Oracle Gece Arastirma — 2026-07-05

Curated by Mahsum Aktaş · Automated daily AI industry scan

Oracle Gece Arastirma — 2026-07-05

Otomatik derleme | v3 pipeline | 5006 ham kaynak | 963 benzersiz item | 79 benzersiz kaynak | 5/5 kaynak ailesi

Gunun Ozeti

Bugunun ana ekseni agent guvenligi ve kurumsal kontrol: Alibaba'nin Claude Code'u yuksek riskli yazilim diye yasaklamasi, JadePuffer saldirisinin AI agent ile otomasyonu ve MCP/tool aciklamalarinin veri sizdirma riski ayni yonde sinyal veriyor. Kaynak: https://techcrunch.com/2026/07/04/alibaba-reportedly-bans-employees-from-using-claude-code/ ve https://www.bleepingcomputer.com/news/security/jadepuffer-ransomware-used-ai-agent-to-automate-entire-attack/ ve https://news.google.com/rss/articles/CBMifEFVX3lxTFA5LS1ZMExFM2I1R2FwWW1scGh4MGJNWGhvMG9yQ0diSUotZzEzTE4yTXZtQ0VvcWVEU1dod3NLQ2FaeVFhU2lhTC0tdVhhczlQRVFQakZ3X2NhY3lKVEJxdVFncW05MXYzNDhSZWRySW9jV29WZ1lBTDdFdm0?oc=5

Ikinci hat acik model ve formal reasoning: Mistral Leanstral 1.5, GLM-5.2, Qwen AgentWorld ve Google TabFM sinyalleri "genel chatbot"tan domain-specific modeller, agent environment simulation ve tabular foundation model'lara kayiyor. Kaynak: https://mistral.ai/news/leanstral-1-5 ve https://huggingface.co/zai-org/GLM-5.2 ve https://huggingface.co/Qwen/Qwen-AgentWorld-35B-A3B ve https://www.reddit.com/r/LocalLLaMA/comments/1un5hyi/googletabfm100/

Ucuncu hat AI-science: Anthropic'in Claude Science Beta ve ilac kesif programlari, AI'nin sadece yazilim uretiminde degil bilimsel is akisi orkestrasyonunda da urunlestigini gosteriyor. Kaynak: https://www.marktechpost.com/2026/07/04/anthropic-launches-claude-science-beta/ ve https://the-decoder.com/anthropic-launches-its-own-drug-discovery-programs-to-tackle-diseases-big-pharma-considers-unprofitable/

TRENDS

• YUKSELIS: Databricks, ElevenLabs ve Midjourney trend penceresinde yukseliyor; bugunku yeni Midjourney sinyali model duyurusu degil, Hollywood davalarinda karsi tarafin AI kullanim detaylarini istemesi. Kaynak: https://techcrunch.com/2026/07/04/midjourney-wants-hollywood-studios-to-reveal-the-details-of-their-ai-usage/
• YUKSELIS: Models, launches ve agents bugunku ilk uc topic; model tarafinda Leanstral, GLM-5.2 ve Qwen AgentWorld, agent tarafinda HORIZON ve Claude Science Beta one cikti. Kaynak: https://mistral.ai/news/leanstral-1-5 ve https://huggingface.co/zai-org/GLM-5.2 ve https://huggingface.co/Qwen/Qwen-AgentWorld-35B-A3B ve https://www.marktechpost.com/2026/07/04/nvidia-horizon-a-hands-free-agent-that-evolves-git-worktrees-and-hits-100-rtl-benchmark-completion/
• AZALIS: Gemini, Gemma ve Grok hacmi yuksek kaldi ama bugunku yeni karar sinyali tasimadi; daha taze aksiyon token maliyeti, routing ve enterprise policy tarafinda. Kaynak: https://the-decoder.com/open-source-tool-pxpipe-hides-text-in-pngs-to-cut-claude-code-and-fable-5-token-costs-up-to-70/

