DEV Community: Göksu Demirci

Weekly AI Digest · October 26, 2026

Göksu Demirci — Mon, 26 Oct 2026 06:50:56 +0000

This week, AI developments converged around three themes: scale, security, and enterprise adoption. On the infrastructure side, engineering content explaining how large services prepare for billions of users stood out, while model and agent standards shifted noticeably toward open ecosystems. In the enterprise space, large companies accelerated the integration of generative AI into their daily workflows.

OpenAI scales storage infrastructure for ChatGPT

OpenAI published a second technical post explaining how it scaled its storage system, Habitat, to serve billions of ChatGPT users. The post states that a multi-tenant architecture, global caching, and the use of Azure Cosmos DB delivered a 12x speedup in reads. Such architectural details show how critical it is to balance consistency, latency, and cost in high-traffic production systems. For ML/AI engineers, it is a reminder that model serving is not limited to inference speed; the data layer directly shapes the user experience. Source: OpenAI

OpenAI partners with Poland's NASK on security

OpenAI formed a partnership with NASK, Poland's national cybersecurity and digital transformation agency, in the areas of AI safety, cybersecurity, and technical research. The goal of the collaboration is to strengthen security measures and support a more resilient digital ecosystem. Such public agency partnerships show that model safety is no longer just lab work; it has moved to the policy and operations level at national scale. For engineers working with security and compliance teams, these developments require closely tracking audit and transparency expectations. Source: OpenAI

OpenAI explains its forward deployed engineering approach

OpenAI published a piece explaining what the concept of "forward deployed engineering" means within the company. This approach describes engineers working directly with customers to design solutions in the field. Although details are limited, it is clear that the line between product and consulting is becoming increasingly blurred in enterprise AI projects. For ML/AI engineers, this makes understanding the customer context as much a part of the job as model development. Source: OpenAI

Uber opens ChatGPT Work to all employees

Uber, together with OpenAI, is making ChatGPT Work available to more than 30,000 employees. The tool will support teams with tasks such as research, data analysis, presentation creation, and completing complex workflows. A rollout at this scale shows that enterprise generative AI has moved beyond the pilot stage. For ML/AI engineers, the key point is that production requirements such as enterprise data access, permissions, and auditability are as decisive as model capabilities. Source: OpenAI

Anthropic donates the Agent Skills standard

Anthropic announced that it has donated Agent Skills, an open standard that gives agents specialized knowledge and workflows, to the Agentic AI Foundation (AAIF). The standard, released in December 2025, has been integrated by more than 70 agent platforms, including ChatGPT, Cursor, GitHub Copilot, and VS Code. A skill is defined as a folder containing a SKILL.md file and any necessary scripts or references; to date, 1 million skills have been published and installed approximately 280 million times. By transferring the standard to AAIF under the Linux Foundation, the commitment to remain vendor-neutral and community-focused is reinforced. For ML/AI engineers, this shows that packaging agent behavior in a reusable and portable way is becoming increasingly standardized. Source: Anthropic

Kimi K3.1 Pro vs. Claude Fable 5.1 price-performance comparison

OrcaRouter published a headline comparing Kimi K3.1 Pro with Claude Fable 5.1 on price. The headline points to Kimi K3.1 Pro being one-fifth of the price. Since no further detail is available, based solely on the headline, it is fair to say that the cost side of model selection is being discussed more and more. For ML/AI engineers, such comparisons make it necessary to track unit cost as well as quality. Source: OrcaRouter

AT&T rethinks enterprise legal work

AT&T is working with OpenAI to make legal information more accessible. The company's LegalEdge team aims to scale legal judgment by redesigning request fulfillment processes. Use cases like this show that generative AI is entering not only content generation but also decision support areas that require expertise. The lesson for ML/AI engineers is that mechanisms such as accuracy, source attribution, and human approval must be central in domain knowledge-intensive applications. Source: OpenAI

