Oracle Gece Arastirma - 2026-07-07

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

Oracle Gece Arastirma - 2026-07-07

Otomatik derleme | v3 pipeline | 5076 kaynak | 821 benzersiz item | 86 benzersiz kaynak | 5/5 kaynak ailesi

Gunun Ozeti

Bugunun ana ekseni agent ekosisteminin "urun" olmaktan cikup runtime, guvenlik, odeme, governance ve veri altyapisi katmanlarina yayilmasi. Vercel'in model-agent ayrimi tartismasi, x402 agent payments ve MCP enterprise auth ayni hatta sinyal verdi. Kaynaklar: https://techcrunch.com/2026/07/06/vercel-ceo-guillermo-rauch-on-the-fight-to-split-off-models-from-agents/ ve https://www.infoq.com/news/2026/07/cloudflare-aws-x402-micropayment/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=AI%2C+ML+%26+Data+Engineering ve https://www.infoq.com/news/2026/07/mcp-ema-enterprise-auth/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=AI%2C+ML+%26+Data+Engineering

Ikinci hat acik model ve altyapi: Tencent Hy3, AWS Bedrock uzerinde MiniMax, Hugging Face Kernels ve NVIDIA open-model analizi API merkezli model pazarinin daha parcali ve portatif hale geldigini gosteriyor. Kaynaklar: https://the-decoder.com/tencent-releases-hy3-open-source-model-that-allegedly-matches-models-up-to-five-times-its-active-size/ ve https://aws.amazon.com/blogs/machine-learning/run-minimax-models-on-amazon-bedrock/ ve https://huggingface.co/blog/revamped-kernels ve https://blogs.nvidia.com/blog/open-models-icml-2026/

Ucuncu hat guvenlik ve governance: SkillCloak, AI sandbox testleri, Alberta hukumetinin Claude ile vulnerability fix akisi ve Cin'in chatbot persona kisitlamasi agent guvenligini sadece prompt injection degil dagitim politikasi meselesi haline getiriyor. Kaynaklar: http://www.bing.com/news/apiclick.aspx?ref=FexRss&aid=&tid=6a4c1ea55479481ea6098fb215a36de1&url=https%3a%2f%2fthehackernews.com%2f2026%2f07%2fnew-skillcloak-technique-lets-malicious.html&c=4423112274932331713&mkt=tr-tr ve http://www.bing.com/news/apiclick.aspx?ref=FexRss&aid=&tid=6a4c1ea55479481ea6098fb215a36de1&url=https%3a%2f%2fwww.ibtimes.sg%2fai-agents-run-their-own-code-researchers-test-sandboxes-meant-contain-them-89167&c=3869947973529818746&mkt=tr-tr ve https://www.anthropic.com/news/alberta-government-claude-cybersecurity ve https://the-decoder.com/china-forces-its-biggest-ai-platforms-to-shut-down-humanlike-chatbot-personas/

TRENDS

• Models, launches ve agents bugunku ilk uc topic; regulation ve security hemen arkadan geliyor. Kaynak: https://huggingface.co/tencent/Hy3
• Trend penceresinde AI Agents hacmi dusuyor ama sinyal kalitesi yuksek: daha az genel hype, daha cok odeme, auth, sandbox ve runtime kontrolu. Kaynak: https://www.infoq.com/news/2026/07/mcp-ema-enterprise-auth/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=AI%2C+ML+%26+Data+Engineering
• ElevenLabs yukselen tek net pozitif entity; buna karsin bugunku karar degeri daha cok agent governance ve infrastructure tarafinda. Kaynak: https://www.infoq.com/news/2026/07/cloudflare-aws-x402-micropayment/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=AI%2C+ML+%26+Data+Engineering

