Oracle Gece Arastirma - 2026-08-05

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

Oracle Gece Arastirma - 2026-08-05

Otomatik derleme | v3 pipeline | 9.109 raw item | 4.150 input unique | 3.823 raporlanabilir item | 108 benzersiz kaynak | 5/5 kaynak ailesi

Gunun Ozeti

Bugunun ana ekseni agent sistemlerinin "model" olmaktan cikarak runtime, odeme, CI/CD, memory ve guvenlik katmanina yayilmasi oldu. Cloudflare agent lifecycle, programmable wallets ve AI destekli engineering standards hattini birlikte duyurdu; AWS Bedrock ise web grounding ve AgentCore ile web insight extraction tarafini genisletti. Kaynaklar: https://blog.cloudflare.com/agent-development-lifecycle/ ve https://blog.cloudflare.com/wallets/ ve https://aws.amazon.com/blogs/machine-learning/introducing-web-search-on-amazon-bedrock-for-foundation-model-grounding/

Guvenlikte odak, onceki gunlerdeki "AI hacking" anlatiminin arkasindaki kontrol katmanlarina kaydi: OpenAI resmi olarak third-party cyber evaluation safeguard'larini anlatti, NVIDIA/Open Secure AI Alliance SAFE guideline hattini acti, BleepingComputer ve Snyk supply-chain/agent-boundary risklerini gorunur kildi. Kaynaklar: https://openai.com/index/third-party-cyber-evaluations-involving-openai-models ve https://blogs.nvidia.com/blog/open-secure-ai-alliance-contributions/ ve https://www.bleepingcomputer.com/news/security/77-open-vsx-extensions-found-harvesting-developer-info/ ve https://snyk.io/blog/remediation-agent-malicious-code-defense/

Arastirma tarafinda self-improving agent memory, agent evaluation ve world model safety cizgisi guclendi: Memory Reward Inflation, CrystalMem, AgentStream, Agentic Coding in the Wild ve WorldExam bugunun temiz paper sinyalleri. Kaynaklar: https://arxiv.org/abs/2608.00017 ve https://arxiv.org/abs/2608.00303 ve https://arxiv.org/abs/2608.00155 ve https://arxiv.org/abs/2608.00101 ve https://huggingface.co/papers/2608.02603

Trend Analizi

Autonomous topic'i spike verdi; AI Agents 1.224 toplam sinyal ile yukselis hattinda kaldi. Bu, sadece yeni agent urunleri degil, agent'larin odeme, CI, tracing, memory ve guvenlik politikasina baglanmasindan kaynaklaniyor. Kaynaklar: https://blog.cloudflare.com/wallets/ ve https://blog.cloudflare.com/ci-workflows/ ve https://blog.cloudflare.com/local-tracing/

Regulation 1.929 topic sayisi ile yine baskin; bugun ozellikle data center izinleri, Chinese open-weight model ban tartismasi, chip hukuku ve workplace AI monitoring onerileri onde. Kaynaklar: https://techcrunch.com/2026/08/04/texas-halts-new-data-centers-as-governor-calls-for-audits/ ve https://the-decoder.com/silicon-valleys-rift-over-open-source-pushes-back-contemplated-white-house-bans-on-chinese-ai/ ve https://www.theregister.com/software/2026/08/04/uk-mulls-making-employers-ask-before-installing-bossware/5282440

Azalan tarafta Amazon/Databricks/Haiku/Sonnet sinyalleri var; fakat AWS Bedrock Web Search ve AgentCore yazilari, Amazon'un "daha az trend mention, daha cok urun katmani" seklinde ilerledigini gosteriyor. Kaynaklar: https://aws.amazon.com/blogs/machine-learning/introducing-web-search-on-amazon-bedrock-for-foundation-model-grounding/ ve https://aws.amazon.com/blogs/machine-learning/automated-web-insight-extraction-with-amazon-bedrock-agentcore/