Top 7

1. Alibaba, Claude Code'u guvenlik riski olarak siniflayip calisanlara yasakladi; bu, coding-agent adoption'inda "tool guveni"nin procurement kapisina takilacagini gosteriyor. Kaynak: https://techcrunch.com/2026/07/04/alibaba-reportedly-bans-employees-from-using-claude-code/
2. JadePuffer ransomware vakasi, LLM agent ile bastan sona otomatik saldiri anlatimi verdi; agent sandbox, secret izolasyonu ve egress kontrolu artik opsiyon degil. Kaynak: https://www.bleepingcomputer.com/news/security/jadepuffer-ransomware-used-ai-agent-to-automate-entire-attack/
3. Mistral Leanstral 1.5, formal math/Lean 4 odakli acik model hattini guclendirdi; kod agent'lari icin "proof-aware" uzman model fikri olgunlasiyor. Kaynak: https://mistral.ai/news/leanstral-1-5
4. Anthropic Claude Science Beta, genomics/proteomics/cheminformatics is akisini multi-agent workbench olarak urunlestiriyor; bilimsel otomasyon yeni rekabet alani. Kaynak: https://www.marktechpost.com/2026/07/04/anthropic-launches-claude-science-beta/
5. pxpipe/ReadBench, uzun dokumani PNG'e cevirip tokenizasyonu atlayarak Claude Code ve Fable maliyetini dusurmeyi deniyor; bu, multimodal context ekonomisinde yeni bir taktik. Kaynak: https://the-decoder.com/open-source-tool-pxpipe-hides-text-in-pngs-to-cut-claude-code-and-fable-5-token-costs-up-to-70/ ve https://github.com/teamchong/pxpipe
6. Midjourney, Hollywood davalarinda stüdyolarin kendi AI kullanimini aciklamasini istiyor; generative AI hukukunda savunma hattı "siz de kullaniyorsunuz" eksenine kayiyor. Kaynak: https://techcrunch.com/2026/07/04/midjourney-wants-hollywood-studios-to-reveal-the-details-of-their-ai-usage/
7. NVIDIA HORIZON, git worktree'lerini evrimlestiren hands-free agent olarak RTL benchmark'ta tamamlama iddiasi tasiyor; self-improving coding loop'lar pratiklesiyor. Kaynak: https://www.marktechpost.com/2026/07/04/nvidia-horizon-a-hands-free-agent-that-evolves-git-worktrees-and-hits-100-rtl-benchmark-completion/

KAT-1 — LLM & Model Guncellemeleri

• Mistral Leanstral 1.5, Lean 4/formal reasoning icin Apache-2.0 hattini guclendiriyor; genel kod modelinden cok matematiksel dogrulama uzmani gibi konumlanmis. Kaynak: https://mistral.ai/news/leanstral-1-5
• GLM-5.2 Hugging Face trendlerinde yukarida; MIT lisansli, cok dilli ve MoE/DSA mimari etiketleriyle acik model ekosisteminde Cin hattinin aktif kaldigini gosteriyor. Kaynak: https://huggingface.co/zai-org/GLM-5.2
• Qwen AgentWorld-35B-A3B, agent/environment simulation etiketiyle "model = simulator + planner" yonunde konumlanmis. Kaynak: https://huggingface.co/Qwen/Qwen-AgentWorld-35B-A3B

KAT-2 — Arastirma & Paper'lar

• ReadBench/pxpipe, uzun-context belgeleri imaj olarak temsil edip VLM'lere okutma fikrini test ediyor; token maliyeti ve retrieval tasarimi icin izlenmeli. Kaynak: https://github.com/teamchong/pxpipe
• Google TabFM 1.0.0, tabular foundation model fikrini LocalLLaMA toplulugunda tekrar gundeme getirdi; operasyonel veri setleri icin zero-shot tabular modeling sinyali. Kaynak: https://www.reddit.com/r/LocalLLaMA/comments/1un5hyi/googletabfm100/
• AI ogrenme maliyeti calismasi, 26 bin ogrencilik veriyle AI'ye fazla yaslanmanin olumsuz etkisinin iki yilda gorundugunu savunuyor; egitimde "kisa vadeli verim, uzun vadeli yetenek kaybi" riski. Kaynak: https://the-decoder.com/a-26000-student-study-shows-ais-hidden-learning-cost-takes-two-full-years-to-surface/

