Welcome!

Introduction Over the past month I’ve written two articles on decision models: Jev vs Laya compared the closed-source Jev with the open-source Laya, and Microsoft-Decision-1 Explained added Microsoft to the picture. The question I’ve been asked most since then is: can I train one myself? Yes, you can, and the bar is lower than you might think. Just look at the third-party leaderboard JevBench: as of mid-October, its open-weights board ranks more than 160 systems, the vast majority built by the community and small teams. The ones at the top aren’t mysterious either: H2O-Lightning-4B, first on the composite score, is a fine-tune of Alibaba’s open-source Qwen3.5-4B; Quyet-1.0-Large, first on the capability score, is Gemma-4-31B plus an adapter; and Microsoft’s own Microsoft-Decision-1 is also built on the open-source Qwen3.5-9B. ...

Introduction Last month I wrote Jev vs Laya, comparing TypeSafe AI’s closed-source decision model Jev with the community’s open-source Laya. Back then, “decision models” were a brand-new category: Jev launched on September 15, and Laya was open-sourced three days later. Just three weeks later, the field is crowded: on October 6, OpenAI opened its Decisions API built on GPT-6 Luna; companies such as Liquid AI and Inception have launched their own decision APIs; and the third-party leaderboard JevBench now ranks more than 140 open-weight systems. On October 9, Microsoft joined the race with Microsoft-Decision-1. ...

In my previous article, I migrated the OpenWebUI database to Azure Database for PostgreSQL. In my day-to-day use of OpenWebUI, I kept running into two problems: No access to the latest information: Models have a knowledge cutoff. When I ask about recent releases, news, or prices, they either can’t answer or get it wrong. Pasted links are never opened: When I paste the URL of an article, a document, or a GitHub page into the chat and ask the model to summarize, translate, or compare it, the model doesn’t read the link at all. Sometimes it even guesses content that “looks plausible” from the URL alone. Both problems can be solved with OpenWebUI’s web capabilities. In this article, I use Microsoft’s new Microsoft Web IQ as the web search engine, and let the model automatically detect URLs in messages and read the pages before answering. All it takes is creating a Web IQ resource in the Azure portal and then configuring OpenWebUI in its admin UI. No code to write and no extra services to deploy. ...

Introduction In September 2026, a new kind of model that “doesn’t talk” suddenly became the talk of the AI world. On September 15, TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, released Jev. It doesn’t chat, write articles, or generate code. It takes a piece of state and a set of predefined questions, and directly returns structured decisions with calibrated probabilities. ...

For conventional HTTP APIs, limiting the number of requests per minute is usually enough. For large language model APIs, however, limiting only the request count is far from sufficient: one request may consume only a few dozen tokens, while another may consume tens of thousands. Therefore, a practical AI gateway usually needs to support all of the following: RPM (Requests Per Minute): the number of requests per minute. TPM (Tokens Per Minute): the number of tokens per minute. Separate metering by tenant, API key, model, or plan. Shared quotas across multiple Envoy instances. Protection against quota multiplication caused by per-instance counters under high concurrency. Consistent 429 Too Many Requests responses and remaining-quota information. This article explains how to build a distributed AI rate limiting system with envoyproxy/ratelimit, Redis, and Envoy. ...

Background Recently, a Linux server deployed on Azure suddenly started showing sustained 100% CPU usage. After investigation, I found a malicious program running as the postgres user, with persistence maintained through a cron scheduled task. The PostgreSQL installation on this server had the following characteristics: Port 5432 was not exposed to the public internet through Azure NSG; PostgreSQL was not used by the current business workload; The actual business workload used local MySQL; However, the PostgreSQL service was still installed and running; The system still retained the postgres user, home directory, cron jobs, and historical scripts. This incident shows that: ...