Top 7

1. Cloudflare, Agent Development Lifecycle ile agent'larin yazma, review, deploy ve maintain surecini platform primitive'lerine bagladi; agent ops artik uygulama yasam dongusu sorunu. Kaynak: https://blog.cloudflare.com/agent-development-lifecycle/
2. OpenAI, third-party cyber evaluation olaylari icin resmi safeguard anlatimini yayinladi; model eval sandbox'lari artik siradan test ortami degil, yuksek riskli guvenlik alani. Kaynak: https://openai.com/index/third-party-cyber-evaluations-involving-openai-models
3. NVIDIA/Open Secure AI Alliance SAFE guidelines, agentic AI cybersecurity icin transparency standardi onermeye basladi. Kaynak: https://blogs.nvidia.com/blog/open-secure-ai-alliance-contributions/
4. Anthropic, Volta ile 10 milyar dolarlik compute kapasitesi anlasmasina gitti; frontier rekabette cloud risk ve kapasite tedariki stratejik avantaj haline geliyor. Kaynak: https://techcrunch.com/2026/08/04/anthropic-signs-10-billion-deal-with-ai-cloud-startup-volta/
5. TechCrunch/SaferAI sinyali, open-weight modellerin frontier yeteneklere yaklasirken safety gap'i kapatamadigini vurguladi. Kaynak: https://techcrunch.com/2026/08/04/open-weight-ai-models-are-catching-up-to-the-frontier-the-safety-gap-remains/
6. Agentic Coding in the Wild, Copilot/Codex/Claude Code benzeri ajan is yuklerini production trace seviyesinde inceleyen yeni bir paper olarak one cikti. Kaynak: https://arxiv.org/abs/2608.00101
7. WorldExam, video/world model degerlendirmesini sadece gorunum kalitesinden fiziksel reactivity ve world-model davranisina tasiyor. Kaynak: https://huggingface.co/papers/2608.02603

KAT-1 - LLM & Model Guncellemeleri

• Liquid AI, LFM2.5-2.6B ile local agent deployment mesajini one cikardi; dusuk maliyetli lokal inference, background agent'lar icin pratiklesiyor. Kaynak: https://huggingface.co/blog/LiquidAI/lfm2-5-2-6b
• DiffusionGemma, token-by-token autoregressive generation yerine discrete diffusion ile hiz odakli open-weight language model cizgisini guclendiriyor. Kaynak: https://huggingface.co/papers/2608.00146
• The Register, 10 dolarlik microcontroller uzerinde LLM calistirma demosunu yazdi; edge tarafinda "yeterince iyi ve ucuz" deneyler artiyor. Kaynak: https://www.theregister.com/edge-and-iot/2026/08/04/dev-proves-llms-will-run-on-anything-even-a-10-microcontroller/5283088

KAT-2 - Arastirma & Paper'lar

• Memory Reward Inflation, self-improving LLM agent'larda dis memory scoring'in kendi odul sinyalini sisirebilecegini gosteriyor; Oracle tipi memory sistemleri icin direkt risk. Kaynak: https://arxiv.org/abs/2608.00017
• CrystalMem, self-evolving agent memory'sini elastic quota ve knowledge crystallization mantigiyla ele aliyor; sinirsiz memory varsayimina saldiri. Kaynak: https://arxiv.org/abs/2608.00303
• AgentStream, streaming task altinda self-evolving LLM agent performansini olcuyor; statik benchmark yerine surekli is akisi testine kayis var. Kaynak: https://arxiv.org/abs/2608.00155
• RAG-TESTER, RAG sistemleri icin end-to-end otomatik test catisi oneriyor; retrieval kalitesi ve response dogrulugunu ayni hatta bagliyor. Kaynak: https://arxiv.org/abs/2608.00054
• SIRIN, memory-grounded ve retrieval-augmented LLM sistemlerinde contextual hallucination tespiti icin toolkit/web UI oneriyor. Kaynak: https://arxiv.org/abs/2608.00033