KAT-3 — Araclar & Framework'ler

• Strix, acik kaynak AI pentesting araci olarak GitHub trending'e girdi; agent'larin hem savunma hem saldiri otomasyonunda cift kullanima kaydigini gosteriyor. Kaynak: https://github.com/usestrix/strix
• herdr, terminalde agent multiplexer yaklasimi sunuyor; coklu coding-agent calismalarinda kontrol paneli ihtiyaci buyuyor. Kaynak: https://github.com/ogulcancelik/herdr
• MailKite, email inbox'u agent iletisim arayuzu yapma fikrini one cikardi; agent workflow'larinda klasik mailbox yeniden control plane olabilir. Kaynak: https://mastodon.social/@beyondthecode/116863124346318369

KAT-4 — Acik Kaynak

• Simon Willison'in Open Source AI Gap Map notu, kamu yararina acik AI secenegi insa etme bosluklarini derliyor; acik ekosistemde "model var, governance/infra eksik" cizgisi belirgin. Kaynak: https://simonwillison.net/2026/Jul/3/open-source-ai-gap-map/#atom-everything
• CubeSandbox, AI agent'lar icin concurrent, guvenli ve hafif sandbox sunuyor; local/enterprise agent runtime icin pratik bir aday. Kaynak: https://github.com/TencentCloud/CubeSandbox
• Astryx, "agent-ready" tasarim sistemi iddiasiyla GitHub trending'de; UI sistemleri bile artik insan+agent ortak uretim akisini hedefliyor. Kaynak: https://github.com/facebook/astryx

KAT-5 — Endustri & Sirketler

• Alibaba'nin Claude Code yasagi, agent araclarinda veri cikisi, kod tabani erisimi ve vendor policy risklerini CISO seviyesine tasiyor. Kaynak: https://techcrunch.com/2026/07/04/alibaba-reportedly-bans-employees-from-using-claude-code/
• CNBC, Macron'dan Modi'ye hukumetlerin AI devlerine altyapi/sermaye alani actigini yazdi; ulkeler model sirketlerini ekonomik kalkinma araci gibi konumluyor. Kaynak: https://www.cnbc.com/2026/07/04/macron-modi-ai-infrastructure-tech-ceos.html
• The Verge'in fanfiction dosyasi, creator topluluklarinda AI detector, telif ve kultur savasinin urunlerden once normlari belirledigini gosteriyor. Kaynak: https://www.theverge.com/tech/960854/ai-fanfiction-ao3-claude-detector

KAT-6 — AI Agent'lar

• NVIDIA HORIZON, git worktree'leriyle self-improving engineering loop kuruyor; agent'in sadece patch yazmasi degil, deneme uzayini yonetmesi one cikiyor. Kaynak: https://www.marktechpost.com/2026/07/04/nvidia-horizon-a-hands-free-agent-that-evolves-git-worktrees-and-hits-100-rtl-benchmark-completion/
• Anthropic Claude Science Beta, bilimsel pipeline'lari multi-agent workbench ile birlestiriyor; agent'lar domain lab asistanina donusuyor. Kaynak: https://www.marktechpost.com/2026/07/04/anthropic-launches-claude-science-beta/
• LocalLLaMA'daki RP/agentic benchmark tartismasi, local modellerde "roleplay + tool use + long context" kombinasyonunun gayriresmi ama canli bir benchmark alanina donustugunu gosteriyor. Kaynak: https://www.reddit.com/r/LocalLLaMA/comments/1unbm45/ran_a_classicmedival_europe_fantasy_rpagentic/