Top 7

1. Tencent Hy3, 295B toplam ve 21B aktif MoE yapisiyla buyuk acik model rekabetini yeniden hareketlendirdi. Kaynak: https://the-decoder.com/tencent-releases-hy3-open-source-model-that-allegedly-matches-models-up-to-five-times-its-active-size/
2. SkillCloak, malicious AI agent skill'lerinin static scanner'lardan self-extracting packing ile kacabilecegini gosteriyor. Kaynak: http://www.bing.com/news/apiclick.aspx?ref=FexRss&aid=&tid=6a4c1ea55479481ea6098fb215a36de1&url=https%3a%2f%2fthehackernews.com%2f2026%2f07%2fnew-skillcloak-technique-lets-malicious.html&c=4423112274932331713&mkt=tr-tr
3. Vercel CEO Guillermo Rauch, model ile agent katmaninin ayrilmasini production price-performance problemi olarak cerceveliyor. Kaynak: https://techcrunch.com/2026/07/06/vercel-ceo-guillermo-rauch-on-the-fight-to-split-off-models-from-agents/
4. AWS, MiniMax modellerini Bedrock'a ekleyerek frontier disi model portfoyunu enterprise dagitima tasiyor. Kaynak: https://aws.amazon.com/blogs/machine-learning/run-minimax-models-on-amazon-bedrock/
5. NVIDIA Kyber NVL144 rack gecikmesi, AI donanim road map'inde uretim sinirlarinin stratejik risk oldugunu gosteriyor. Kaynak: https://www.cnbc.com/2026/07/06/nvidia-kyber-rack-system-delays-manufacturing-taiwan-rubin-chips-.html
6. LeRobot v0.6.0, robotik gelistirme akisini imagine, evaluate, improve dongusune cekiyor. Kaynak: https://huggingface.co/blog/lerobot-release-v060
7. Cin'in leading chatbot'larda humanlike persona'lari kapattirmasi, AI companion ve karakter bot pazarinda regulator riskini artiriyor. Kaynak: https://the-decoder.com/china-forces-its-biggest-ai-platforms-to-shut-down-humanlike-chatbot-personas/

KAT-1 - LLM & Model Guncellemeleri

• Hy3, Apache 2.0 sinyali ve MoE verimliligiyle Cin acik model hattinda yeni benchmark baskisi yaratiyor. Kaynak: https://huggingface.co/tencent/Hy3
• MiniMax'in Bedrock'a gelmesi, kurumsal ekiplerin tek vendor yerine model portfoyu yonetmesine imkan veriyor. Kaynak: https://aws.amazon.com/blogs/machine-learning/run-minimax-models-on-amazon-bedrock/
• Zhipu AI ZCode, Claude Code ve Codex'e dusuk maliyetli agent-first coding alternatifi olarak konumlaniyor. Kaynak: https://the-decoder.com/zhipu-ai-launches-zcode-to-challenge-claude-code-and-openai-codex-at-a-fraction-of-the-cost/
• GPT-4 donemiyle karsilastirildiginda lider modellerin zirvede kalma suresinin kisalmasi, model seciminde "kalici lider" varsayimini zayiflatiyor. Kaynak: https://the-decoder.com/gpt-4s-dominance-lasted-a-year-while-todays-top-models-barely-survive-seven-weeks-at-the-top/

KAT-2 - Agent Ekosistemi

• Vercel'in model-agent ayrimi tartismasi, agent runtime'in application layer'da ayri optimizasyon gerektirdigini netlestiriyor. Kaynak: https://techcrunch.com/2026/07/06/vercel-ceo-guillermo-rauch-on-the-fight-to-split-off-models-from-agents/
• Reddit, LLM kaynakli spam problemini LLM'lerle temizlemeye calisiyor; platform moderasyonu artik model-vs-model savasina kayiyor. Kaynak: https://techcrunch.com/2026/07/06/reddit-is-using-llms-to-solve-a-problem-llms-largely-created/
• OfficeCLI, ajanlarin Microsoft Office dosyalarini okuyup duzenlemesi icin agent-native ofis arayuzu sunuyor. Kaynak: https://github.com/iOfficeAI/OfficeCLI
• OpenAI roon'un tool-oriented AI yerine autonomous agent savi, tool orchestration paradigmasina karsi sosyal sinyal uretti. Kaynak: https://x.com/deepfates/status/2074000607984439409