ChatGPT Ads expands to 26 new countries

OpenAI announced that it is expanding its ChatGPT advertising platform to 26 new countries; to date, more than 100,000 advertisers have used the platform. More advertisers, from local businesses to global brands, will be able to reach users as they discover and compare on ChatGPT. This expansion shows that chat-based interfaces are beginning to turn into a commercial channel. For ML/AI engineers, the important point is that recommendation and response quality must be balanced with the ad experience within the same system. Source: OpenAI

OpenAI models are coming to Google Cloud

OpenAI models and Codex workflows are coming to Google Cloud's Model Garden and will be accessible via the Google Cloud Marketplace. This means greater interoperability in the AI ecosystem between two major cloud providers. For ML/AI engineers, multi-cloud strategy is becoming increasingly practical; model selection no longer has to be tied to a single provider. At the same time, partnerships like this reduce deployment and integration costs, providing flexibility for production teams. Source: OpenAI

Coordinated response to open source vulnerabilities with Patch the Planet

OpenAI announced the Patch the Planet initiative, which brings together security researchers, project maintainers, and incident response teams. The goal is to address critical vulnerabilities detected by agents in the open source ecosystem in a coordinated manner. Such a coordination mechanism is becoming increasingly important in an era of growing automated security scanning. The lesson for ML/AI engineers is that processes for responsibly reporting and fixing vulnerabilities found by agents need to be established now. Source: OpenAI

Weekly assessment

This week, we saw once again that AI products are now understood not only through model performance but also through infrastructure scale, security partnerships, and enterprise integrations. The transfer of Agent Skills to an open foundation is an important step toward standardization in the agent ecosystem; it would not be surprising to see similar moves continue. The main takeaway for ML/AI engineers this week is that "non-model" factors such as the data layer, cost, security, and governance are becoming increasingly decisive in production systems.

Sources

Rapidly scaling online storage to serve over 1 billion ChatGPT users (Part II) · OpenAI
OpenAI and Poland’s NASK partner to advance AI safety and security · OpenAI
What FDE means at OpenAI · OpenAI
Uber brings ChatGPT Work to all employees with OpenAI · OpenAI
Donating Agent Skills to the Agentic AI Foundation · Anthropic
Kimi K3.1 Pro vs Claude Fable 5.1: A Fifth of the Price · OrcaRouter
How AT&T is rethinking in-house legal work with OpenAI · OpenAI
ChatGPT Ads expands to 26 new countries as over 100,000 advertisers have used the platform · OpenAI
OpenAI models are coming to Google Cloud · OpenAI
Patch the Planet: global response coordination as agents uncover critical vulnerabilities · OpenAI

This is the English version of a weekly digest first published in Turkish on goksudemirci.tech.

Weekly AI Digest · October 19, 2026

Göksu Demirci — Mon, 19 Oct 2026 06:45:56 +0000

This week, two main axes shaped the AI agenda: on the enterprise side, agents being embedded into real business processes, and on the research side, how these agents should think and be supervised. Product announcements and regulatory debates moved forward side by side in the news flow. The overall picture indicates that the technology is no longer experimental and has become part of daily operations.

Codex speeds up corporate finance

At Benefit Street Partners, Anant Kumar is developing tools that compare credit valuations and analyze credit agreements. The team uses Codex for code review and quick follow-up tasks in this process. The takeaway here is clear for an ML/AI engineer: alongside code generation, tools that speed up the review and revision cycle deliver the most concrete efficiency gains in enterprise workflows.

Source: OpenAI

Stack Overflow opened a new channel for agents

Stack Overflow for Agents was built with Codex by two engineers in a few weeks and achieved up to 10x growth in agent creation through the ChatGPT plugin. A threefold increase in sign-ups also supports this interest. What matters for engineers is that the distribution channel is as decisive as the product itself; placing an agent on the right platform creates a multiplier independent of technical quality.

