Oracle Gece Arastirma — 2026-04-12

Curate eden: Mahsum Aktaş · Günlük otomatik AI sektör taraması

Oracle Gece Arastirma — 2026-04-12

Otomatik derleme | v3 pipeline | 88 kaynak | 1515 benzersiz item

Gunun Ozeti

Bugunun ana resmi net: AI ekosistemi yeniden uc yone ayrildi, agent runtime ve memory, supply-chain/security, on-device ve local inference. Anthropic tarafinda Claude Code icin cloud-planlama ve Word entegrasyonu, agent urunlerinin daha kurumsal yuzeylere aktigini gosteriyor. Kaynaklar: https://the-decoder.com/claude-codes-new-ultraplan-feature-moves-task-planning-to-the-cloud/ | https://www.technopat.net/2026/04/11/claude-microsoft-word-entegrasyonunu-duyurdu/

Ayni anda guvenlik hatti sertlesiyor. Axios developer tool compromise sonrasi OpenAI’nin resmi incident yaniti, Trivy/Axios vakalarinin supply-chain savasini yeni normale cevirdigini teyit ediyor. Kaynaklar: https://openai.com/index/axios-developer-tool-compromise | https://go.theregister.com/feed/www.theregister.com/2026/04/11/trivy_axios_supply_chain_attacks/

Ucuncu eksen ise local ve edge. Gemma 4’un telefonda agentic/on-device anlatisi, Apple Silicon uzerinde speculative decoding denemeleri ve SQLite 3.53.0 gibi altyapi guncellemeleri, “kucuk ama hizli ve yakin” stack’in yeniden guclendigini gosteriyor. Kaynaklar: https://the-decoder.com/googles-gemma-4-puts-free-agentic-ai-on-your-phone-and-no-data-ever-leaves-the-device/ | https://www.reddit.com/r/LocalLLaMA/comments/1simszl/dflash_speculative_decoding_on_apple_silicon_85/ | https://simonwillison.net/2026/Apr/11/sqlite/#atom-everything

Trend Analizi

7 gun trendinde xAI yukselis, Claude 4 spike, AI Safety dusus, AI Regulation dusus gorunuyor. Bugunku akis bu tabloyu kismen dogruluyor: Claude ekosistemi cloud planning, managed agents ve Office entegrasyonu ile momentumunu koruyor. Kaynaklar: https://the-decoder.com/claude-codes-new-ultraplan-feature-moves-task-planning-to-the-cloud/ | https://www.reddit.com/r/artificial/comments/1si5evu/anthropic_launches_claude_managed_agents/

AI Safety ve Regulation tarafi ise tamamen kaybolmadi, ama ana hikaye buyuk yasa/policy degil, daha cok supply-chain, data access ve guvenilirlik katmanina kaydi. Kaynaklar: https://openai.com/index/axios-developer-tool-compromise | https://go.theregister.com/feed/www.theregister.com/2026/04/10/ai_agents_spider_man_snowflake/ | https://the-decoder.com/when-ai-models-cant-see-they-just-make-something-up/

Local model ve infra tarafinda ise belirgin yukselis var. GLM/Gemma/Qwen etrafindaki topluluk akisina Apple Silicon, SSD offload ve template hygiene tartismalari eklendi. Kaynaklar: https://huggingface.co/unsloth/GLM-5.1-GGUF | https://www.reddit.com/r/LocalLLaMA/comments/1siug5y/analysis_of_spilling_moe_weights_onto_ssd_glm5_is/ | https://www.reddit.com/r/LocalLLaMA/comments/1sis1vn/the_definitive_qwen_35_jinja_template/