KAT-3 - Guvenlik & Risk

• SkillCloak, skill paketlerinin sadece static tarama ile guvenli sayilamayacagini gosteriyor. Kaynak: http://www.bing.com/news/apiclick.aspx?ref=FexRss&aid=&tid=6a4c1ea55479481ea6098fb215a36de1&url=https%3a%2f%2fthehackernews.com%2f2026%2f07%2fnew-skillcloak-technique-lets-malicious.html&c=4423112274932331713&mkt=tr-tr
• AI agent'larin kendi kodunu calistirdigi sandbox'larda eski veya patchsiz bilesen riski one cikiyor. Kaynak: http://www.bing.com/news/apiclick.aspx?ref=FexRss&aid=&tid=6a4c1ea55479481ea6098fb215a36de1&url=https%3a%2f%2fwww.ibtimes.sg%2fai-agents-run-their-own-code-researchers-test-sandboxes-meant-contain-them-89167&c=3869947973529818746&mkt=tr-tr
• Alberta hukumeti, Claude'u kamu sistemlerinde vulnerability bulma ve fix akisi icin kullaniyor; AI security artik defense operasyonuna giriyor. Kaynak: https://www.anthropic.com/news/alberta-government-claude-cybersecurity
• Job interview temali phishing kampanyalari Google hesap hirsizligini hedefliyor; AI is akisi guvenligi identity layer'a bagli kalmaya devam ediyor. Kaynak: https://www.bleepingcomputer.com/news/security/phishing-poses-as-big-brand-job-interview-to-steal-google-accounts/

KAT-4 - Arastirma & Paper'lar

• Anthropic global workspace arastirmasi, language model icindeki bilgi paylasim yapisini cognitive architecture acisindan yeniden tartismaya acti. Kaynak: https://mastodon.social/@h4ckernews/116874621954304430
• MultAttnAttrib, uzun dokuman QA'da multimodal attribution icin training-free yontem oneriyor. Kaynak: https://huggingface.co/papers/2607.01420
• Measuring the Gap Between Human and LLM Research Ideas, LLM fikirlerinin insan arastirma fikirlerinden ne kadar uzaklastigini olcmeye calisiyor. Kaynak: https://huggingface.co/papers/2607.01233
• WARP, model agirliklarindan training data portfolio izlerini geri okumaya calisarak veri seffafligi tartismasina teknik aci ekliyor. Kaynak: https://huggingface.co/papers/2607.01686

KAT-5 - Acik Kaynak & GitHub

• LeRobot v0.6.0, robotik deney tasarimi ve degerlendirmeyi daha tekrar edilebilir hale getiriyor. Kaynak: https://huggingface.co/blog/lerobot-release-v060
• Hugging Face Kernels update'i, model serving performansini repo-native kernel katmanina daha yakin tasiyor. Kaynak: https://huggingface.co/blog/revamped-kernels
• Planning-with-files, uzun agent gorevlerinde crash-proof markdown plan ve completion gate fikrini yayginlastiriyor. Kaynak: https://github.com/OthmanAdi/planning-with-files
• Taste-Skill, frontend agent'larinda "anti-slop" tasarim kurali olarak skill katmaninin yalniz otomasyon degil kalite kontrol icin de kullanildigini gosteriyor. Kaynak: https://github.com/Leonxlnx/taste-skill

KAT-6 - Endustri & Sirketler

• Station F, Avrupa AI startup pipeline'ini F/ai accelerator ile guclendirmeye calisiyor. Kaynak: https://techcrunch.com/2026/07/06/station-f-ramps-up-as-a-launchpad-for-europes-hottest-ai-startups/
• TeraWulf-Anthropic 19 milyar dolarlik AI infrastructure lease haberleri, model sirketlerinin enerji ve compute kontratlarini ana stratejiye tasidigini gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMivwFBVV95cUxOU0llTW12ck1BS3BsV2ZPR3BNWVFYcEFpcERKajZfUENDRkpZNkQ0ODhDVllBV01uQU9CWjVJNmVsZDV3amthMjJpVThYbHlnNHFsd3BfdW02OEZrVjNhczVQdVRlSDFYMzloa25ibDRJMzgtU09jY0hWTW02elEzOVFKZjN6Tnctd0ZKenFoeU9UZkdqM2xWZ2RzMXFlWVNlY3NUNVE0S3o4WWIzNG16cGRJZ1NlZ204a2NzcjhzTQ?oc=5
• OpenAI icin onerilen "aile basina 300 dolar pay" tartismasi, AI getirilerinin halka dagitimi narratifini tekrar gundeme tasidi. Kaynak: https://www.technologyreview.com/2026/07/06/1140176/your-familys-300-stake-in-openai/
• Even Realities, Tencent ve Meituan destekli smart-glasses rekabetinde 1 milyar dolar degerlemeye ulasti. Kaynak: https://www.cnbc.com/2026/07/06/apple-veteran-takes-on-meta-with-1-billion-smart-glasses-maker.html