KAT-3 - Araclar & Framework'ler

• AWS Bedrock Web Search, foundation model cevaplarini guncel web bilgisiyle server-side grounding yapacak sekilde GA oldu. Kaynak: https://aws.amazon.com/blogs/machine-learning/introducing-web-search-on-amazon-bedrock-for-foundation-model-grounding/
• AWS Bedrock AgentCore Browser, coklu web sitesinden otomatik insight extraction icin referans mimari yayinladi. Kaynak: https://aws.amazon.com/blogs/machine-learning/automated-web-insight-extraction-with-amazon-bedrock-agentcore/
• Cloudflare CI Workflows, milyonlarca repo icin sandboxed CI/CD pipeline'i platform uzerine almaya calisiyor. Kaynak: https://blog.cloudflare.com/ci-workflows/
• Cloudflare local tracing, coding agent'in Workers hatasini deploy etmeden trace API ile bulmasini hedefliyor. Kaynak: https://blog.cloudflare.com/local-tracing/

KAT-4 - Acik Kaynak

• GitHub trending'de AirLLM, 70B inference'i tek 4GB GPU ile calistirma iddiasiyla lokal inference hafifletme hattinda one cikti. Kaynak: https://github.com/lyogavin/airllm
• TencentDB Agent Memory, team-level agent memory hub olarak chat memory, skill, LLM-wiki ve code-graph varliklarini birlestiriyor. Kaynak: https://github.com/TencentCloud/TencentDB-Agent-Memory
• pdf-inspector, PDF routing ve scanned/text ayrimi icin Rust tabanli hafif bir inspection kutuphanesi olarak dikkat cekti. Kaynak: https://github.com/firecrawl/pdf-inspector
• DeepSeek-Reasonix, terminalde DeepSeek-native coding agent yaklasimini prefix-cache stabilitesi etrafinda kuruyor. Kaynak: https://github.com/esengine/DeepSeek-Reasonix

KAT-5 - Endustri & Sirketler

• Anthropic, Volta ile 10 milyar dolarlik compute kapasitesi anlasmasi imzaladi; compute tedarigi artik model kalitesi kadar kritik. Kaynak: https://techcrunch.com/2026/08/04/anthropic-signs-10-billion-deal-with-ai-cloud-startup-volta/
• Google, Anthropic chip/data-center riskinin milyarlarca dolarlik kismini balance sheet disina tasiyan finansman yapisi uzerinde calisiyor. Kaynak: https://the-decoder.com/google-moves-billions-in-anthropic-chip-risk-off-its-balance-sheet/
• SpaceX, xAI data center'lari icin 2026'da 329 milyon dolarlik Tesla Megapack alimi yapti; AI altyapisi enerji storage tedarik zincirine dogrudan baglaniyor. Kaynak: https://techcrunch.com/2026/08/04/spacex-has-bought-329m-worth-of-tesla-megapacks-so-far-this-year/
• Palantir CEO Alex Karp, AI labs/enterprise guven tartismasini sertlestirdi; enterprise AI satisinda guvenlik ve politik durus marka riski oluyor. Kaynak: https://techcrunch.com/2026/08/03/after-killer-quarter-palantir-ceo-alex-karp-calls-ai-industry-marxist/

KAT-6 - AI Agent'lar

• Cloudflare Wallets, agent'lara x402 ile programmable wallet, kimlik ve otonom API/content satin alma altyapisi vermeyi hedefliyor. Kaynak: https://blog.cloudflare.com/wallets/
• Cloudflare engineering standards enforcement, AI agent'larin structured RFC ve agentic review ile internal standardlara baglanmasini anlatiyor. Kaynak: https://blog.cloudflare.com/engineering-standards-enforcement/
• Shared Organizational Memory for Enterprise Coding Agents, kurum icindeki informal bilgi ve code context'i agent memory tasarimina tasiyor. Kaynak: https://arxiv.org/abs/2608.00122
• Personalizing Large Language Model Agents with Small Policy Models, agent'in soru sorma, memory cekme ve tool kullanma kararlarini kucuk policy modellerle kisisellestirmeyi inceliyor. Kaynak: https://arxiv.org/abs/2608.00215
• LongHorizon-Harness, gercek dunyaya yakin uzun-horizon agent task'leri icin harness ihtiyacini gosteriyor. Kaynak: https://huggingface.co/papers/2608.01964