KAT-7 — Multimodal

• pxpipe, metni PNG'e gomerek VLM okuma hattini kullaniyor; multimodal sadece gorsel uretim degil, token ekonomisi ve context tasarrufu araci. Kaynak: https://github.com/teamchong/pxpipe
• baidu/Unlimited-OCR, vision-language OCR hattinda trend oldu; belge isleme ve cok dilli OCR agent'lari icin izlenmeli. Kaynak: https://huggingface.co/baidu/Unlimited-OCR
• Krea-2-Turbo, text-to-image tarafinda hiz/kalite ekseninde Hugging Face trendlerinde; kreatif uretim pazarinda model cesitliligi artiyor. Kaynak: https://huggingface.co/krea/Krea-2-Turbo

KAT-8 — Edge & Cihaz

• Huawei Kirin 2026 paper iddiasi, LogicFolding ile %41 dusuk guc ve %37,5 daha kucuk die alani sinyali verdi; mobile AI silicon rekabeti sertlesiyor. Kaynak: https://x.com/poezhao0605/status/2073300079923069104
• Samsung Exynos 2700 tasarim degisikligi, bellek/sicaklik dengesini hedefliyor; edge AI performansinda sustained thermal envelope kritik. Kaynak: https://www.donanimhaber.com/samsung-exynos-2700-de-apple-in-izinden-gidiyor--207663
• Intel 18A uretiminde kritik esigin asilmasi iddiasi, AI compute tedarik zincirinde Foundry tarafinin tekrar izlenmesi gerektigini gosteriyor. Kaynak: https://www.technopat.net/2026/07/04/intel-18a-uretiminde-kritik-esik-asildi/

KAT-9 — Veri & Altyapi

• Cloudflare Town Lake/Skipper anlatimi, billing, security ve operasyonel datayi tek analytics platformunda topluyor; AI analytics agent icin veri platformu once geliyor. Kaynak: https://www.infoq.com/news/2026/07/cloudflare-unified-data-platform/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=AI%2C+ML+%26+Data+Engineering
• Hardwood, JVM tarafinda Parquet isleme icin sifir mandatory dependency ve multi-threaded yol vadediyor; AI veri boru hatlarinda Java ekosistemi icin pratik bir sinyal. Kaynak: https://www.infoq.com/news/2026/07/hardwood-java-parquet/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=AI%2C+ML+%26+Data+Engineering
• Elasticsearch agent memory yazisi, hybrid retrieval ve document-level security ile agent memory'nin enterprise search tarafina baglanabilecegini gosteriyor. Kaynak: https://www.elastic.co/search-labs/blog/agent-memory-elasticsearch

SECURITY

• JadePuffer vakasi, LLM agent'in ransomware zincirini otomatiklestirebildigini iddia ediyor; detection sadece imza degil, agent davranis telemetrisi de istemeli. Kaynak: https://www.bleepingcomputer.com/news/security/jadepuffer-ransomware-used-ai-agent-to-automate-entire-attack/
• Microsoft'un MCP tool description uyarisi, tool metadata'sinin prompt injection yuzeyi oldugunu gosteriyor; agent runtime'da tool schema review zorunlu olmali. Kaynak: https://news.google.com/rss/articles/CBMifEFVX3lxTFA5LS1ZMExFM2I1R2FwWW1scGh4MGJNWGhvMG9yQ0diSUotZzEzTE4yTXZtQ0VvcWVEU1dod3NLQ2FaeVFhU2lhTC0tdVhhczlQRVFQakZ3X2NhY3lKVEJxdVFncW05MXYzNDhSZWRySW9jV29WZ1lBTDdFdm0?oc=5
• Clean GitHub repo'nun AI coding agent'i malware calistirmaya kandirmasi, repo trust modelinin agent caginda yetersiz kaldigini gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMifEFVX3lxTE5SSGU1Y3pjd3BjRGllY3NDT2x5ZWI3MEpza1A5UGFRVDBNUEE2OEZ4RDZfc2RJeVhJOUNZUEVEb2t0d1RzVzFmUkhwTnhRZ29vMHVuT1RwS0ZzYXFFTXhlVlk2TE1PZ0kxNWdpc3RtY1U4VU9YTDdoNWdZbTU?oc=5