KAT-7 - Multimodal

• Siri'nin pace ve expressivity ayarlari, voice assistant deneyiminde kisilestirmeyi model yetenegi kadar UX parametresi haline getiriyor. Kaynak: https://techcrunch.com/2026/07/06/you-can-now-customize-siris-pace-and-expressivity-in-the-latest-ios-27-beta/
• Smart glasses privacy tartismasi, kamera ve AI wearable kombinasyonunun sosyal kabul riskini buyutuyor. Kaynak: https://www.theverge.com/column/961707/smart-glasses-ai-wearables-meta-surveillance-privacy
• Amazon Nova ile image PII redaction pipeline'i, multimodal modellerin privacy workflow'larinda kullanilabilecegini gosteriyor. Kaynak: https://aws.amazon.com/blogs/machine-learning/automatically-redact-pii-in-images-with-amazon-nova/
• AnyBokeh, tek goruntuden fizik-guided bokeh editing ile generative photo editing hattinda ilerleme sunuyor. Kaynak: https://huggingface.co/papers/2606.31959

KAT-8 - Robotik & Embodied AI

• VLA-Corrector, vision-language-action modellerinde action horizon'u hafif detect-and-correct inference ile ayarlamayi hedefliyor. Kaynak: https://huggingface.co/papers/2607.01804
• Embodied.cpp, farkli robot platformlarinda portable embodied AI inference runtime ihtiyacini hedefliyor. Kaynak: https://huggingface.co/papers/2607.02501
• NVIDIA destekli Generalist AI degerleme haberi, robot foundation model pazarinin finansman tarafinda hizlandigini gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMitgFBVV95cUxQUkl0dWo1cmVmQTAwa3FPOUlhUGdKUU45U1N0STUxWEtodnpqT1FnaHlGQ1ZNTFo4M3d2S3IxSGFYVHE5X2hZQ05rMUFZUFYtQ1g5Wjc4MVhsSW40SVBWS3otU1hBMjRlOWRaZFZaM2hVMTd6elFPY0pSa2V0V3F5XzlrTVhRT1lzcEUtRXhNLVJtZV82QmJvUFY2OUFuZ3FVNDFSYUxFSFVnbHpzZ2k3YkJ5TGh5QQ?oc=5

KAT-9 - Veri & Operasyon

• Google'in privacy ayari degisikligi, kullanici medya ve dosyalarinin AI training'e dahil edilmesi tartismasini yeniden acti. Kaynak: https://techcrunch.com/2026/07/06/if-you-use-google-youre-training-its-ai-heres-how-to-opt-out/
• SageMaker AI'nin MLflow streaming entegrasyonu, benchmark ve inference recommendation sonuclarini deney takibine daha dogrudan bagliyor. Kaynak: https://aws.amazon.com/blogs/machine-learning/streaming-benchmark-and-recommendation-results-to-mlflow-with-amazon-sagemaker-ai/
• TRACE tartismasi, flat RAG yerine hiyerarsik agent memory'nin community tarafinda guclenen ihtiyac oldugunu gosteriyor. Kaynak: https://www.reddit.com/r/MachineLearning/comments/1uoz5jo/trace_opensource_hierarchical_memory_for_llm/
• Google Chrome'un lokal 4GB AI modeli tartismasi, browser icine gomulu model dagitiminda seffaflik sorusunu buyutuyor. Kaynak: https://oztalking.com/en/issues/hidden-4gb-ai-model

