ve Stability AI trend penceresinde sicrama verdi; bugunku Sonnet 5 ve Claude Science haberleri "Anthropic rising" sinyalini besledi. Kaynak: https://www.anthropic.com/news/claude-sonnet-5 ve https://www.anthropic.com/news/claude-science-ai-workbench
• Yukselis: AI Agents 973 toplam iz, Anthropic 521, Amazon 52; bu, model release haberinden ziyade deployment, FDE ve agent ops katmaninin buyudugunu gosteriyor. Kaynak: https://techcrunch.com/2026/06/30/amazon-launches-new-1-billion-fde-org-following-openai-and-anthropic/
• RAG 148 toplam ile halen sicak; bugunku arastirma tarafinda AB-RAG, PIXELRAG, Regime-Aware Peer Specialization ve Efficient RAG papers kaynak kalitesinin "retrieval budget" ve "context type" eksenine kaydigini gosteriyor. Kaynak: https://arxiv.org/abs/2606.29090 ve https://arxiv.org/abs/2606.28344 ve https://arxiv.org/abs/2606.30518

KAT-1 — Meta / Ekosistem

• AI startup'larinda "model + servis + saha ekibi" kombinasyonu hizlaniyor: Amazon FDE organizasyonu, OpenAI ve Anthropic'in enterprise deployment yaklasimini AWS icine kopyaliyor. Kaynak: https://www.cnbc.com/2026/06/30/aws-amazon-ai-forward-deployed-engineers.html
• Etched vakasi, inference donaniminda "genel GPU" yerine model odakli accelerator stratejisinin ticari alici buldugunu gosteriyor. Kaynak: https://techcrunch.com/2026/06/30/nvidia-competitor-etched-hits-5b-valuation-1b-in-sales-for-ai-chip/
• 8090 Labs'in $135M yatirimi, AI software development pazarinda coding-agent tooling'in hala sermaye cekebildigini gosteriyor. Kaynak: https://webrazzi.com/2026/06/30/yapay-zeka-destekli-yazilim-gelistirme-platformu-8090-135-milyon-dolar-yatirim-aldi/

KAT-2 — LLM & Model Guncellemeleri

• Claude Sonnet 5, Sonnet 4.6'ya gore coding/reasoning iyilestirmesi ve agent maliyeti mesajiyla geldi; Simon Willison tokenizer maliyet notunda Ingilizce token maliyetinin ~1.4x etkilenebilecegini isaret etti. Kaynak: https://www.anthropic.com/news/claude-sonnet-5 ve https://simonwillison.net/2026/Jun/30/claude-sonnet-5/
• Nemotron-Labs-Diffusion-Image, masked discrete diffusion ile high-resolution text-to-image synthesis tarafinda yeni bir NVIDIA arastirma hattı sundu. Kaynak: https://arxiv.org/abs/2606.29814
• Meituan LongCat-2.0, 1.6T toplam / ~48B active MoE anlatısıyla Cin chip ekosisteminin Nvidia bagimsiz model egitimi iddiasini guclendiriyor. Kaynak: https://the-decoder.com/meituans-longcat-2-0-shows-china-can-train-massive-ai-models-without-nvidia/

KAT-3 — Araclar & Framework'ler

• Simon Willison shot-scraper video ile coding agent'larin YAML storyboard'dan video demo kaydedebilmesini saglayan pratik bir automation araci duyurdu. Kaynak: https://simonwillison.net/2026/Jun/30/shot-scraper-video/
• Cursor, coding agent'larini iOS uzerinden baslatma/izleme/review etme akisini mobil beta ile genisletti. Kaynak: https://webrazzi.com/2026/06/30/cursor-yapay-zeka-destekli-kodlama-ajanlari-icin-ios-uygulamasi-yayinladi/
• Apify'nin AI agent tool sayisini 10x artirdigi haber akisi, browser/data extraction araclarinin agent runtime icin standart parca olmaya basladigini gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMiogFBVV95cUxQR1BNeGF6WnpYYXBIbFdjN1BheW5YV3RiSFJ6aldjUS1WSGQycFVyV01SeGkzYllkN0RTcXAxMTRoS1lGb0UxTm9xX1ZNYTZvOUJ2MmYyMmkwUjl5ZC1wa1N3VUp3QnZWdHllb3FOYVJaMk9qMTBQeTkxTmZlZGVGSFU5Y1B1UHA3emo4SU1iQTU3Y3ZvV1VPVkNETFVMekE2dXc?oc=5