KAT-7 - Multimodal

• MiniMax-H3 video generation modeli, Simon Willison tarafinda local/MLX denemesiyle pratik multimodal playground sinyali verdi. Kaynak: https://simonwillison.net/2026/Aug/4/minimax-h3-mlx/
• DeepVoyager-VL, vision-in-the-loop search ile long-horizon multimodal agent'lari tesvik etmeye calisiyor. Kaynak: https://huggingface.co/papers/2608.01827
• UEmbed, sparse ve dense multimodal embeddings'i tek retrieval katmaninda birlestirmeyi hedefliyor. Kaynak: https://huggingface.co/papers/2608.02583
• SwanTale, multi-speaker speech/audio generation'i instruct ve zero-shot senaryolara tasiyor. Kaynak: https://huggingface.co/papers/2608.02023

KAT-8 - Robotik & Embodied AI

• NVIDIA World Action Models, robot manipulation'da VLA politikalarinin otesine gecip future dynamics'i action-aligned uzayda modellemeyi one cikariyor. Kaynak: https://developer.nvidia.com/blog/beyond-vlas-how-world-action-models-reshape-robot-manipulation/
• NVIDIA Alpamayo 2 Super, robotaxi/autonomous vehicle icin trajectory, reasoning trace ve auto-label uretimini birlestiriyor. Kaynak: https://developer.nvidia.com/blog/generate-trajectories-reasoning-traces-and-auto-labels-with-nvidia-alpamayo-2-super/
• WCM, VLA reinforcement learning icin World Critic Model yaklasimini oneriyor. Kaynak: https://huggingface.co/papers/2607.29613
• Weights or Skills?, robot learning'i "frozen policy weights" ve "robotun kendi skill yazmasi" ayrimina tasiyan survey olarak izlenmeli. Kaynak: https://arxiv.org/abs/2608.01851

KAT-9 - Veri & Altyapi

• NVIDIA AI storage yazisi, memory/context window buyumesinin sadece disk kapasitesi degil, AI-native storage path sorunu oldugunu vurguluyor. Kaynak: https://blogs.nvidia.com/blog/ai-storage-fms/
• NVIDIA Vera Storage Benchmarks, encryption, compression, integrity checking ve recovery'yi AI-native storage performans problemi olarak olcuyor. Kaynak: https://developer.nvidia.com/blog/nvidia-vera-storage-benchmarks-faster-encryption-compression-integrity-checking-and-recovery-for-ai-native-storage/
• The Register, hyperscaler'larin AI talebi nedeniyle capex'i yaklasik 600 milyar dolar seviyesine tasidigini raporladi. Kaynak: https://www.theregister.com/off-prem/2026/08/04/cloud-giants-pour-nearly-600b-into-capex-as-ai-demand-surges/5282679
• Request-Level Energy Attribution, batched LLM serving'de enerji maliyetini request seviyesine atfetme problemini aciyor. Kaynak: https://arxiv.org/abs/2608.00026

SECURITY

• OpenAI'nin third-party cyber evaluation post'u, sandbox/proxy sinirlarinin model eval'lerinde daha sert tasarlanmasi gerektigini resmi olarak kabul eden sinyal. Kaynak: https://openai.com/index/third-party-cyber-evaluations-involving-openai-models
• BleepingComputer, 77 Open VSX extension'in developer bilgisi topladigini raporladi; agent/coding IDE ekosisteminde extension hygiene kritik. Kaynak: https://www.bleepingcomputer.com/news/security/77-open-vsx-extensions-found-harvesting-developer-info/
• ChainDrop npm saldirisi, yuzlerce package ve milyarlarca aylik download uzerinden supply-chain riskinin hala en buyuk attack surface oldugunu gosteriyor. Kaynak: https://www.bleepingcomputer.com/news/security/massive-chaindrop-npm-supply-chain-attack-infects-hundreds-of-packages/
• Snyk Agentic AppSec, remediation agent ve malicious code defense'i ayni urun hattina bagladi. Kaynak: https://snyk.io/blog/remediation-agent-malicious-code-defense/
• Exposed by Design, internet-facing MCP server'larin buyuk olcekte security assessment'ini yapiyor; MCP yayilimi dogrudan attack surface'e donusuyor. Kaynak: https://arxiv.org/abs/2608.00150