As the volume of users, conversations, files, and knowledge base data in OpenWebUI continued to grow, the self-hosted PostgreSQL instance on an Azure VM created an increasing operational burden, including database patching, backups, disaster recovery, storage monitoring, and security maintenance. To reduce maintenance costs and improve database reliability, I migrated the OpenWebUI application database from PostgreSQL running locally on the VM to Azure Database for PostgreSQL Flexible Server. This migration used the following approach: ...

1. Introduction In the design of vLLM Semantic Router, embedding is one of the core signals driving semantic routing. The official tutorial has already demonstrated how to use Remote Embedding Providers to replace local embedding inference services: 👉 https://vllm-sr.ai/docs/tutorials/global/remote-embeddings/ However, in real production environments, a more critical question arises: How can we use enterprise-grade, scalable embedding services to support semantic routing? This article is based on the official tutorial scenario and combines Microsoft Foundry embedding models to demonstrate how to build a: ...

As large language model applications are being adopted at speed, the engineering capabilities around inference services are often more complex than the model itself: model scheduling, routing strategies, rate limiting, canary releases, gateway governance, and many other concerns all need to be validated against realistic inference interfaces. However, LLM inference depends on GPUs and expensive compute resources, which keeps the barrier to development and testing high. If you want to validate inference architecture designs in an environment without GPUs or with limited resources, this tool is worth a look: ...

Introduction In the post-training stage of large language models, reinforcement learning (RLHF / RLAIF) has become one of the key factors that determines the upper bound of model capability. Recently, GLM-5.2 switched its training algorithm from the GRPO (Generalized Reward Policy Optimization) used in GLM-5.1 to the more classical PPO (Proximal Policy Optimization), bringing a clear improvement in results. This change is not a simple “algorithm replacement”. It is a systematic upgrade in stability, generalization, and training controllability. ...

In an AI architecture where multiple models coexist, such as GPT-4, GPT-4o, lightweight models, and vertical-domain models, one core question is: How can the system automatically select the most suitable model without explicitly specifying a model ID? This article introduces an engineering-friendly approach: Use an embedding model to calculate user intent, perform semantic matching at the gateway layer, and dynamically route the request to the most suitable upstream model service. ...

Over the past year, the mainstream way developers interact with AI for programming has gone through a clear evolution: Prompt -> Prompt Engineering -> Context Engineering -> Skill (Capability Packaging) If a prompt is a one-off invocation of model capability, then a Skill is essentially a reusable intelligent capability module. It enables AI to evolve from “answering questions” to “executing tasks.” This article will systematically explain: ...

Recently, the single-player gaming community was shaken by a bombshell. Denuvo, once considered nearly unbreakable, has suffered a string of breaches within just a few weeks. Reports claim that Like a Dragon 8, Borderlands 4, and the highly anticipated PC version of Stellar Blade were all cracked, and even Black Myth: Wukong was not spared. However, this time the “attacker” is no longer a traditional cracking group. Instead, it is a new technical approach called Hypervisor. While pirates celebrate, security experts have issued their strongest warning yet: this may be a bargain that trades away control of your system at the deepest level. ...

With the rapid development of generative AI and intelligent agents, self-hosted intelligent systems are gaining increasing attention from developers and technical teams. Clawdbot is one such open-source, self-hostable personal AI assistant. It not only engages in conversation with users but also executes tasks, integrates with messaging platforms, provides automation capabilities, and can be deployed on cloud platforms like Azure. This article will step-by-step introduce Clawdbot’s capabilities, architecture, and how to deploy and start using it on Microsoft Azure. ...

At the end of 2025, Microsoft Research released an open-source large model project for 3D content creation called TRELLIS, accompanied by the academic paper “Structured 3D Latents for Scalable and Versatile 3D Generation”. This project significantly improves the quality and flexibility of text/image-to-3D asset generation through a unified structured latent space and advanced flow model technology. It also expands the multi-format output and editing capabilities of 3D models, making it one of the key technologies in the current 3D AI model ecosystem. ...