KAT-4 — Arastirma & Paper'lar

• GeneBench-Pro, biology/genomics benchmark tarafinda "complex real-world datasets" vurgusu ile bilimsel AI degerlendirmesini tekil quiz benchmark'larindan uzaklastiriyor. Kaynak: https://openai.com/index/introducing-genebench-pro
• SEVA, fact attribution verifier'larinin self-evolving process reward ile iyilestirilmesini oneriyor; hallucination savunmasinda "son katman verifier" daha aktif hale geliyor. Kaynak: https://arxiv.org/abs/2606.29713
• Complexity Ceiling Benchmark, sequential reasoning derinligi arttikca model performansinin nasil kirildigini kontrollu sekilde olcmeye calisiyor. Kaynak: https://arxiv.org/abs/2606.29278

KAT-5 — Endustri & Sirketler

• Anthropic, Claude Science ile biopharma ve bilimsel arastirma pazarina dogrudan urunlestirilmis workflow satma yoluna girdi. Kaynak: https://techcrunch.com/2026/06/30/anthropics-claude-science-bets-on-workflow-not-a-new-model-to-win-over-scientists/
• Base44'un kendi modelini roll out etmeye baslamasi, vibe-coding platformlarinin sadece wrapper kalmamak icin model defensibility aradigini gosteriyor. Kaynak: https://techcrunch.com/2026/06/29/vibe-coding-platform-base44-launches-own-model-as-ai-startups-seek-defensibility/
• Globant-Anthropic "Claude-powered AI Pods" isbirligi, system integrator'larin model laboratuvarlariyla daha paketli teslimat modeli kurdugunu gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMi9gFBVV95cUxOdkRtbURfT0h1ZkQ4Tkxxb1hNZUtzUWNfUUE3WDdvdnlueWl5ZThsTVNoeU52N200V3loQmY4M0hEY0pzR1o1N0tjVXhpenVnNHlyZUI3WmMzMEdxN1VvWGg1ZkttUTdHR3EzbVdsWllOSE1GVmdYWDhMMW1zd2pxTEFMQnl6THdkeFZNUmRsc1IyUERuY0FnU2IxTlZBNUFmRzNLcVk0dTJSYzd1cWhLWWhGQUh0RzdTV3Z3aFBBNUxRSDMyRDZuMHlIU3k5QUJSV0oyMWhGTVNIdzRQYTlSWXRIN3A5T0VGaGpIZC1DMWRFUnFIZnc?oc=5

KAT-6 — AI Agent'lar

• Agent-as-a-Router, coding task'leri icin hangi modelin secilecegini agentic routing problemi olarak ele aliyor; bu, multi-provider maliyet/kalite orkestrasyonu icin dogrudan uygulanabilir. Kaynak: https://arxiv.org/abs/2606.22902
• LLM Agents Are Latent Context Managers, uzun ufuklu tool agent'larda context'in kendisinin yonetilmesi gereken bir runtime kaynagi oldugunu vurguluyor. Kaynak: https://arxiv.org/abs/2606.30005
• Scaling the Horizon, Not the Parameters, 35B agent'in horizon scaling ile daha buyuk model performansina yaklasabilecegini iddia ediyor; parametre yerine planlama ufku optimizasyonu one cikiyor. Kaynak: https://arxiv.org/abs/2606.30616