REGULATION

• Anthropic'in Tino Cuellar'i Chief Global Affairs Officer olarak almasi, frontier lab'larda politika/diplomasi kapasitesinin core leadership'e girdigini gosteriyor. Kaynak: https://www.anthropic.com/news/tino-cuellar
• ABD'de Chinese open-weight AI modellerine ban/cloud restriction tartismasi, open-source ekosistem icindeki cikar ayriliklari nedeniyle geri itiliyor. Kaynak: https://the-decoder.com/silicon-valleys-rift-over-open-source-pushes-back-contemplated-white-house-bans-on-chinese-ai/
• Texas, yeni data center projelerini audit cagrisiyla durdurdu; AI altyapi buyumesi enerji ve yerel izin politikasina carpiyor. Kaynak: https://techcrunch.com/2026/08/04/texas-halts-new-data-centers-as-governor-calls-for-audits/
• UK, bossware icin isverenlerin izin alma zorunlulugunu tartisiyor; AI productivity scores ve biometrics calisan mahremiyeti dosyasina girdi. Kaynak: https://www.theregister.com/software/2026/08/04/uk-mulls-making-employers-ask-before-installing-bossware/5282440

AI-SCIENCE

• AutoFOAM, OpenFOAM kullanan CFD is akislari icin self-refining autonomous agent oneriyor; bilimsel yazilim kullanimi agent'lasiyor. Kaynak: https://arxiv.org/abs/2608.00003
• Agentic Bayesian Optimization, surrogate-augmented autoresearch ile BO surecini agentik arastirma dongusune bagliyor. Kaynak: https://arxiv.org/abs/2608.00316
• DGA2D, LLM'lerle directed graph-guided automated algorithm design yaklasimini oneriyor; heuristic design alaninda LLM kullanimini sistemlestiriyor. Kaynak: https://arxiv.org/abs/2608.00700
• Abduction Without a Body?, bilimsel hipotez uretiminde continuous sensorimotor embodiment gerekip gerekmedigini tartisiyor. Kaynak: https://tldr.takara.ai/p/2608.02505

INFRA

• Cloudflare, Astro issue triage icin izole subagent'lari GitHub Actions icinde kullanarak open issue count'u %85 azaltan software factory anlatti. Kaynak: https://blog.cloudflare.com/astro-issue-triage/
• InfoQ, platform engineering maturity'nin enterprise AI success icin ayirt edici faktor haline geldigini yazdi. Kaynak: https://www.infoq.com/news/2026/08/perforce-maturity-ai-success/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=AI%2C+ML+%26+Data+Engineering
• Nova, deep learning icin end-to-end MLIR compiler olarak model performansinin compiler katmanina bagimli oldugunu hatirlatiyor. Kaynak: https://arxiv.org/abs/2608.00029
• Kernel Foundry, GPU kernel generation'i diagnosis-driven evolutionary optimizer ve multi-expert yapisiyla ele aliyor. Kaynak: https://arxiv.org/abs/2605.30359