Top 7

1. OpenAI, Axios developer tool compromise olayina resmi yanit verdi. Bu, AI tooling supply-chain riskinin artik ana akim vendor response seviyesine ciktigini gosteriyor. Kaynak: https://openai.com/index/axios-developer-tool-compromise
2. Trivy ve Axios vakalari birlikte, open-source tool zincirinin yeni saldiri yuzeyi oldugunu kanitladi. SBOM ve signing artik opsiyon degil. Kaynak: https://go.theregister.com/feed/www.theregister.com/2026/04/11/trivy_axios_supply_chain_attacks/
3. Claude Code Ultraplan, planlamayi cloud’a tasiyor. Agent runtime’larda terminalden ayrisan “planner layer” standarda donusebilir. Kaynak: https://the-decoder.com/claude-codes-new-ultraplan-feature-moves-task-planning-to-the-cloud/
4. Gemma 4’un on-device agent anlatisi local AI’yi yeniden ciddi kategoriye tasidi. Telefon ustunde multimodal + tool use fikri artik demo degil urun anlatisi. Kaynak: https://the-decoder.com/googles-gemma-4-puts-free-agentic-ai-on-your-phone-and-no-data-ever-leaves-the-device/
5. MolmoWeb ve ClawBench gibi acik benchmark/agent calismalari gunluk dijital gorevleri merkeze aldi. Web ve everyday-task automation tarafinda acik veri ivmesi var. Kaynaklar: https://huggingface.co/papers/2604.08516 | https://huggingface.co/papers/2604.08523
6. DFlash’in Apple Silicon uzerinde 3.3x speculative decoding iddiasi, local inference hiz tartismasini yeniden atesledi. Kaynak: https://www.reddit.com/r/LocalLLaMA/comments/1simszl/dflash_speculative_decoding_on_apple_silicon_85/
7. Snowflake’in “Spider-Man theory of AI agent data access” anlatimi, asil darboğazin model degil veri yetkilendirmesi oldugunu acik sekilde ortaya koydu. Kaynak: https://go.theregister.com/feed/www.theregister.com/2026/04/10/ai_agents_spider_man_snowflake/

KAT-1

• OpenAI incident response, geliştirici araci ekosistemindeki imza, sertifika ve dagitim guvenliginin dogrudan marka riski oldugunu gosterdi. Kaynak: https://openai.com/index/axios-developer-tool-compromise
• Project Glasswing, kritik yazilim ve AI doneminde savunma odakli tedarik zinciri vizyonunu guclendiriyor. Kaynak: https://www.anthropic.com/glasswing

KAT-2

• Claude Code Ultraplan, execution ile planning katmanini ayirarak agent workflow’larinda yeni bir desen oneriyor. Kaynak: https://the-decoder.com/claude-codes-new-ultraplan-feature-moves-task-planning-to-the-cloud/
• Claude for Word, AI’nin bagimsiz uygulama degil mevcut office yuzeylerine gomulen servis oldugunu teyit ediyor. Kaynak: https://www.technopat.net/2026/04/11/claude-microsoft-word-entegrasyonunu-duyurdu/

KAT-3

• Managed agents anlatisi, state, sandboxing ve credentials yonetimini urun seviyesinde satan yeni katmani guclendiriyor. Kaynak: https://www.reddit.com/r/artificial/comments/1si5evu/anthropic_launches_claude_managed_agents/
• Multica, open-source managed agents platformu olarak “agent teammate” pazarini hedefliyor. Kaynak: https://github.com/multica-ai/multica

KAT-4

• Gemma 4 on-device vizyonu, telefon icinde multimodal ve gizlilik dostu inference talebinin arttigini gosteriyor. Kaynak: https://the-decoder.com/googles-gemma-4-puts-free-agentic-ai-on-your-phone-and-no-data-ever-leaves-the-device/
• AMD GAIA masaustu agent builder hamlesi, desktop-local agent deneyiminin ticari urune donustugunu gosteriyor. Kaynak: https://www.reddit.com/r/artificial/comments/1sitbvu/amds_gaia_now_allows_building_custom_ai_agents/

KAT-5

• SkillClaw, skill’lerin statik prompt dosyasi degil kolektif evrilen agent varliklari olarak ele alinmasi gerektigini savunuyor. Kaynak: https://huggingface.co/papers/2604.08377
• Graph of Skills, buyuk skill havuzlarinda dependency-aware retrieval ihtiyacini one cikariyor. Kaynak: https://huggingface.co/papers/2604.05333

KAT-6

• SQLite 3.53.0, hafif veri katmaninin AI tooling icinde yeniden one ciktigini gosteriyor. Kaynak: https://simonwillison.net/2026/Apr/11/sqlite/#atom-everything
• SQLite Query Result Formatter, local analysis tooling tarafinda UX iyilestirmelerinin hiz kazandigini gosteriyor. Kaynak: https://simonwillison.net/2026/Apr/11/sqlite-qrf/#atom-everything

KAT-7

• ClawBench, AI agent’larin gunluk online gorevleri ne kadar gercekci sekilde tamamlayabildigini olcen yeni benchmark olarak dikkat cekiyor. Kaynak: https://huggingface.co/papers/2604.08523
• KnowU-Bench, mobil agent’larda kisilestirme ve proaktif yardim kalitesini olcmeye odaklaniyor. Kaynak: https://huggingface.co/papers/2604.08455