KAT-7 — Veri & Altyapi

• Core dump epidemiology yazisi, OpenAI'in buyuk olcekli core dump analiziyle hem hardware fault hem 18 yillik software bug buldugunu anlatiyor; AI altyapisi artik klasik reliability engineering'e daha fazla bagimli. Kaynak: https://openai.com/index/core-dump-epidemiology-data-infrastructure-bug
• Speculative Pre-Positioning, stateful session'larda bir sonraki karar noktasina off-critical-path decode yaparak vLLM/SGLang/TensorRT-LLM benzeri server'larda bos GPU zamanini azaltmayi hedefliyor. Kaynak: https://arxiv.org/abs/2606.29565
• One-Step Gradient Delay, asenkron pipeline parallel LLM pretraining'de GPU bubble kaybini azaltmak icin gecikmenin teorik bariyer olmadigini savunuyor. Kaynak: https://arxiv.org/abs/2606.30634

KAT-8 — Guvenlik & Alignment

• Microsoft'un MCP tool description poisoning uyarisı, agent'larin sadece user prompt'tan degil, tool metadata'sindan da etkilenebildigini gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMifEFVX3lxTFA5LS1ZMExFM2I1R2FwWW1scGh4MGJNWGhvMG9yQ0diSUotZzEzTE4yTXZtQ0VvcWVEU1dod3NLQ2FaeVFhU2lhTC0tdVhhczlQRVFQakZ3X2NhY3lKVEJxdVFncW05MXYzNDhSZWRySW9jV29WZ1lBTDdFdm0?oc=5
• 282 iOS AI app'inin API key/proxy access leak ettigi calisma, consumer AI app tedarik zincirinde client-side secret hijyeninin hala zayif oldugunu gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMihAFBVV95cUxQVE9NbWVPY3o5VDI0empPM0daa1lfbkctSHU2SHZhVmFJMU8waUFvZzVnc2UwQTFmVWRTbW1pZFFiSTVqUDNka0lVb0lpLUZ1QnlDTklSOHpjbENpZjRsa01iMk5sT1Y3TzVnX3QtQWF5Z0VVWWJOT0xiMjZlUGN0M25VQ3I?oc=5
• Claude Code request steganography tartismasi, agent request'lerinin izlenebilirlik/filigran/telemetry katmaninda opak riskler tasiyabilecegini gundeme tasidi. Kaynak: https://thereallo.dev/blog/claude-code-prompt-steganography

KAT-9 — Topluluk & Tartismalar

• HN'de Claude Sonnet 5, 681 puan ve 358 yorumla gunun ana teknik tartismasi oldu; benchmark kadar fiyat/tokenizer davranisi da konusuldu. Kaynak: https://www.anthropic.com/news/claude-sonnet-5
• HN'de Claude Science, bilimsel workflow urunlesmesi uzerinden "AI workbench mi, model mi" tartismasini tetikledi. Kaynak: https://claude.com/product/claude-science
• Reddit LocalLLaMA'da LongCat-2.0 ve DeepSeek V4 Flash GGUF basliklari local/open-weight inference tarafindaki ilgiyi canli tuttu. Kaynak: https://www.reddit.com/r/LocalLLaMA/comments/1uj7egu/introducing_longcat20_a_largescale_moe_language/ ve https://www.reddit.com/r/LocalLLaMA/comments/1ujlwbm/bartowski_has_delivered_ds4_gguf/

SECURITY

• Langflow RCE'nin Monero miner deployment icin kullanildigi haber akisi, AI app endpoint'lerinin klasik internet-exposed app risklerinden muaf olmadigini tekrar gosterdi. Kaynak: https://news.google.com/rss/articles/CBMihAFBVV95cUxNZ19oSkZ4UXFybk82Rm9xWW1DMHF0a0pJaWdwSGhpcXBDcVBld1MwWUxBaGh4dkh1TUN0enZzbkF2c1RGTDNWNG9EX0JmSndVZy1DM2w2RHk3N3g1V2o0YzlIbG40Tk1KUFY2Wm83OEV5ZzBka2Jpb0Z2SktOVmV0Q1kybF8?oc=5
• An Empirical Evaluation of Prompt Injection Vulnerabilities in LLMs, multilingual ve obfuscated attack senaryolarinda prompt injection riskini sistematik test ediyor. Kaynak: https://arxiv.org/abs/2606.29602
• SrDetection, Code LLM benchmark'larinda data leakage tespitini self-referential framework ile ele aliyor; coding benchmark skorlarina guven problemi buyuyor. Kaynak: https://arxiv.org/abs/2606.29815