REGULATION

• Midjourney-Hollywood hattinda discovery talebi, AI telif davalarinda kullanici/uretici ayrimini bulaniklastiriyor; stüdyolarin kendi AI kullanimi hukuki argumana donusuyor. Kaynak: https://techcrunch.com/2026/07/04/midjourney-wants-hollywood-studios-to-reveal-the-details-of-their-ai-usage/
• Hukumetlerin AI devlerine altyapi acmasi, regülasyonun sadece fren degil sanayi politikasi oldugunu gosteriyor. Kaynak: https://www.cnbc.com/2026/07/04/macron-modi-ai-infrastructure-tech-ceos.html
• ARTCI'nin dijital governance ve AI forumu, Frankofon regulator tarafinda AI yonetisiminin yayildigini gosteriyor; Batı disi regülasyon sinyalleri takip edilmeli. Kaynak: https://news.google.com/rss/articles/CBMi2gFBVV95cUxPV0dVLVBYWUNNTlF0b29uMmtBbEZxVXBDR1AtblhDd0p3eXJlUlB2bnhtd3dSb0tRVkF3elg5X1E5bEZYMHdsY0lGT0ZqN1QtUDFLbkVLZTNKbWw4VFZKTi01VW53SFM0NmJzUDlZVGxtRWFZNWVHNXF5Z3VINUdERTZ4QmdoR1RLQmUxSGFtcXk3RE1PWWhQVGtCSTNNeHhKRVpiZWpZY085dGJVX1lSV0xHQXdoWl9QZ1JodGxLVjkxU0hQMi1uRmNnY2FVdkVUNnJQUnhYY24wdw?oc=5

AI-SCIENCE

• Claude Science Beta, genomics/proteomics/cheminformatics icin agentic workbench sunuyor; lab otomasyonu ile coding-agent paradigmasi yakinlasiyor. Kaynak: https://www.marktechpost.com/2026/07/04/anthropic-launches-claude-science-beta/
• Anthropic'in neglected disease programlari, frontier lab'lerin sadece model platformu degil, bilimsel portfoy sahibi kurumlara donusebilecegini gosteriyor. Kaynak: https://the-decoder.com/anthropic-launches-its-own-drug-discovery-programs-to-tackle-diseases-big-pharma-considers-unprofitable/
• Life sciences paperwork otomasyonu icin Collate yatirimi, bilimsel operasyonlarda AI'nin dokuman ve compliance katmanindan basladigini gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMixAFBVV95cUxQM1VjcnJEaWVLSFdVbG95VHE4SVRfbTByVU1sUHZldWNMbDFHR05tbDJ2LWlIUGd3SFV6eDNqWDUwa3pGMWYwLXhaMVlTMlZUN1Z6VnBEZ3FsM1Z1YkI5WVc2bXphUnhTQlU3UVZwZkItMTJPQzhTN1pIeXFxcGdDMDhnVDdYMWR6MTdIN3dGNmNwdFJORGE1Y0ZfSWxVQmZSWURtVkJFYWkxNC1ZVUhBZlFQNzZXbUxmQTNOdV9qNWg3azh2?oc=5