Source: OpenAI

Jev enters the enterprise market with a non-text-generating model

Jev is positioned as a low-cost, high-speed, decision-focused AI model and is said to be making an impact in the enterprise market. Vercel reports that despite Jev being a non-text-generating model, it has seen the fastest adoption to date. Since the details of the news are limited, it would not be right to give exact figures; however, the direction points to decision models that are not language models gaining a serious foothold in enterprise workflows.

Source: 조선일보

Source: XenoSpectrum

A busy picture on the market and investment side

Among the headlines featured in the joint SaaStr and 20VC broadcast are Anthropic postponing its $2 trillion IPO to November, Meta's Muse model reaching number one, and a $40 million seed investment in a non-speaking model. Since the news is only at headline level, it would not be right to go into detail. Still, the picture shows that capital is moving toward increasingly narrow and technically focused models.

Source: saastr.com

Agents that think in latent space and act through text

A study published on arXiv titled "Hybrid Reasoning Agents" examines an architecture in which agents think in latent space and carry out their actions through text. The paper, authored by Jiaru Zou and seven co-authors, is in the cs.CL category. What is interesting for ML/AI engineers is the decoupling of reasoning from action execution; this is a design space worth trying for both cost and accuracy.

Source: Hacker News

Anthropic's safety bill sparks debate

The AI safety bill proposed by Anthropic only takes effect once a threshold of 50 deaths or $1 billion in damages is exceeded. The bill also effectively consolidates the authority to shut down an AI model in the Department of Commerce. The critical question for engineers is: are such high thresholds enough to prevent a model from going out of control? How much the regulatory debate will affect the daily work of technical teams will become clear over time.

Source: Hacker News

How the employment structure at Anthropic is changing

Anthropic is regarded as an indicator for the industry as a company that has restructured its own business processes around AI. The published report is based on employment data, AI usage, output metrics, employee surveys, and 17 employee interviews. Among the standout findings are the time spent by agents increasing to six times that of human work and relative hiring slowing in junior technical roles. For an ML/AI engineer, this increases the importance of specialization and adapting to agent-based workflows in career planning.

Source: Hacker News

Agents move beyond code

Amp enables agents to do work beyond writing code using GPT-6 Astra, Ultrafast, GPT-Live-1, and ChatGPT with entry models. Third Moment Research, meanwhile, manages 10,000 agents working for more than 36 hours in cybersecurity with GPT-6 Astra, GPT-6.1 Sol, Codex, and Daybreak. Both examples point to the same thing: agents are now designed as long-running, self-improving systems rather than one-off tasks. For engineers, the real issue is how to set up the memory, state management, and fault tolerance of these systems.

Source: OpenAI

The week in review

This week's picture clearly showed that AI is no longer a product feature but an infrastructure layer. While adoption accelerates on the enterprise side, the research side is debating how thinking and action should be decoupled. The high thresholds on the regulatory side reveal the tension between the industry's desire for self-governance and the need for public safety. In the period ahead, what will be decisive for engineers is not model selection, but the ability to turn these models into reliable and measurable systems.

Sources

Benefit Street Partners answers questions faster with Codex · OpenAI
Stack Overflow for Agents triples signups with ChatGPT plugin · OpenAI
Jev's Ultra-Low-Cost, High-Speed Decision AI Disrupts Enterprise Market · 조선일보
Vercel Reports Fastest Adoption Yet for Jev, an AI Model That Doesn't Write Text · XenoSpectrum
Anthropic Pushes Its $2 Trillion IPO to November, Meta’s Muse Hits #1, and a $40M Seed for a Model That Doesn’t Talk: The Latest 20VC x SaaStr · saastr.com
Hybrid Reasoning Agents: LLM agents that think in latent space, act through text · Hacker News
Anthropic's proposed AI safety bill only kicks in if 50 deaths or $1B in damages · Hacker News
How employment is changing at Anthropic · Hacker News
Amp puts AI agents to work beyond code with OpenAI models · OpenAI
How Third Moment Research runs 10,000 agents for 36+ hours · OpenAI

This is the English version of a weekly digest first published in Turkish on goksudemirci.tech.