REGULATION

• 35 news publisher'in OpenAI'a scraping iddiasiyla dava actigi haber, telif/regulasyon ekseninde veri kaynagi tartismasinin surdugunu gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMifkFVX3lxTE5GU1JGR0RBd0s3Tk9mekl0Y05ESDFqM3dmOEhuYTQwTi15NFRSZ0pORW9TcTMwbXdTeHlTU1JiWjRRWWlQNzVibmdwUnB5eHhTeVRuNkJUdmJtT05VU3FjSWtHWE1zSUU3SXlIY0tGRlY2U1R4bHhTa1FVVG9Xdw?oc=5
• CIA'nin AI cagina gore tech/acquisition ofislerini yeniden yapilandirdigi haber, kamu tedarik ve istihbarat operasyonlarinda AI'nin organizasyonel degisim tetikledigini gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMiiAFBVV95cUxPdEVwSUdjeUNEV2JjSndETFRBTnZSVV96SDI1ZUw1aUxRMms2Uy11a0ZVXzhLbFZBVHpGMTV6Z3NYNng0MUR3T1QwdzVtOGhoNXplMGZxeGNMODljTFZXeUttMnp4M3ZOc2RwRzJJN0RMQW8wbVctbUJ3T0tEblFqV2pISWJmV0F3?oc=5
• Modernizing Global Vulnerability Standards for the Age of AI haberi, CVE/CVSS benzeri guvenlik standartlarinin AI-agent risklerini kapsayacak sekilde baski altinda oldugunu gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMihgFBVV95cUxOWTlfcU5NNmZ1ZEhIbmhXcnk4RjlKRTA2R0Vad080Ui1ESWdiWUFyQUZrdkd4ZFZzcENiVVZRN2JvaUx4Yi04amVTUXZrTzVhOHoyWkpocmNVNVRsaV9OdGN0WGdMSWl3Y3pHeWtIY2ZWelc3NVEwdXRLNk5HV0lySDdhd3JFZw?oc=5

AI-SCIENCE

• Claude Science, genomics, computational chemistry ve research workflow'larini tek workbench altinda topluyor; verification agent vurgusu bilimsel hallucination riskine dogrudan cevap. Kaynak: https://www.anthropic.com/news/claude-science-ai-workbench
• OpenAI GeneBench-Pro, genomics/biology benchmark'ini daha zor gercek dunya veri setleriyle kuruyor. Kaynak: https://openai.com/index/introducing-genebench-pro
• LUMEN, systematic review/meta-analysis surecini cost-transparent multi-agent pipeline olarak ele aliyor; bilimsel literatur taramasi agent pazarinin sert use-case'lerinden biri olmaya devam ediyor. Kaynak: https://arxiv.org/abs/2606.28362

INFRA

• Etched'in inference chip cikisi, hardware roadmap'inde "model-specific accelerator" anlatısını guclendirdi. Kaynak: https://techcrunch.com/2026/06/30/nvidia-competitor-etched-hits-5b-valuation-1b-in-sales-for-ai-chip/
• Coverage-Driven KV Cache Eviction, long-context inference'ta bellek-verimlilik dengesini KV cache kararlarina indiriyor. Kaynak: https://arxiv.org/abs/2606.29563
• Predict, Reuse, and Repair, dynamic sparse attention'da top-K KV block secimini hizlandirmaya odaklaniyor. Kaynak: https://arxiv.org/abs/2606.30389

SAFETY

• CAREBench, child-safety risk escalation'ini explicit harm oncesinde yakalamaya calisan benchmark yaklasimi sunuyor. Kaynak: https://arxiv.org/abs/2606.29685
• EvalSafetyGap, benchmark/reward/safety metric iyilesirken gercek guvenligin geride kalabilecegi olcum boslugunu tarif ediyor. Kaynak: https://arxiv.org/abs/2606.30219
• MemLeak, multimodal agent memory'de "silindi" denen bilginin image/multimodal temsil uzerinden sizmaya devam edebilecegini gosteriyor. Kaynak: https://arxiv.org/abs/2606.29788