SECURITY

• MCP Enterprise-Managed Authorisation stable oldu; enterprise agent erisiminde merkezi auth katmani standartlasiyor. Kaynak: https://www.infoq.com/news/2026/07/mcp-ema-enterprise-auth/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=AI%2C+ML+%26+Data+Engineering
• Securing the AI Agent paper'i, MCP, model serving ve agent platformlarini kapsayan multi-layer red teaming cercevesi oneriyor. Kaynak: https://huggingface.co/papers/2606.31227
• InfoQ'nun AI Security & Privacy Engineering cohort'u, regulated industries tarafinda AI guvenliginin ayri bir meslek pratigi haline geldigini gosteriyor. Kaynak: https://www.infoq.com/news/2026/07/online-cohort-ai-security/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=AI%2C+ML+%26+Data+Engineering

REGULATION

• Cin'in AI persona kisitlamasi, chatbot urunlerinde antropomorfik baglilik riskine regulator mudahalesi olarak okunmali. Kaynak: https://asia.nikkei.com/business/technology/artificial-intelligence/china-s-leading-chatbots-to-ditch-ai-personas-as-beijing-tightens-rules
• Meta'nin Hindistan'da Instagram child abuse ads nedeniyle regulator baskisi gormesi, platform AI moderation riskini yukseltir. Kaynak: https://www.cnbc.com/2026/07/06/meta-instagram-india-warning-whatsapp.html
• ABD'de data center ban karari, AI altyapisinda yerel belediye ve enerji kapasitesi riskinin buyudugunu gosteriyor. Kaynak: https://asia.nikkei.com/business/technology/artificial-intelligence/how-an-asian-majority-city-passed-the-us-first-permanent-data-center-ban
• DMA'nin AI inovasyonu etkisi tartismasi, Avrupa'da platform regulation ile AI rekabetinin ayni dosyada birlestigini gosteriyor. Kaynak: https://news.google.com/rss/articles/CBMilAFBVV95cUxOVGpYU3BxWjZIM0VoM3N3Y0wyNHR0dDF0ZDNBWWNHcllBcmxnZ252dEZTRzRnRDFRX2ctRXcyTzRiOFN6aHdrRDFGUk9yZ0RWOVE5b3ZPRHExWUtFLUdWaUltZDZUZy1teDhlenFxUXlmNTBINEQyM1RqeFgwSlpMb3ZRUEU4aHRvajU4ckw1QXV1UFlm?oc=5

AI-SCIENCE

• LLM'lerin underrepresented epilepsy care icin recommend/defer davranisini ogrenmesi, klinik karar destek sistemlerinde guvenli erteleme yetenegini one cikariyor. Kaynak: https://huggingface.co/papers/2606.31036
• Plant disease identification calismasi, smartphone tabanli AI'nin tarla seviyesinde erken tani araci olarak yayilabilecegini gosteriyor. Kaynak: http://www.bing.com/news/apiclick.aspx?ref=FexRss&aid=&tid=6a4c1ea5128c46a8bd172af8ca38537e&url=https%3a%2f%2fwww.cirad.fr%2fen%2fcirad-news%2fnews%2f2026%2fidentifying-plant-diseases&c=10503222178499496601&mkt=tr-tr
• ML destekli superconductor taramasi, material discovery'de candidate filtering icin AI kullanimini guclendiriyor. Kaynak: http://www.bing.com/news/apiclick.aspx?ref=FexRss&aid=&tid=6a4c1ea67fcf47eda891f393764dc909&url=https%3a%2f%2fwww.azoquantum.com%2fNews.aspx%3fnewsID%3d11178&c=15257464609591554958&mkt=tr-tr