SAFETY

• Open-weight frontier safety gap, Z.ai GLM-5.2 gibi acik modellerin yetenek olarak yaklasirken safety mitigation tarafinda geri kaldigi uyarisini veriyor. Kaynak: https://techcrunch.com/2026/08/04/open-weight-ai-models-are-catching-up-to-the-frontier-the-safety-gap-remains/
• Quality-Diversity Red-Teaming, LLM saldirganlarini hem kaliteli hem cesitli uretmek icin otomatik red-team hattini inceliyor. Kaynak: https://arxiv.org/abs/2506.07121
• More Debate, Same Evidence, homogeneous multi-agent judging'in groundedness'i yapisal olarak sinirlayabilecegini gosteriyor. Kaynak: https://arxiv.org/abs/2608.00243
• When Prompts Control Robots, multi-agent robotic systems'de prompt injection'in fiziksel kontrol riskine donustugunu aciyor. Kaynak: https://arxiv.org/abs/2608.00747
• CompanionBench, AI emotional companionship icin gercek dunya temelli benchmark ihtiyacini vurguluyor. Kaynak: https://arxiv.org/abs/2608.02046

WATCHLIST

• Cloudflare Wallets + x402: Agent'larin odeme yapmasi izin, limit, audit ve fraud risklerini birlikte getirecek. Kaynak: https://blog.cloudflare.com/wallets/
• MCP exposure: Internet-facing MCP server sayisi buyudukce prompt injection, auth bypass ve data exfiltration riskleri artacak. Kaynak: https://arxiv.org/abs/2608.00150
• Anthropic/Volta compute financing: Yeni cloud startup'lari uzerinden capacity lock-in ve counterparty risk izlenmeli. Kaynak: https://the-decoder.com/anthropic-locks-in-10-billion-of-compute-from-volta-a-cloud-startup-that-didnt-exist-six-months-ago/
• Developer extension marketplaces: Open VSX ve npm olaylari, coding agent ekosisteminde "trusted tool" varsayimini zayiflatiyor. Kaynaklar: https://www.bleepingcomputer.com/news/security/77-open-vsx-extensions-found-harvesting-developer-info/ ve https://www.bleepingcomputer.com/news/security/massive-chaindrop-npm-supply-chain-attack-infects-hundreds-of-packages/

TRENDS

• Launches 2.085 topic sayisiyle baskin; ama bugunun kaliteli launch'lari model degil, agent lifecycle, wallet, CI/CD ve grounding urunleri. Kaynaklar: https://blog.cloudflare.com/agent-development-lifecycle/ ve https://blog.cloudflare.com/wallets/ ve https://aws.amazon.com/blogs/machine-learning/introducing-web-search-on-amazon-bedrock-for-foundation-model-grounding/
• Models 1.285 topic sayisinda kaldi; LFM2.5 ve DiffusionGemma disinda en guclu sinyal safety/eval etrafinda donuyor. Kaynaklar: https://huggingface.co/blog/LiquidAI/lfm2-5-2-6b ve https://huggingface.co/papers/2608.00146
• Tooling 553, agents 537 topic sayisiyla birbirine yaklasti; agent tooling artik ayri kategori degil, ana platform rekabeti. Kaynaklar: https://blog.cloudflare.com/local-tracing/ ve https://aws.amazon.com/blogs/machine-learning/automated-web-insight-extraction-with-amazon-bedrock-agentcore/