KAT-8

• HY-Embodied-0.5, embodied foundation model cizgisini daha kullanisli bir urun ailesine ceviriyor. Kaynak: https://huggingface.co/papers/2604.07430
• ViVa, robot reinforcement learning icin video-generative value modeling cizgisini guclendiriyor. Kaynak: https://huggingface.co/papers/2604.08168

KAT-9

• QEIL v2, edge intelligence tarafinda enerji, kalite ve guvenilirligi ayni anda optimize etmeye calisiyor. Kaynak: https://huggingface.co/papers/2602.06057
• Flux Attention, uzun baglamli inference icin context-aware hybrid attention ile verimlilik kazanci hedefliyor. Kaynak: https://huggingface.co/papers/2604.07394

SECURITY

• Axios compromise sonrasi OpenAI’nin sertifika rotasyonu ve guncellemeleri, vendor-level hygiene’in zorunlu hale geldigini gosterdi. Kaynak: https://openai.com/index/axios-developer-tool-compromise
• Trivy ve Axios saldirilari, supply-chain kompromisinin gelecegin en tehlikeli open-source tehdidi oldugunu acik sekilde isaret ediyor. Kaynak: https://go.theregister.com/feed/www.theregister.com/2026/04/11/trivy_axios_supply_chain_attacks/
• Hungarian government creds sizintisi, devlet hesaplarinda temel credential hijyeninin hala zayif halka oldugunu gosteriyor. Kaynak: https://go.theregister.com/feed/www.theregister.com/2026/04/11/hungary_government_logins_breach/

REGULATION

• xAI’nin Mississippi power plant izni sonrasi yeni itirazlarla karsilasmasi, AI compute buyumesinin enerji-politik cephede daha fazla surtusme uretecegini gosteriyor. Kaynak: https://www.cnbc.com/2026/04/10/musks-xai-draws-more-opposition-over-mississippi-power-plant-permit.html
• Vercel ToS guncellemesi, AI urunleri etrafinda platform kosullarinin daha agresif yeniden yazildigi bir doneme girildigini hatirlatiyor. Kaynak: https://vercel.com/changelog/updates-to-terms-of-service-march-2026

AI-SCIENCE

• ImplicitMemBench, explicit recall yerine davranissal/implicit memory tarafini olcmeye odaklanarak agent memory literaturune yeni aci getiriyor. Kaynak: https://huggingface.co/papers/2604.08064
• Master Key Hypothesis, capability transfer’i linear subspace alignment ile ele alarak model arasi bilgi aktarimini yeniden cerceveliyor. Kaynak: https://huggingface.co/papers/2604.06377
• Rethinking Generalization in Reasoning SFT, SFT genellemez ezberler anlatisini daha kosullu bicimde sorguluyor. Kaynak: https://huggingface.co/papers/2604.06628

INFRA

• TriAttention, KV cache compression ile full attention kalitesine yakin kalip throughput artisi hedefliyor. Kaynak: https://www.marktechpost.com/2026/04/11/researchers-from-mit-nvidia-and-zhejiang-university-propose-triattention-a-kv-cache-compression-method-that-matches-full-attention-at-2-5x-higher-throughput/
• Cloudflare browser rendering + CDP/MCP anlatisi, browser automation’in yeni altyapi katmanina donustugunu gosteriyor. Kaynak: https://www.reddit.com/r/artificial/comments/1sieggh/cloudflare_just_turned_browser_rendering_into_a/
• Snowflake data-access vizyonu, agent stack’in merkezindeki asıl problemin veri erisim governance’i oldugunu soyluyor. Kaynak: https://go.theregister.com/feed/www.theregister.com/2026/04/10/ai_agents_spider_man_snowflake/

SAFETY

• ProactiveBench bulgulari, modellerin eksik gorsel bilgi oldugunda yardim istemek yerine uydurmaya daha yatkin oldugunu gosteriyor. Kaynak: https://the-decoder.com/when-ai-models-cant-see-they-just-make-something-up/
• AI companion oyuncaklar etrafindaki hikayeler, duygusal baglanma ve yanlis bilgi riskinin consumer AI’da buyudugunu gosteriyor. Kaynak: https://www.theverge.com/ai-artificial-intelligence/910008/fawn-friends-ai-companion
• Bullshit detector erozyonu, AI-gorsel ve sentetik medya caginda dogrulama sistemlerinin geride kaldigini vurguluyor. Kaynak: https://www.wired.com/story/how-the-internet-broke-everyones-bullshit-detectors/