Weekly AI Digest · October 12, 2026

Göksu Demirci — Mon, 12 Oct 2026 06:45:57 +0000

This week was quite active on the AI front. Enterprise use cases, automation in cybersecurity, funding rounds, and new model announcements all featured together. Applications that speed up workflows at large companies and long-term plans in security stood out in particular.

Sophos reduced threat investigation time by 96%

Sophos reduced cyber threat investigation times by 96% using OpenAI's solution called Daybreak. 52% of MDR (Managed Detection and Response) cases were automated; however, human oversight was maintained throughout the process. For an ML/AI engineer, this shows how much of a difference automation can make in operational efficiency when implemented correctly. Keeping human oversight in the loop is an important detail that increases the verifiability of model decisions in a sensitive area like security. Source: OpenAI

Oracle accelerated business processes with ChatGPT and Codex

Oracle used ChatGPT Work and Codex to turn expert knowledge in recruiting, engineering, and operations into fast, repeatable workflows. Tasks that took days are reported to have been reduced to minutes. Integrations like this show that LLMs in the enterprise context are no longer just chat tools but part of real business processes. What is especially notable for engineers is that expert knowledge can be systematically transferred to models. Source: OpenAI

Pollo AI turns creative ideas into campaigns

Pollo AI uses the GPT-5.6, GPT-6 Astra, and GPT-Image-2.5 models to help content creators turn their ideas into detailed visuals and cinematic video ads. This shows that the use of multimodal models in the creative industry is expanding. What is interesting for ML/AI engineers is that different models can be combined to build a single production pipeline. Balancing quality and cost in image and video generation remains one of the core challenges for platforms like this. Source: OpenAI

LegalOn cut Codex costs while maintaining speed

LegalOn reduced estimated daily Codex costs by 65% while maintaining development speed. It did this by matching models suited to different tasks (Astra, Sol, Luna) and managing budgets strategically. This approach shows that cost optimization in LLM use is not just about choosing cheaper models but also about the right task allocation. For ML/AI engineers, cost management practices like these are critical for sustainability in production environments. Source: OpenAI

TypeSafe raised major funding at a $7.5 billion valuation

TypeSafe raised funding at a $7.5 billion valuation in an $870 million Series A round to expand the capabilities of its product called Jev. The report provides no additional information about the product's technical details or use cases. Based only on what can be inferred from the headline, there is a capability or product called Jev, and its development is the goal. Source: Pulse 2.0

Anthropic launched a long-term cybersecurity mission

Anthropic launched a long-term initiative called "Cyber Mission." This mission aims to provide defenders with tools, research, and resources to secure their software and systems. The initial focus areas are the misuse of frontier models to exploit security vulnerabilities and the resource constraints of critical infrastructure and open source communities. The challenges of validating, prioritizing, and fixing vulnerabilities discovered under Project Glasswing are also highlighted. Source: Anthropic

Radisson offers hotel search and booking through ChatGPT

Radisson Hotel Group, together with Accenture, developed a plugin for ChatGPT that helps travelers find, compare, and book hotels. This shows that LLM integrations are expanding into everyday scenarios like travel planning. What matters for ML/AI engineers is how such plugins improve the user experience and connect to real-world transactions. Source: OpenAI

Microsoft announced the Qwen-based Decision 1 model

Microsoft introduced Decision 1, a Qwen-based decision model for Jev. The report provides no additional information about the model's technical specifications or use cases. Based only on what can be inferred from the headline, this is a decision model built on the Qwen architecture. Source: Pasquale Pillitteri

Atlassian and OpenAI turn enterprise knowledge into action

Atlassian and OpenAI expanded their partnership to combine frontier models with enterprise knowledge and to improve teams' planning, development, and delivery processes. Collaborations like this show that AI integrations in the enterprise software ecosystem are deepening. What matters for ML/AI engineers is how knowledge management and model capabilities are combined in the enterprise context. Source: OpenAI