WATCHLIST

• MCP metadata poisoning: Tool description guven zinciri icin allowlist, signed manifests ve runtime policy gerekli hale geliyor. Kaynak: https://news.google.com/rss/articles/CBMifEFVX3lxTFA5LS1ZMExFM2I1R2FwWW1scGh4MGJNWGhvMG9yQ0diSUotZzEzTE4yTXZtQ0VvcWVEU1dod3NLQ2FaeVFhU2lhTC0tdVhhczlQRVFQakZ3X2NhY3lKVEJxdVFncW05MXYzNDhSZWRySW9jV29WZ1lBTDdFdm0?oc=5
• Agent memory confounds: MemDelta ve Reclaim Evaluation, memory benchmark'larinda yanlis/eksik ozete dayali guven hatasini tekrar gundeme aliyor. Kaynak: https://arxiv.org/abs/2606.29914 ve https://arxiv.org/abs/2606.25449
• Mobile AI secret leakage: iOS AI app API key leak calismasi, mobile-side proxy ve key storage denetimini watchlist'e aliyor. Kaynak: https://news.google.com/rss/articles/CBMihAFBVV95cUxQVE9NbWVPY3o5VDI0empPM0daa1lfbkctSHU2SHZhVmFJMU8waUFvZzVnc2UwQTFmVWRTbW1pZFFiSTVqUDNka0lVb0lpLUZ1QnlDTklSOHpjbENpZjRsa01iMk5sT1Y3TzVnX3QtQWF5Z0VVWWJOT0xiMjZlUGN0M25VQ3I?oc=5

TRENDS

• Launches ve Regulation bugunku en yogun iki topic; "urun cikariyoruz" ve "bunu nasil yonetecegiz" ayni anda buyuyor. Kaynak: https://techcrunch.com/2026/06/30/anthropic-launches-claude-sonnet-5-as-a-cheaper-way-to-run-agents/ ve https://news.google.com/rss/articles/CBMifkFVX3lxTE5GU1JGR0RBd0s3Tk9mekl0Y05ESDFqM3dmOEhuYTQwTi15NFRSZ0pORW9TcTMwbXdTeHlTU1JiWjRRWWlQNzVibmdwUnB5eHhTeVRuNkJUdmJtT05VU3FjSWtHWE1zSUU3SXlIY0tGRlY2U1R4bHhTa1FVVG9Xdw?oc=5
• Agent tooling trendi consumer surface'e indi: Cursor mobile, Acti keyboard agent ve shot-scraper video ayni gun farkli arayuzlerde agent kontrolunu gosterdi. Kaynak: https://webrazzi.com/2026/06/30/cursor-yapay-zeka-destekli-kodlama-ajanlari-icin-ios-uygulamasi-yayinladi/ ve https://techcrunch.com/2026/06/30/acti-puts-ai-agents-directly-into-your-smartphone-keyboard/ ve https://simonwillison.net/2026/Jun/30/shot-scraper-video/
• Scientific agents trendi: Claude Science, GeneBench-Pro ve LUMEN ayni yonde; AI bilimi sadece cevap uretmekten workflow + verification + evidence synthesis hattina kayiyor. Kaynak: https://www.anthropic.com/news/claude-science-ai-workbench ve https://openai.com/index/introducing-genebench-pro ve https://arxiv.org/abs/2606.28362

Multimodal

• DAIN, statik MoE yerine dinamik agent-based interaction network ile multimodal reasoning'i daha adaptif hale getirmeyi hedefliyor. Kaynak: https://arxiv.org/abs/2606.30189
• Latent Noise Mask, MLLM'lerde gereksiz gorsel token yogunlugunun fine-grained reasoning'i bozmasini azaltmaya odaklaniyor. Kaynak: https://arxiv.org/abs/2606.30168
• BrainJanus, brain-vision-language ekseninde anlama ve uretimi tek modelde birlestirmeye calisiyor. Kaynak: https://arxiv.org/abs/2606.30319