INFRA

• Cloudflare unified data platform dosyasi, AI analytics agent'in arkasinda 91K+ query/s gibi operasyonel olceklerin oldugunu hatirlatiyor. Kaynak: https://www.infoq.com/news/2026/07/cloudflare-unified-data-platform/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=AI%2C+ML+%26+Data+Engineering
• Seltz'in AI agent search icin seed yatirimi, klasik web search'in agent query pattern'lerine gore yeniden tasarlanacagini gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMiyAFBVV95cUxQNFJMUVptQk9VejhvaDJxY3hoUEZ0YUJ1ZWY3X1N1b3g0Q3RGZk5DNjdadUd6TGE0NWh5WXQ0Q2hIVVo2QUVxNVlHdkFqb2VKdjNEVzJQMGJCOUh4ZWNraWhpcEpNU19LWjFwTldJQXF1a1Y4SHdZNHd6S2RYZGZXck1BU3lXMlZZSjY0UmVDcEJSLUxCVDdSTTBpYWlScDNXekhrUXhCQkZKQXkweTI0aGRHTnFwdUhnMnNJWGtzRHFfVDhHQmlmYQ?oc=5
• Samsung RAM fiyat sinyali, AI memory supercycle'in consumer ve enterprise donanim maliyetine yayildigini gosteriyor. Kaynak: https://www.technopat.net/2026/07/04/samsung-ram-fiyatlarina-yeni-zam/

SAFETY

• Anthropic Responsible Scaling Policy guncellemesi, frontier model governance'in daha ayrintili risk esikleriyle kurumsallasmasi anlamina geliyor. Kaynak: https://www.anthropic.com/news/announcing-our-updated-responsible-scaling-policy
• LocalLLaMA'daki "literal prompt injection by Anthropic" tartismasi, kullanici topluluklarinin model davranis degisimlerini guven sorunu olarak okudugunu gosteriyor. Kaynak: https://www.reddit.com/r/LocalLLaMA/comments/1unif51/possible_evidence_of_literal_prompt_injection_by/
• Fanfiction topluluklarindaki AI detector kavgası, safety'nin sadece zarar azaltma degil kultur, emek ve guven normlariyla da ilgili oldugunu gosteriyor. Kaynak: https://www.theverge.com/tech/960854/ai-fanfiction-ao3-claude-detector

WATCHLIST

• Meta Watermelon AI'nin GPT-5.5 parity iddiasi dogrulanmamis benchmark basliklariyla geliyor; kaynak zayif, dogrulama beklenmeli. Kaynak: https://news.google.com/rss/articles/CBMiywFBVV95cUxQR04wMmJZaGVodXlKYTZWTjlRRkpkUmVxS0dPTjh4WkJ6WUdKakdoTXphZTV0SWh5U0FBSUtxSm45cE9ieDl0NHBVVXN6UEZWZ0o1eFFxQjNIekMxVGJLYmtJMlpHUnJLRF9wdC1qOWN3QkpZeTRMekZrZm1Kb2hxclUyRi1SNk5WMGhESE8tZFNIYzYyTDdzYlpxdnE0UzVFTkYtRTlheFl5SzJTTnFZaGZrVnkteGF4SkFERVlXTFU0Z0d1Q1hROWdndw?oc=5
• Claude/Fable routing tartismalari, kullanicilarin silent backend model routing'e hassas oldugunu gosteriyor; model secimi UI'da seffaf olmali. Kaynak: https://x.com/giffmana/status/2073342166135369831/photo/1
• AI data center isi ve mahalle sicakligi tartismalari, altyapi buyumesinin public health/regulation tarafina tasabilecegini gosteriyor. Kaynak: https://rbfirehose.com/2026/07/04/colorado-sun-do-data-centers-make-nearby-neighborhoods-hotter/