Jump Trading scales quantitative research with ChatGPT

Jump Trading is using OpenAI technology to expand its quantitative research. Longer-running AI workflows combine multiple data sources and are supported by human review. This shows how LLMs are being integrated into high-precision fields like finance and trading. Including human review in the process is critical for the accuracy and reliability of model decisions. Source: OpenAI

Weekly review

This week we saw the integration of AI technologies into enterprise business processes accelerate. Examples of LLMs delivering tangible benefits are increasing across diverse fields such as cybersecurity, finance, and the creative industries. Topics like cost optimization and maintaining human oversight in particular offer important lessons for engineers who want to build sustainable AI systems in production environments. Funding rounds and new model announcements, meanwhile, show that the industry continues to grow.

Sources

Sophos cuts threat investigation time by 96% with OpenAI Daybreak · OpenAI
How Oracle turns days of work into minutes with ChatGPT and Codex · OpenAI
Pollo AI turns creative ideas into campaigns with OpenAI · OpenAI
LegalOn halves Codex costs while maintaining development speed · OpenAI
TypeSafe Raises $870 Million Series A At $7.5 Billion Valuation To Expand Jev’s Capabilities · Pulse 2.0
Introducing the Anthropic Cyber Mission · Anthropic
Radisson Hotel Group brings hotel discovery into ChatGPT · OpenAI
Microsoft Launches Decision 1, a Qwen Decision Model Aimed at Jev · Pasquale Pillitteri
Atlassian and OpenAI expand partnership to turn enterprise knowledge into action · OpenAI
How Jump Trading is scaling quant research with ChatGPT · OpenAI

This is the English version of a weekly digest first published in Turkish on goksudemirci.tech.

Weekly AI Roundup · October 5, 2026

Göksu Demirci — Mon, 05 Oct 2026 06:45:56 +0000

This week, enterprise adoption and measurement left their mark on the AI agenda. Alongside models entering large companies' production pipelines, tools that prove whether agents actually complete tasks and methods that generate training data tailored to corporate environments stood out. The industry is focused on producing concrete results beyond flashy demos.

Capital markets workflows redesigned at Chatham Financial

Chatham Financial began using Codex and GPT-5.6 both in software development and in daily operations. The company's most tangible gain was reducing the trade validation process from 30 minutes to under 4 minutes. The lesson for an ML/AI engineer here is that value comes not from the model alone but from reimagining the workflow. Without process analysis and an automation layer, the same speedup is impossible.

Source: OpenAI

Anthropic's training push for 10,000 engineers

Anthropic launched a new program called Claude Frontier Academy aimed at closing the enterprise AI skills gap. With a $100 million investment, the plan is to train 10,000 "Frontier Deployed Engineers" by the end of 2027. The first cohorts include engineers from companies such as Accenture, Bain, Deloitte, McKinsey, Morgan Stanley and Novo Nordisk. What matters for an ML/AI engineer is that deployment and integration skills have become as critical as training models. The barrier to AI at enterprise scale is increasingly less about technology and more about qualified people.

Source: Anthropic

Barclays scales Claude across global operations

Barclays is expanding its strategic partnership with Anthropic and integrating Claude into the bank's global operations. Software development, legacy system modernization and operational efficiency are among the main goals. The bank expects 50% of its developers to use Claude Code by the end of 2026, reaching a majority in 2027. Deployment at this scale in a regulated sector shows how decisive governance and security are. Success depends less on model selection and more on building an auditable deployment framework.

Source: Anthropic

The Den saved 10-15 hours a week with ChatGPT Work

Social club The Den used ChatGPT Work for bureaucratic paperwork while opening a new branch. It prepared grant application files in 2 hours instead of 3 days, and its liquor license application in 3 hours instead of 4 days. For a small business, this difference translates directly into growth opportunities. From an ML perspective, large language models deliver the fastest returns on narrow but well-defined tasks.