CikCik Paketi

• Harrison Chase tarafinda managed deepagents public beta sinyali var; agent infra'nin boring/undifferentiated katmanini urunlestirme cabasi net. Kaynak: https://news.google.com/rss/articles/CBMiX0FVX3lxTFBQVjAzX2FPWHpjeEl6aEUxUjZ3MFBNMU1leTVDZVJmT2REOV96WVM4MHktUGVNZkU5Y29wMkNBWm9IRVJnYUhucm5yUk93NjNnTkNJTC03bGVZaGRxZG00?oc=5
• Fei-Fei Li sinyali, world model'lerin bilgi riski degil fiziksel risk urettigini vurguluyor; embodied AI safety anlatimi gucleniyor. Kaynak: https://news.google.com/rss/articles/CBMiYkFVX3lxTFAtVkJhMVR0blN1NkQyZTZXQlh4TEdnOXF6YWpvdlFiZzRWWEJVSms2SXBDdG1jSWcyNDd5VnFKVXp4NDc4bF9LcWtWTEtUZG42d1dGU3VydGFNRzZreWo5ZWpR?oc=5
• Philipp Schmid sinyali, agent'larin kaynak kullanmasi ama gereksiz citation gosterimiyle akisi bozmasi arasindaki urun tasarimi gerilimini isaret ediyor. Kaynak: https://news.google.com/rss/articles/CBMiZEFVX3lxTFBldmJOXzk5Q3hYRXpkSjBlcjdEcldIMmdqeTlnWm5SZ1dOT3R5Tzh4aF9EWGgyOWpjaWM5dW9yZ2ZWYTgyNlJ3Y0F1UVM5ZVNULTNqTDVIQlVZWUVGQVZsRGlVYjg?oc=5
• Margaret Mitchell sinyalleri, AI tanimini kimin yaptigi ve AI guvenlik katmanlarinin birbirini untrusted kabul etmesi gerektigi konusunu one cikardi. Kaynaklar: https://news.google.com/rss/articles/CBMiYEFVX3lxTFBLV1FDaUJjQWZhYllaRW1WSmpkS2hYX1FHMVVSZXd0QnBiS212NHA2cngybzJ1ZWEzU01GZ1dsb1NOMDdlZXBnQjQ1TzdEWF8zUFlYWTZRd2t0eVk0U3Rkdw?oc=5 ve https://news.google.com/rss/articles/CBMiZEFVX3lxTE5WQjR0czVYUm5fNzhid0d5QlR2Z0t4c0tIUWRZSmF2QVFQWm83ZXFLUnFEZXIzdmVqRmQtWVlsZXBBR1RFcThCc3JqUzdvY3lLN3JISnNnS3lTRlJoT2JYUm5TaTY?oc=5
• Mastodon tarafinda Threadverse, "I don't know" demeyen modelin daha bilgili degil daha overconfident oldugunu tartisti; eval'larda calibration sinyali izlenmeli. Kaynak: https://mastodon.social/@threadverse/117038508342392615
• Simon Willison, MiniMax-H3 video generation'i local Mac denemesiyle anlatti; multimodal modellerde developer-playground anlatimi sosyalde ilgi topluyor. Kaynak: https://simonwillison.net/2026/Aug/4/minimax-h3-mlx/

Rehberler & Kaynaklar

• Google, 353.000 kisilik AI agents intensive course recap'i yayinladi; agent egitimi kitlesel developer funnel'a donusuyor. Kaynak: https://blog.google/innovation-and-ai/technology/developers-tools/ai-agents-intensive-recap-2026/
• OpenAI, ChatGPT Work ve Codex icin egitim odakli yeni kullanim yollarini duyurdu. Kaynak: https://openai.com/index/learn-teach-chatgpt-work-codex
• InfoQ, AI maturity'nin bes asamasini ve ekiplerin nerede takildigini anlatan sunumu one cikardi. Kaynak: https://www.infoq.com/presentations/ai-sdlc-maturity-framework-bottlenecks/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=AI%2C+ML+%26+Data+Engineering
• Docker/Omdia supply-chain raporu, developer'larin savunmadaki ilk hat oldugunu tekrar gosteriyor. Kaynak: https://www.docker.com/blog/software-supply-chain-security-omdia-2026-report/

Oracle Self-Improvement Sinyalleri

• Memory Reward Inflation ve CrystalMem, Oracle memory skorlamasinda "her deneyimi yuksek onemli kaydetme" egilimini riskli hale getiriyor; memory yazarken decay, negative evidence ve quota politikasini sertlestirmek gerekir. Kaynaklar: https://arxiv.org/abs/2608.00017 ve https://arxiv.org/abs/2608.00303
• AgentStream ve LongHorizon-Harness, cron/agent kalitesinin tek cevapla degil streaming task altinda olculmesi gerektigini gosteriyor. Kaynaklar: https://arxiv.org/abs/2608.00155 ve https://huggingface.co/papers/2608.01964
• RAG-TESTER ve SIRIN, memory-grounded cevaplarda otomatik hallucination/grounding testi eklemek icin dogrudan aday. Kaynaklar: https://arxiv.org/abs/2608.00054 ve https://arxiv.org/abs/2608.00033
• More Debate, Same Evidence, homogeneous multi-agent critique'in fazla guven yaratabilecegini gosteriyor; review agent'lari farkli evidence kaynaklariyla zorlanmali. Kaynak: https://arxiv.org/abs/2608.00243

