Training-Inference Mismatch: Why Your LLM Reinforcement Learning Is Optimizing the Wrong Policy
Reinforcement learning has become the defining ingredient of modern LLM post-training. GRPO, PPO, and their variants drive the reasoning capabilities
Tag: LLM
GLM-5.2 and Tencent Hy3: Two Different Bets on the Open-Weight Frontier
The open-weight frontier has been moving fast. Over the past few weeks, two major releases have landed on HuggingFace that
Continue readingGLM-5.2 and Tencent Hy3: Two Different Bets on the Open-Weight Frontier
Program-as-Weights: Compiling Natural Language Into Local Neural Programs
There’s a class of programming tasks that resists clean implementation: deciding whether a log line is “important,” repairing malformed JSON
Continue readingProgram-as-Weights: Compiling Natural Language Into Local Neural Programs
vLLM v0.23.0: Model Runner V2, Multi-Tier KV Offloading, and the Growing Rust Frontend
The vLLM v0.23.0 release landed last week with 408 commits from 200 contributors, and it packs several changes that directly
LoopCoder-v2: Why Two Loops Beat Four in Test-Time Compute Scaling
The dominant scaling narrative in large language models has been straightforward: more parameters, more data, more compute. But there’s a
Continue readingLoopCoder-v2: Why Two Loops Beat Four in Test-Time Compute Scaling
GLM-5.2: The New #1 Open-Weight LLM and Why IndexShare Matters
The open-source LLM landscape just got a new heavyweight contender. Z.ai (Zhipu AI) released GLM-5.2, a 753B-parameter mixture-of-experts model that
Continue readingGLM-5.2: The New #1 Open-Weight LLM and Why IndexShare Matters
How MiniMax Sparse Attention Achieves 28x Compute Reduction at 1M Context Length
The attention mechanism is the backbone of every transformer model, but it carries a brutal cost: quadratic complexity with respect
Continue readingHow MiniMax Sparse Attention Achieves 28x Compute Reduction at 1M Context Length
5 Trending GitHub Repos: Apple’s Container Runtime Hits 1.0, LLM Token Compression, and AI Skill Security
The GitHub trending page this week is dominated by AI agent tooling, but tucked between the skills and plugins are
Microsoft’s MAI Models at Build 2026: Seven New AI Models and What They Mean for Developers
Microsoft’s Build 2026 conference delivered a move that had been anticipated for months but still landed with weight: the company
SkillOpt: Training AI Agent Skills Like Neural Networks
AI agents have a skill problem. You give a language model a system prompt — or “skill” — and it
Continue readingSkillOpt: Training AI Agent Skills Like Neural Networks