INFRA

• Cloudflare Workers Cache, Worker entrypoint'leri onune regional cache koyarak edge compute maliyet ve latency denklemini iyilestiriyor. Kaynak: https://blog.cloudflare.com/workers-cache/
• x402 agent payments, agent-to-service odemelerde HTTP 402'yi edge seviyesinde yeniden kullanima sokuyor. Kaynak: https://www.infoq.com/news/2026/07/cloudflare-aws-x402-micropayment/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=AI%2C+ML+%26+Data+Engineering
• DigitalOcean'un hidden GPU charge analizi, real-time inference maliyetinin list price'in yaklasik ustune cikabilecegini gosteriyor. Kaynak: https://www.digitalocean.com/community/tutorials/four-hidden-gpu-charges
• Inferentia2, TPU, Groq LPU ve Tenstorrent karsilastirmasi, GPU disi inference silicon seciminin workload'a bagli oldugunu netlestiriyor. Kaynak: https://www.digitalocean.com/community/tutorials/inferentia2-tpu-groq-lpu-vs-gpu-llm-serving

SAFETY

• Anthropic'in Responsible Scaling Policy guncellemesi, frontier model risk governance'i icin RSP hattini yeniden konumlandiriyor. Kaynak: https://www.anthropic.com/news/announcing-our-updated-responsible-scaling-policy
• Multimodal attribution ve PII redaction haberleri, safety'nin artik sadece model refusal degil evidence ve data handling meselesi oldugunu gosteriyor. Kaynaklar: https://huggingface.co/papers/2607.01420 ve https://aws.amazon.com/blogs/machine-learning/automatically-redact-pii-in-images-with-amazon-nova/
• Reddit'teki model sameness tartismasi, uzun kullanimda model ciktisi cesitliligi ve epistemik daralma riskini community gundemine tasiyor. Kaynak: https://www.reddit.com/r/MachineLearning/comments/1uon503/does_anyone_have_a_name_for_that_subtle_sameness/

WATCHLIST

• Hy3: acik MoE performans iddialari benchmark bagimsiz dogrulamaya muhtac. Kaynak: https://the-decoder.com/tencent-releases-hy3-open-source-model-that-allegedly-matches-models-up-to-five-times-its-active-size/
• SkillCloak: skill registry ve install pipeline'larinda runtime unpacking kontrolu sart. Kaynak: http://www.bing.com/news/apiclick.aspx?ref=FexRss&aid=&tid=6a4c1ea55479481ea6098fb215a36de1&url=https%3a%2f%2fthehackernews.com%2f2026%2f07%2fnew-skillcloak-technique-lets-malicious.html&c=4423112274932331713&mkt=tr-tr
• x402: agent payments yayilirsa billing, abuse prevention ve audit log tasarimi kritik olacak. Kaynak: https://www.infoq.com/news/2026/07/cloudflare-aws-x402-micropayment/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=AI%2C+ML+%26+Data+Engineering
• Data center bans: compute supply artik sadece GPU degil yerel izin, enerji ve su siyasetiyle kilitleniyor. Kaynak: https://asia.nikkei.com/business/technology/artificial-intelligence/how-an-asian-majority-city-passed-the-us-first-permanent-data-center-ban
• Google AI training opt-out: consumer data policy degisiklikleri enterprise DLP ve privacy egitimine yansitilmali. Kaynak: https://techcrunch.com/2026/07/06/if-you-use-google-youre-training-its-ai-heres-how-to-opt-out/

CikCik Paketi

• "Tool-oriented AI kaybeder, autonomous agents kazanir" tartismasi kisa ve keskin thread adayi. Kaynak: https://x.com/deepfates/status/2074000607984439409
• Simon Willison'un AI ile uzmanlik hissi yorumu, "AI seni tembellestiriyor mu yoksa kapasiteni mi uzatiyor?" acisindan iyi quote. Kaynak: https://bsky.app/profile/simonwillison.net/post/3mpwuokdch22t
• Gemma 4'un lokal calistigini inkar etmesi, fine-tuning contamination ve self-awareness yanilsamasi icin guzel teknik meme. Kaynak: https://x.com/Strife212/status/2073815736925966675
• TeraWulf-Anthropic 19B lease, AI'nin yeni enerji politikasina donustugunu gosteren finansal sinyal. Kaynak: https://x.com/Techmeme/status/2074107474261807577
• TRACE hierarchical memory post'u, "flat RAG yetmiyor" tartismasi icin community acisi veriyor. Kaynak: https://www.reddit.com/r/MachineLearning/comments/1uoz5jo/trace_opensource_hierarchical_memory_for_llm/
• AI marketing backlash yazisi, "AI-first" etiketinin artik marka gucu degil risk olabilecegini gosteriyor. Kaynak: https://www.breef.com/breefingroom/articles/the-ai-marketing-backlash-why-ai-first-brands-are-starting-to-fall-flat