Kaynak Ozeti

• Toplam raw item: 9.109; input unique: 4.150; raporlanabilir unique: 3.823; distinct unique source: 108. Pipeline metrikleri collector/dedupe ciktisindan hesaplandi.
• RSS/news: 2.232 raw, 256 unique; baskin kaynaklar DonanimHaber, Nikkei Asia, Planet AI. Pipeline metrikleri collector/dedupe ciktisindan hesaplandi.
• Search: 518 raw, 298 unique; Google News AI/companies/security baskin. Pipeline metrikleri collector/dedupe ciktisindan hesaplandi.
• Social: 4.121 raw, 1.434 unique; Twitter fallback, Mastodon ve Bluesky kapsandi. Pipeline metrikleri collector/dedupe ciktisindan hesaplandi.
• Academic/API: 1.599 raw, 1.786 unique; arXiv cs.AI, cs.CV, cs.LG baskin. Pipeline metrikleri collector/dedupe ciktisindan hesaplandi.

Coverage / Blind Spots

• Family coverage tam: rss/news, search, community, social, academic/api 5/5 kapsandi. Pipeline metrikleri collector/dedupe ciktisindan hesaplandi.
• Missing family yok; empty topic yok; thin topic yok. Pipeline metrikleri collector/dedupe ciktisindan hesaplandi.
• RSS tarafinda Microsoft AI Blog 410, Apple ML 404, Tech in Asia 403, Reddit RSS 429/403 gibi engeller goruldu; pipeline alternatif kaynaklarla devam etti.
• Academic tarafta Semantic Scholar 429 ve PapersWithCode JSON parse hatasi var; arXiv/HF yeterli hacim sagladi.
• Search akisi hala Google News/Bing aggregator linklerine dayaniyor; canonical URL cozumleme gelecek run icin kalite artirir.

Bu Gece Sistem Ne Ogrendi

• Onceki artifact'taki "Launches, Regulation, Models" temasi bugun de surdu; fakat launches artik model duyurusu degil agent lifecycle, wallet, CI/CD ve grounding primitive'leri. Kaynaklar: https://blog.cloudflare.com/agent-development-lifecycle/ ve https://blog.cloudflare.com/wallets/ ve https://aws.amazon.com/blogs/machine-learning/introducing-web-search-on-amazon-bedrock-for-foundation-model-grounding/
• Rising entities listesinde Google, AI Agents, Anthropic ve RAG guclu kaldi; bugun Cloudflare ve NVIDIA security/infra katmanlari bu listeye aday oldu. Kaynaklar: https://blog.cloudflare.com/engineering-standards-enforcement/ ve https://blogs.nvidia.com/blog/open-secure-ai-alliance-contributions/
• Kalici ders: Agent sistemlerinde memory, payment, CI ve security ayri is degil; tek governance yuzeyi olarak dusunulmeli. Kaynaklar: https://arxiv.org/abs/2608.00122 ve https://blog.cloudflare.com/wallets/ ve https://snyk.io/blog/why-we-rebuilt-evo-ai-model-risk-scoring/

Dedupe & Kalite Notu

Bu rapordaki maddeler onceki 3 gunun ana tekrarlarindan elendi/dedupe edildi. Merge sonucu 5 dosyada 9.109 giris 7.572 cikisa indi; dedupe sonucu 7.572 giris 4.150 yeni unique kayda indi. Raporlanabilir unique item: 3.823; fp-tekrar: 3.391; semantik-tekrar: 31; guncelleme: 0.