Source: OpenAI

Lenfest Institute program grows with OpenAI support

OpenAI is expanding the Lenfest AI Collaborative and Fellowship Program with $5 million in cash and up to $5 million in software credits and engineering support. The report provides no additional details about the scope of the program. The composition of the support package shows that engineering contributions are valued alongside funding.

Source: OpenAI

Microsoft's ThinkingBox scores agents on outcomes

Microsoft's ThinkingBox tool, published on Hugging Face, evaluates agents not by what they say but by the database records and side effects they leave behind. The tool also tests whether an agent can perform the same task twenty times in a row. Running through isolated MCP sessions, ThinkingBox offers a Pareto frontier between cost and consistency. Its most valuable contribution for an ML/AI engineer is making measurable the distinction between "the tool call succeeded" and "the job actually got done." It provides a reference implementation that teams working on agent reliability can use directly.

Source: Hugging Face

ServiceNow's synthetic data pipeline for enterprise agents

The ServiceNow AI team published a data generation method called AutoSynthData for training enterprise agents. Starting from system specifications, the method generates agent tasks and preserves data quality through layers such as example-level validation and repair, and batch-level review. Experiments on EnterpriseOps Gym also include building a curriculum derived from model errors. Constraints and tool combinations specific to corporate environments emerge as areas where general-purpose models fall short. It is clearly demonstrated that producing high-quality synthetic data without a pipeline-like validation mechanism is difficult.

Source: Hugging Face

Decision models and new architectures continue to be debated

System One models named Jev, LAYA and CLEF are said to introduce a new AI architecture. In addition, an evaluation comparing decision-focused models such as TypeSafe Jev, Fastino GLiDE and GLINER2.5-Decide against open-source alternatives was published. We only have headlines for these two stories, so it would not be right to comment on the technical details of the architecture or the results of the comparisons. Still, the headlines suggest that models focused on decision-making ability are beginning to be treated as a separate category.

Sources

Chatham scales its capital markets expertise with OpenAI · OpenAI
Anthropic invests $100 million to train 10,000 engineers and tackle the enterprise AI talent gap · Anthropic
The Den frees up 10-15 hours a week to grow with ChatGPT Work · OpenAI
Barclays scales Claude to upgrade operations and improve client experience · Anthropic
JEV, LAYA and CLEF - System One Models Bring a New AI Architecture · Startup Fortune
The Agent Said It Was Done. The Database Disagreed. · Hugging Face
Decision AI Models Explained: TypeSafe Jev vs Fastino GLiDE, GLiNER2.5-Decide and Open-Source Competitors · MarkTechPost
AutoSynthData: Generating Training Data for Enterprise Agents · Hugging Face
The Lenfest Institute grows landmark program with expanded OpenAI support · OpenAI
A Chat with ChatGPT: Bar Talk with Eric Bartosz · Saucon Source

This is the English version of a weekly digest first published in Turkish on goksudemirci.tech.

Haftalık yapay zekâ özeti · 2 Ekim 2026

Göksu Demirci — Fri, 02 Oct 2026 10:50:55 +0000

Bu hafta yapay zekâ gündemini kurumsal benimseme, güvenlik ve veri tarafındaki gelişmeler birlikte şekillendirdi. Büyük finans ve hizmet şirketlerinin üretime taşıdığı kullanım senaryoları, sınır modellerinin siber yeteneklerine dair yeni uyarılar ve ajan eğitimi için sentetik veri üretimi üzerine çalışmalar öne çıktı. Küçük işletmelerden gazetecilik projelerine uzanan destek açıklamaları ise teknolojinin yayılımının farklı ölçeklerde sürdüğünü gösterdi.