Rehberler & Kaynaklar

• DigitalOcean GPU cost rehberi, inference maliyeti hesaplarken egress, idle time, cold start ve noisy neighbor etkilerini ayri kalem yapmayi oneriyor. Kaynak: https://www.digitalocean.com/community/tutorials/four-hidden-gpu-charges
• Inferentia2/TPU/Groq/Tenstorrent rehberi, LLM serving icin silicon secimini latency, compilation ve memory ceiling uzerinden yapiyor. Kaynak: https://www.digitalocean.com/community/tutorials/inferentia2-tpu-groq-lpu-vs-gpu-llm-serving
• AWS Nova RL infrastructure yazisi, multi-turn RL egitim altyapisini HyperPod uzerinde event-driven pipeline olarak kuruyor. Kaynak: https://aws.amazon.com/blogs/machine-learning/deploying-multi-turn-rl-infrastructure-for-amazon-nova-on-amazon-sagemaker-hyperpod/

Oracle Self-Improvement Sinyalleri

• Pipeline bugun 5 kaynak ailesinin tamamini kapsadi; en zayif nokta API key olmayan Currents/NewsData kaynaklari ve Reddit 429'lari.
• Social collector 4099 ham item uretirken benzersiz katkisi 338 item oldu; sosyal kaynakta noise yuksek, fakat CikCik sinyali icin hala degerli.
• Dedupe 5039 merged item'dan 821 unique item cikardi; 4216 fingerprint tekrari ve 2 semantik tekrar temizlendi.
• Kalici ders: source canonicalization hala gerekli; Google/Bing aggregator linkleri rapor kalitesini dusuruyor.

Kaynak Ozeti

• RSS/news: 316 ham, 217 unique, baskin kaynaklar Nikkei Asia, DonanimHaber, Technopat.
• Search: 519 ham, 226 unique, baskin kaynaklar google_news/ai, google_news/companies, google_news/funding.
• Community: 72 ham, 25 unique, baskin kaynaklar GitHub Trending, Hacker News AI, Reddit MachineLearning.
• Social: 4099 ham, 338 unique, baskin kaynaklar mastodon/#AI, mastodon/#MachineLearning, mastodon/#LLM.
• Academic/API: 70 ham, 15 unique, baskin kaynaklar Hugging Face daily papers ve trending models.

Coverage / Blind Spots

• Coverage tam: 5/5 family, 10/10 topic kapsandi.
• Eksik family yok; bos topic yok.
• Thin family yok; ancak academic/API tarafinda Semantic Scholar 429 ve PapersWithCode parse hatasi nedeniyle akademik kaynak cesitliligi sinirli kaldi.
• En baskin konu basliklari: models 173, launches 136, agents 114, regulation 93, security 83.

Bu Gece Sistem Ne Ogrendi

• Launches, Regulation ve Models kalici yuksek hacimli akislari suruyor; ancak bugunku karar degeri Agents, Security ve Infra kesismesinde.
• Search family aggregator linklerine dayanmaya devam ediyor; canonical URL cozumleme bir sonraki pipeline iyilestirmesi olmali.
• Rising entities listesinde Meta, AI Regulation, Google, AI Agents, Anthropic, Claude, OpenAI ve RAG var; bugunku haberler bu listenin regulation/security agirligini dogruladi.

Dedupe & Kalite Notu

Bu rapordaki haber secimi onceki 3 gunun raporlarindaki MTurk, S3 Annotations, Strix, codex-plugin-cc, Alibaba/Claude Code, Leanstral, GLM, Qwen AgentWorld, Chrome DevTools MCP ve AISI benchmark tekrarlarindan arindirildi. Toplam 5076 ham item islendi, 5039 merged item uzerinden 821 benzersiz item rapor havuzuna alindi.