Robotik & Embodied AI

• Fast Enough to Act, low-latency robotic VLM/VLA icin spatio-temporal visual token merging oneriyor; robotikte inference hizinin task basarisina dogrudan etkisi var. Kaynak: https://arxiv.org/abs/2606.29350
• Early Warning Signals for OpenVLA Failure, VLA modellerinde visual distribution shift altinda hata oncesi sinyal yakalamayi hedefliyor. Kaynak: https://arxiv.org/abs/2606.29699
• Vision-Language-Action Models: Experimental Insights from a Real-World UR5 Platform, VLA modellerinin gercek robot platformuna transferini test ediyor. Kaynak: https://arxiv.org/abs/2606.30456

Edge & Cihaz

• JuZhou 1.0, China-developed AI accelerators uzerinde egitilen edge-native text-to-image foundation model olarak yerel accelerator ekosistemi sinyali veriyor. Kaynak: https://arxiv.org/abs/2606.28421
• Acti, AI agent'lari smartphone keyboard icine koyarak app-switching yerine input-layer agent deneyimi deniyor. Kaynak: https://techcrunch.com/2026/06/30/acti-puts-ai-agents-directly-into-your-smartphone-keyboard/
• HASTE, resource-constrained cihazlarda CNN compression'i training-free ve steerable hale getirmeye calisiyor. Kaynak: https://arxiv.org/abs/2606.30516

Acik Kaynak

• Claude Sonnet 5 GitHub Copilot'a genel kullanima geldi; coding workflow'unda model secimi artik IDE ve agent katmaninda hizli degisiyor. Kaynak: https://github.blog/changelog/2026-06-30-claude-sonnet-5-is-generally-available-for-github-copilot
• LAMP, Lean + MCP + proof repair ile formal proof agent'lari icin acik bir agentic framework tasarlıyor. Kaynak: https://arxiv.org/abs/2606.28841
• SimpleX Chat, GitHub trending'de gizlilik odakli messaging altyapisiyla one cikti; AI disi ama privacy-first altyapi sinyali guclu. Kaynak: https://github.com/simplex-chat/simplex-chat

CikCik Paketi

• Simon Willison, Claude Sonnet 5'in tokenizer degisimi nedeniyle Ingilizce/İspanyolca maliyet etkisini not etti; model fiyatina sadece list price degil tokenizer davranisi da giriyor. Kaynak: https://simonwillison.net/2026/Jun/30/claude-sonnet-5/
• Simon Willison, shot-scraper video ile agent'larin kendi demo videolarini uretebilmesini gosterdi; CikCik icin "agent output proof" formatina donusturulebilir. Kaynak: https://simonwillison.net/2026/Jun/30/shot-scraper-video/
• Karpathy aynasindan Etched'in stealth'ten cikisi ve $1B+ kontrat duyurusu yayıldi; chip narrative sosyalde teknik merak cekiyor. Kaynak: https://x.com/Etched
• Latent Space akisi "Sonnet 5 ships, X gets an MCP server, Codex gets buttons" paketini gundeme tasidi; X MCP konusu sosyal operasyonlar icin izlenmeli. Kaynak: https://mastodon.social/@Steenbergen_apps/116840729412655238
• HN'de Claude Code steganography yazisi yuksek ilgi gordu; CikCik icin "AI tool trust" temasinda guclu tartisma malzemesi. Kaynak: https://thereallo.dev/blog/claude-code-prompt-steganography

Rehberler & Kaynaklar

• Context Engineering for RAG, tek bir RAG cevabinin arkasindaki typed inputs modelini anlatiyor; internal RAG tasariminda yararli. Kaynak: https://towardsdatascience.com/context-engineering-for-rag-the-four-typed-inputs-behind-every-rag-answer/
• Stop Choosing Between Local and Cloud LLMs, Gemma 4 + GPT-5.4 ile hybrid local-cloud workflow rehberi sunuyor. Kaynak: https://towardsdatascience.com/stop-choosing-between-local-and-cloud-llms-a-field-guide-to-hybrid-patterns/
• HTML table extractor, tarayicidan kopyalanan rich text/HTML tablolarini donusturmek icin pratik bir utility. Kaynak: https://simonwillison.net/2026/Jun/29/html-table-extractor/