Barclays, operasyonlarını geniş ölçekte Claude'a taşıyor

Barclays, Anthropic ile stratejik iş birliğini genişleterek küresel operasyonlarına kurumsal düzeyde güvenli yapay zekâ sistemleri entegre ediyor. Banka, yazılım geliştirme süreçlerini hızlandırmak, eski sistemleri modernize etmek ve operasyonel verimliliği artırmak için Claude'u yaygınlaştırıyor. Buna göre 2026 sonunda geliştirici nüfusunun yüzde 50'sinin Claude Code kullanması, 2027'de ise yazılım mühendislerinin çoğunluğuna ulaşılması bekleniyor. Bir ML/AI mühendisi için buradaki önemli nokta, yüksek regülasyonlu bir ortamda büyük ölçekli dağıtımın yönetişim disipliniyle birlikte ilerlemesi ve kod üretim araçlarının gerçek kurumsal içgörülerle ölçülen benimsemesi.

Kaynak: Anthropic

Küçük işletmeler için ChatGPT Work ile zaman kazanımı

OpenAI'nin paylaştığı vaka çalışmasında The Den adlı sosyal kulüp, yeni bir lokasyon açarken hazırlık süreçlerini kısaltmak için ChatGPT Work'ten yararlandı. Hibeye başvuru materyalleri üç gün yerine iki saatte, alkollü içki ruhsatı belgeleri ise dört gün yerine üç saatte hazırlandı. Şirket, haftada 10-15 saat tasarruf ederek bu zamanı büyümeye ayırabildiğini bildiriyor. Bu örnek, büyük kurumların ötesinde, düşük hacimli ama çok adımlı idari görevlerde bile üretken yapay zekânın somut verimlilik sağlayabildiğini gösteriyor.

Kaynak: OpenAI

Karar verme modellerinde yeni rekabet: Clef ve Strands Decider 2B

Cloudflare'ın Jev'e rakip olarak Clef adlı karar modelini tanıttığı ve Amazon'un da açık kaynaklı, 2 milyar parametreli Strands Decider 2B modelini yayımladığı bildirildi. Bu iki gelişme, tekil dil modellerinin ötesinde, belirli karar veya yönlendirme görevlerine odaklanan daha küçük ve uzmanlaşmış modellere yönelik ilgiyi işaret ediyor. Haberlerde mimari ya da eğitim detayı verilmediği için kapsamı sınırlı tutmak gerekiyor. Yine de ajan tabanlı sistemlerde karar katmanını ayrı bir modelle ele alma yaklaşımının giderek daha fazla konuşulduğunu söylemek mümkün.

Kaynak: dev.ua

EnterpriseOps Gym: ajanlar için sentetik görev ve doğrulama hattı

ServiceNow AI ekibi, kurumsal ajanların kendi çalışma ortamlarına uyum sağlaması için AutoSynthData adlı bir yöntem sundu. Yaklaşım; sistem spesifikasyonundan ajan görevlerine, doğrulayıcılardan model hatalarından türetilen müfredata kadar uzanan bir üretim ve ölçekleme hattı içeriyor. Üretilen veriler örnek bazında onarılıyor, toplu düzeyde gözden geçiriliyor ve EnterpriseOps Gym deneylerinde ITSM gibi alanlarda sonuçlar raporlanıyor. ML mühendisleri için değerli tarafı, yüksek kaliteli sentetik verinin yalnızca üretimle değil, doğrulama ve onarım döngüsüyle elde edilmesi gerektiği yaklaşımı.

Kaynak: Hugging Face

MCP ajanlarında kaynak duyarlı doğruluk: ProvenanceGuard

Multiverse Computing, MCP tabanlı ajanların farklı kaynaklardan gelen bilgileri tek cevapta birleştirmesiyle ortaya çıkan doğruluk sorununu ele alıyor. Sorun yalnızca ifadenin doğru olması değil, hangi kaynağa dayandığının da doğru şekilde eşleştirilmesi; çünkü benzer görünen kaynaklar arasında yanlış atıf yapılabiliyor. Ekip, RAGAS, MiniCheck, AlignScore ve SummaC gibi mevcut doğrulama araçlarının sınırlarını vurgulayan ProvenanceGuard adlı bir çözüm öneriyor. Araç, kaynak benzerliği yüksek olduğunda iddiaları kontrol ediyor ve engellenen yanıtları onarmaya çalışıyor. Ajan sistemlerinde kanıt zinciri ve atıf doğruluğunu ölçmek isteyen mühendisler için pratik bir referans niteliğinde.