CikCik Paketi

• Aci: "Agent guvenligi artik teorik degil: JadePuffer vakasi LLM agent ile ransomware zincirini otomatiklestirdi." Kaynak: https://www.bleepingcomputer.com/news/security/jadepuffer-ransomware-used-ai-agent-to-automate-entire-attack/
• Aci: "Alibaba'nin Claude Code yasagi, enterprise AI'da asil savasin model kalitesi degil tool trust oldugunu gosteriyor." Kaynak: https://techcrunch.com/2026/07/04/alibaba-reportedly-bans-employees-from-using-claude-code/
• Aci: "pxpipe ilginc: context'i metin degil gorsel olarak okutmak token ekonomisini tersine cevirebilir." Kaynak: https://github.com/teamchong/pxpipe
• Aci: "Simon Willison'in Fable yorumu net: hunch degil data. Agent benchmark'lari tekrar kosulmali." Kaynak: https://bsky.app/profile/simonwillison.net/post/3mpt7gnenkc2a
• Aci: "Karpathy'nin playable worlds yorumu, LLM+code+knowledge birlesiminin oyun/prototype tarafinda nereye gidecegini gosteriyor." Kaynak: https://bsky.app/profile/karpathy-mirr.selfhosted.social/post/3mptxv3px6p2l
• Aci: "Clem Delangue'in acik AI milestone vurgusu, acik kaynak tarafinin sadece model degil tarih ve narratif de insa ettigini gosteriyor." Kaynak: https://news.google.com/rss/articles/CBMiZ0FVX3lxTE1kUkY2Uk5mTTFhUlg0Zy1wVlhRX2JtZUhNa191ZzF4eHdxcjBWU25ieWgzX2drV3l4M0FsS041Q1Z2S2ducGpLRGozdEgxNlBGdzR2eGJtYnF6R2VKbFVzX1RMVTQ1WHc?oc=5

Rehberler & Kaynaklar

• Anthropic developer prompting ipuclari, Fable 5 icin once kendi blind spot'unu bulma prensibini vurguluyor; prompt review checklist'e cevrilebilir. Kaynak: https://the-decoder.com/anthropic-developer-shares-prompting-tips-for-fable-5-that-focus-on-finding-your-own-blind-spots-first/
• Simon Willison Open Source AI Gap Map, acik AI stratejisi icin okunacak kaynak. Kaynak: https://simonwillison.net/2026/Jul/3/open-source-ai-gap-map/#atom-everything
• PyTorch test infrastructure yazisi, import-time generated tests ve local debug pratikleri icin yararli. Kaynak: https://pytorch.org/blog/understanding-pytorchs-test-infrastructure/

Oracle Self-Improvement Sinyalleri

• Agent tool guveni icin policy: yeni tool/connector eklenmeden once tool description prompt-injection scan ve egress risk kontrolu zorunlu olmali. Dayanak: https://news.google.com/rss/articles/CBMifEFVX3lxTFA5LS1ZMExFM2I1R2FwWW1scGh4MGJNWGhvMG9yQ0diSUotZzEzTE4yTXZtQ0VvcWVEU1dod3NLQ2FaeVFhU2lhTC0tdVhhczlQRVFQakZ3X2NhY3lKVEJxdVFncW05MXYzNDhSZWRySW9jV29WZ1lBTDdFdm0?oc=5
• Pipeline icin ders: social collector 4095 ham kayittan 587 unique uretti; X fallback yuksek gürültü uretiyor, canonical URL cozumleme ve eski tarih filtresi sertlestirilmeli. Dayanak: https://news.google.com/rss/articles/CBMiO0FVX3lxTFBMR2tuN0FwY2NyZFdTTnhJaHphUzhQRm5Xby04MVlmc3l0YXM5LW5MQTI5aWVQclBGRmww?oc=5
• Raporlama icin ders: Google News aggregator linkleri fazla; publish oncesi canonical resolver eklemek kaynak kalitesini yukseltir. Dayanak: https://news.google.com/rss/articles/CBMiyAFBVV95cUxQNFJMUVptQk9VejhvaDJxY3hoUEZ0YUJ1ZWY3X1N1b3g0Q3RGZk5DNjdadUd6TGE0NWh5WXQ0Q2hIVVo2QUVxNVlHdkFqb2VKdjNEVzJQMGJCOUh4ZWNraWhpcEpNU19LWjFwTldJQXF1a1Y4SHdZNHd6S2RYZGZXck1BU3lXMlZZSjY0UmVDcEJSLUxCVDdSTTBpYWlScDNXekhrUXhCQkZKQXkweTI0aGRHTnFwdUhnMnNJWGtzRHFfVDhHQmlmYQ?oc=5