WATCHLIST

• Waypoint-1.5, consumer hardware uzerinde AI-generated 3D worlds anlatisini urune cevirebilir. Kaynak: https://the-decoder.com/overworlds-waypoint-1-5-brings-ai-generated-3d-worlds-to-mac-and-windows-on-consumer-hardware/
• VimRAG, memory graph kullanan multimodal RAG cizgisi icin izlenmesi gereken bir framework. Kaynak: https://www.marktechpost.com/2026/04/10/alibabas-tongyi-lab-releases-vimrag-a-multimodal-rag-framework-that-uses-a-memory-graph-to-navigate-massive-visual-contexts/
• Rowboat, email + notes + knowledge graph baglantisini tek open-source coworker urununde topluyor. Kaynak: https://github.com/rowboatlabs/rowboat

TRENDS

• Agent runtime katmanlasiyor. Planner, memory, sandbox ve execution tek urun yerine ayrik ama bagli modul haline geliyor. Kaynaklar: https://the-decoder.com/claude-codes-new-ultraplan-feature-moves-task-planning-to-the-cloud/ | https://github.com/multica-ai/multica | https://blog.langchain.com/your-harness-your-memory/
• Memory, tek ozellik degil rekabet alani oldu. Coding assistant, RAG ve agent harness’lerde kalici baglam ana farklastirici olarak one cikiyor. Kaynaklar: https://towardsdatascience.com/why-every-ai-coding-assistant-needs-a-memory-layer/ | https://blog.langchain.com/your-harness-your-memory/ | https://huggingface.co/papers/2604.08064
• Local-first geri donuyor. Apple Silicon, Gemma 4, Qwen template ve SSD offload tartismalari pahali cloud inference’a net alternatif uretiyor. Kaynaklar: https://www.reddit.com/r/LocalLLaMA/comments/1simszl/dflash_speculative_decoding_on_apple_silicon_85/ | https://the-decoder.com/googles-gemma-4-puts-free-agentic-ai-on-your-phone-and-no-data-ever-leaves-the-device/ | https://www.reddit.com/r/LocalLLaMA/comments/1siug5y/analysis_of_spilling_moe_weights_onto_ssd_glm5_is/

CikCik Paketi

• DFlash on Apple Silicon tartismasi, local inference hizlanmasinin toplulukta ciddi heyecan urettigini gosteriyor. Kaynak: https://www.reddit.com/r/LocalLLaMA/comments/1simszl/dflash_speculative_decoding_on_apple_silicon_85/
• Qwen 3.5 icin “definitive Jinja template” paylasimi, local stack’te template correctness’in performans kadar kritik oldugunu gosteriyor. Kaynak: https://www.reddit.com/r/LocalLLaMA/comments/1sis1vn/the_definitive_qwen_35_jinja_template/
• GLM-5 SSD spill analizi, buyuk MoE modellerin eksik VRAM ile bile kullanilabilir kalabildigine dair pratik bir sinyal verdi. Kaynak: https://www.reddit.com/r/LocalLLaMA/comments/1siug5y/analysis_of_spilling_moe_weights_onto_ssd_glm5_is/
• Cloudflare browser rendering -> MCP altyapisi paylasimi, browser-use agent’larda remote CDP fikrini yeniden gundeme tasidi. Kaynak: https://www.reddit.com/r/artificial/comments/1sieggh/cloudflare_just_turned_browser_rendering_into_a/
• Claude Code x n8n tartismasi, low-code otomasyon ile coding-agent workflow’larinin birlesme istahini gosteriyor. Kaynak: https://www.reddit.com/r/artificial/comments/1sik4by/claude_code_x_n8n/

Oracle Self-Improvement Sinyalleri

• Unique veri guclu ama category_hint asiri genel. Ozellikle akademik item’lar CAT-04 altina yigildigi icin ikinci semantic pass gerekli. Kaynaklar: https://huggingface.co/papers/2604.08377 | https://huggingface.co/papers/2604.08064 | https://huggingface.co/papers/2604.08516
• Search/source karisimi wrapper agirlikli. Google News ve community linklerinden canonical URL cozumleme pipeline’a eklenirse rapor kalitesi yukselir. Kaynaklar: https://www.cnbc.com/2026/04/10/musks-xai-draws-more-opposition-over-mississippi-power-plant-permit.html | https://www.reddit.com/r/artificial/comments/1si5evu/anthropic_launches_claude_managed_agents/
• Memory ve agent trendi kalici. Son gunlerde tekrar eden pattern, Oracle tarafinda memory-first siniflandirma ve agent runtime scoring’i agirlastirmayi gerektiriyor. Kaynaklar: https://blog.langchain.com/your-harness-your-memory/ | https://towardsdatascience.com/why-every-ai-coding-assistant-needs-a-memory-layer/ | https://huggingface.co/papers/2604.08064
• Topluluk akisi erken sinyal uretiyor ama gürültulu. Reddit/LocalLLaMA post’lari productizasyon once erken performans ve template sinyali veriyor, bu akisa ayri agirlik mantigi faydali olur. Kaynaklar: https://www.reddit.com/r/LocalLLaMA/comments/1simszl/dflash_speculative_decoding_on_apple_silicon_85/ | https://www.reddit.com/r/LocalLLaMA/comments/1sis1vn/the_definitive_qwen_35_jinja_template/