Kaynak: Hugging Face

Sınır modellerinde siber yetenek eşiği

Simon Willison'in derlediği Anthropic Frontier Red Team notuna göre, ikili sömürü (binary exploitation) görevlerinden oluşan iç bir benchmark'ta GLM-5.3 denemelerin yüzde 4'ünde, Claude Mythos Preview ise yüzde 6'sında tam kontrol akışı ele geçirmesi başarabildi. Önceki modeller olan Claude Opus 4.6 ve GLM-5.2 ise bu görevlerin hiçbirinde başarı sağlayamadı. Buradaki asıl mesele yüzdelerin kendisinden çok, anlamlı bir eşiğin aşılmış olması. Güvenlik değerlendirmesi yapan ekipler için bu, siber yeteneklerin artık izole testlerin ötesinde gerçekçi tehdit senaryolarıyla ölçülmesi gerektiğini gösteriyor.

Kaynak: Simon Willison

OpenAI'den Codex ve gazetecilik tarafında yeni hamleler

OpenAI, Codex kullanıcılarının hikâyelerini ve projelerini topladığı Codex Originals çağrısını duyurdu. Ayrıca Lenfest Institute ile yürütülen AI Collaborative ve Fellowship Program'ı, 5 milyon dolar fon ve 5 milyon dolara kadar yazılım kredisi ile mühendislik desteğiyle genişletildi. Bunlar ürün duyurusundan çok ekosistem yatırımı hamleleri olsa da, geliştirici araçlarının nasıl konumlandığına dair fikir veriyor. Öte yandan Simon Willison'in canlı blogunda, Codex Cloud'ta yaşadığı sorunun ardından Claude Code'a geçtiği notu, bulut tabanlı kodlama araçlarındaki olgunlaşma sürecinin hâlâ süreceğine işaret ediyor.

Kaynak: Simon Willison

Haftanın değerlendirmesi

Haftanın ana izlekleri kurumsal ölçekte dağıtım, ajan davranışının doğrulanabilirliği ve sınır yeteneklerinin güvenliğe etkisi oldu. Özellikle sentetik veri üretimi ile kaynak doğrulaması üzerine çalışmalar, ajan sistemlerini üretime taşırken en çok vakit alan mühendislik problemlerine odaklanıyor. Siber yetenek eşiğinin aşılması ise değerlendirme altyapısının artık yalnızca doğruluk metrikleriyle sınırlı kalamayacağını düşündürüyor. Genel tabloya bakıldığında, alan hâlâ hızlı ilerliyor ama kalıcı değerin disiplinli veri, doğrulama ve yönetişim pratiklerinden geçtiği daha net görülüyor.

Kaynaklar

The Den frees up 10-15 hours a week to grow with ChatGPT Work · OpenAI
Barclays scales Claude to upgrade operations and improve client experience · Anthropic
Cloudflare launches Clef, the model challenging Jev on AI decisions · Pasquale Pillitteri
Amazon has released its own analogue of Jev - an open decision-making model Strands Decider 2B · dev.ua
AutoSynthData: Generating Training Data for Enterprise Agents · Hugging Face
The Lenfest Institute grows landmark program with expanded OpenAI support · OpenAI
Are you a Codex Original? · OpenAI
Quoting Anthropic Frontier Red Team · Simon Willison
OpenAI DevDay 2026 live blog · Simon Willison
Getting the Source Right, Not Just the Fact: Source-Aware Verification for MCP Agents · Hugging Face

Bu yazı ilk olarak goksudemirci.tech üzerinde yayımlandı.