Kaynak Ozeti

• Kapsam: 5/5 kaynak ailesi calisti; RSS/news 250 ham, search 522 ham, community 69 ham, social 4095 ham, academic/API 70 ham. Kaynak: https://techcrunch.com/2026/07/04/alibaba-reportedly-bans-employees-from-using-claude-code/
• Benzersiz sonuc: 5006 ham kayit merge edildi, 963 yeni kayit rapor havuzuna girdi; baskin aile social, baskin topic models/launches/agents. Kaynak: https://the-decoder.com/open-source-tool-pxpipe-hides-text-in-pngs-to-cut-claude-code-and-fable-5-token-costs-up-to-70/
• Kaynak kalitesi notu: Reddit ve Semantic Scholar rate limit verdi; Google News aggregator linkleri canonical cozumleme gerektiriyor. Kaynak: https://www.reddit.com/r/MachineLearning/comments/1ungvxu/well_benchmark_an_open_weights_llm_on_any_gpu_you/

Coverage / Blind Spots

• Eksik family yok: rss/news, search, community, social ve academic/API calisti; ancak academic unique sadece 16 item'e indi, paper coverage sosyal/search gürültüsü altinda zayif kaldi. Kaynak: https://huggingface.co/zai-org/GLM-5.2
• Thin family yok ama Semantic Scholar 429 ve Reddit 429 gozlemlendi; API anahtari ve backoff stratejisi iyilestirilmeli. Kaynak: https://www.reddit.com/r/MachineLearning/comments/1un3lsf/barygraph_knowledge_graph_where_every/
• Bos topic yok; thin topic yok; buna ragmen REGULATION/SAFETY maddeleri cogunlukla haber aggregation uzerinden geldi, resmi dokuman agirligi artirilmali. Kaynak: https://www.anthropic.com/news/announcing-our-updated-responsible-scaling-policy

Bu Gece Sistem Ne Ogrendi

• Onceki artifact'taki "Launches, Regulation, Models" yogunlugu bugun de korundu; ancak yeni agirlik agent security ve science workbench tarafina kaydi. Kaynak: https://www.marktechpost.com/2026/07/04/anthropic-launches-claude-science-beta/
• "Search akisi aggregator linklerine dayaniyor" dersi tekrar dogrulandi; canonical URL cozumleme bir sonraki run icin yuksek oncelik. Kaynak: https://news.google.com/rss/articles/CBMijAFBVV95cUxPTmdPd09KeWhyel9IOE5DUVI5ejlZSGFwaElxSk91RmoteTJ3RDZyVUp4X21YaGQyN3ZKUzgwNEpDYmRzMUVSa1o2Q0twVHRxOTJnVzRBbU1xTFhmTnY4bTJLWmdrTWZva1BsMXFncm44SlBxS3dQMldvUUhPOE5XN21Kc01ZZl9mWDAybQ?oc=5
• Rising entities icinde Anthropic/OpenAI/Google devam ediyor ama bugun pratik karar sinyali Alibaba, Mistral ve Anthropic Science uzerinden geldi. Kaynak: https://techcrunch.com/2026/07/04/alibaba-reportedly-bans-employees-from-using-claude-code/ ve https://mistral.ai/news/leanstral-1-5

Dedupe & Kalite Notu

Bu rapordaki maddeler onceki 3 gunun ana tekrarlarindan elendi/dedupe edildi. Toplam 5006 item islendi; merge sonrasi 4986 kayit, dedupe sonrasi 963 yeni benzersiz kayit kaldi. 4013 fingerprint tekrari ve 10 semantik tekrar atildi.