Oracle Self-Improvement Sinyalleri

• Agent memory paper yogunlugu, Oracle memory sisteminde "silindi/ozetlendi" bilgisinin dogruluk ve sızıntı testinden gecmesi gerektigini gosteriyor. Kaynak: https://arxiv.org/abs/2606.29788 ve https://arxiv.org/abs/2606.29914
• Context management paper'lari, uzun sessionlarda sadece compaction degil, agent'in kendi context dashboard sinyalini uretmesi gerektigini destekliyor. Kaynak: https://arxiv.org/abs/2606.30005
• PolicyGuard, sub-agent verifier modelini destekliyor: agent aksiyonlari yalnizca output kalitesiyle degil, system/policy uyumuyla da kontrol edilmeli. Kaynak: https://arxiv.org/abs/2606.29225

Kaynak Ozeti

• RSS/news: 2.106 ham, 232 unique; baskin kaynaklar DonanımHaber, The Register, Technopat. Kaynak: https://www.donanimhaber.com/ ve https://www.theregister.com/
• Search: 519 ham, 309 unique; Google News funding/AI/company akislari baskin, canonical URL cozumleme halen gerekli. Kaynak: https://news.google.com/
• Academic/API: 1.550 ham, 1.514 unique; arXiv cs.CV, cs.LG, cs.CL yogun. Kaynak: https://arxiv.org/
• Social: 4.131 ham, 656 unique; Twitter fallback gürültülü, Bluesky/Mastodon daha temiz sinyal verdi. Kaynak: https://bsky.app/ ve https://mastodon.social/
• Community: 65 ham, 47 unique; GitHub Trending, Lobsters ve Reddit ML ana sinyal verdi. Kaynak: https://github.com/trending ve https://lobste.rs/

Coverage / Blind Spots

• Coverage tam: 5/5 kaynak ailesi kapsandi, 10/10 topic kapsandi, missing family yok. Kaynak: https://arxiv.org/ ve https://news.google.com/ ve https://github.com/trending
• Thin family yok; ancak social collector X fallback tarafinda eski x.com / Posts aggregator sonucunu fazla topluyor, kalite skoru dusurulmeli. Kaynak: https://news.google.com/rss/articles/CBMiPEFVX3lxTFBHOE1tS1NSS05lUTBrZGcyTmxyYVd1Wl9hZ3htTnRwU1lJUVhEOWktUHJoNzJ1bmNnc2xDbA?oc=5
• Reddit tarafinda birkac kaynak 429 verdi; community coverage yine de GitHub/Lobsters/Reddit ML ile kapanmis durumda. Kaynak: https://www.reddit.com/r/MachineLearning/

Bu Gece Sistem Ne Ogrendi

• Kalici ders: "Launches + Regulation + Models" uclusu tekrar baskin; pipeline bu uc ekseni ayri skorlamali. Kaynak: https://techcrunch.com/2026/06/30/anthropic-launches-claude-sonnet-5-as-a-cheaper-way-to-run-agents/ ve https://news.google.com/
• Kalici ders: Search akisi aggregator linklerine dayaniyor; publish oncesi canonical URL resolver eklemek rapor kalitesini artirir. Kaynak: https://news.google.com/
• Kalici ders: Rising entities bugun Anthropic, AI Agents, RAG ve Amazon etrafinda toplandi; Oracle watchlist'e "agent runtime security", "scientific AI workbench" ve "FDE deployment model" eklenmeli. Kaynak: https://www.anthropic.com/news/claude-science-ai-workbench ve https://techcrunch.com/2026/06/30/amazon-launches-new-1-billion-fde-org-following-openai-and-anthropic/

Dedupe & Kalite Notu

Bu rapordaki ana maddeler onceki 3 gunun raporlarindaki tekrarlar elenerek secildi. Ozellikle DSpark, California-Claude, Azure-GB300 ve temiz repo/malware maddeleri tekrar ana haber yapilmadi. Toplam 8.371 item islendi; merge sonrasi 6.894 kayit, dedupe sonrasi 2.758 benzersiz item kaldı. Dedupe sonucu: 4.127 fingerprint tekrar, 8 semantik tekrar, 1 guncelleme.