Kaynak Ozeti

• Dogrudan haber/blog sinyali en cok OpenAI, The Register, The Decoder, Simon Willison ve Wired tarafindan tasindi. Kaynaklar: https://openai.com/index/axios-developer-tool-compromise | https://go.theregister.com/feed/www.theregister.com/2026/04/11/trivy_axios_supply_chain_attacks/ | https://the-decoder.com/claude-codes-new-ultraplan-feature-moves-task-planning-to-the-cloud/ | https://simonwillison.net/2026/Apr/11/sqlite/#atom-everything | https://www.wired.com/story/how-the-internet-broke-everyones-bullshit-detectors/
• Academic akis agent benchmark, multimodal reasoning, memory ve embodied AI ekseninde yogunlasti. Kaynaklar: https://huggingface.co/papers/2604.08523 | https://huggingface.co/papers/2604.08516 | https://huggingface.co/papers/2604.08064 | https://huggingface.co/papers/2604.07430 | https://huggingface.co/papers/2604.08168
• Community akis local inference, template hygiene, browser automation ve managed agents ekseninde one cikti. Kaynaklar: https://www.reddit.com/r/LocalLLaMA/comments/1simszl/dflash_speculative_decoding_on_apple_silicon_85/ | https://www.reddit.com/r/LocalLLaMA/comments/1sis1vn/the_definitive_qwen_35_jinja_template/ | https://www.reddit.com/r/artificial/comments/1sieggh/cloudflare_just_turned_browser_rendering_into_a/ | https://www.reddit.com/r/artificial/comments/1si5evu/anthropic_launches_claude_managed_agents/

Coverage / Blind Spots

• Coverage durumu iyi: 5/5 family ve 10/10 topic kapsandi. Kaynak: /Users/mahsum/clawd/research/nightly-v3/raw-sources/2026-04-12/unique.json
• Eksik family yok, thin family yok, empty topic yok. Bu, veri toplama tarafinin bugun dengeli calistigini gosteriyor. Kaynak: /Users/mahsum/clawd/research/nightly-v3/raw-sources/2026-04-12/unique.json
• Baskin kaynaklar: social family icinde Twitter fallback; search family icinde Google News; academic family icinde HF Daily Papers baskin. Bu, wrapper/canonical riskini artiriyor. Kaynak: /Users/mahsum/clawd/research/nightly-v3/raw-sources/2026-04-12/unique.json

Bu Gece Sistem Ne Ogrendi

• Kalici pattern degismedi: launches, regulation, models, agents, tooling hala ana tema. Ama bugun bunlarin icinde en guclu alt tema memory + runtime + local inference oldu. Kaynaklar: https://blog.langchain.com/your-harness-your-memory/ | https://the-decoder.com/claude-codes-new-ultraplan-feature-moves-task-planning-to-the-cloud/ | https://the-decoder.com/googles-gemma-4-puts-free-agentic-ai-on-your-phone-and-no-data-ever-leaves-the-device/
• Learning artifact’taki “Google, AI Agents, Anthropic yukseliyor” dersi bugun de calisti. Google local/on-device, Anthropic runtime/productization, agents ise benchmark + platform katmaninda tekrar merkeze geldi. Kaynaklar: https://the-decoder.com/googles-gemma-4-puts-free-agentic-ai-on-your-phone-and-no-data-ever-leaves-the-device/ | https://www.technopat.net/2026/04/11/claude-microsoft-word-entegrasyonunu-duyurdu/ | https://huggingface.co/papers/2604.08523
• Search akisinin canonical zayifligi suruyor. En temiz sinyal yine dogrudan blog ve paper linklerinden geldi. Kaynaklar: https://openai.com/index/axios-developer-tool-compromise | https://simonwillison.net/2026/Apr/11/sqlite/#atom-everything | https://huggingface.co/papers/2604.08064

Dedupe & Kalite Notu

Bu rapordaki tum maddeler onceki 3 gunun raporlarindan elendi/dedupe edilmistir. Toplam 1515 item islendi, secili olarak 45+ benzersiz sinyal kullanildi.