Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Anthropic paused its planned June 15 billing overhaul for the Claude Agent SDK, which would have created a separate credit pool for programmatic usage, after developer backlash and amid reports of OpenAI preparing price …
Z ai's GLM-5.2, a 744B-parameter open weights model with 40B active parameters, achieved a score of 51 on the Artificial Analysis Intelligence Index v4.1, surpassing competitors MiniMax-M3 and DeepSeek V4 Pro. It also le…
Anthropic's Claude Code now features 27 specialized skills, transforming it from a coding assistant into a full AI engineering team capable of handling complex software development tasks autonomously.
Ucp-Local, an open-source offline RAG server, launched as a single-binary MCP tool that indexes local files for use with Claude Desktop, Cursor, and LM Studio. The tool provides hybrid retrieval, tree-sitter code chunkin…
OpenAI published new work on predicting model behavior before release by simulating deployment, creating realistic pressure tests that mimic how users, teams, and attackers might actually use the model. The approach shif…
GitHub launched the Copilot desktop app in technical preview, a control center for agent-native development that lets engineers direct and observe multiple AI coding agents running in parallel. The app uses git worktrees…
A developer modularized their AI assistant's instruction system by splitting a monolithic file into a lean entry point and specialized modules, reducing token costs and improving signal-to-noise ratio. The entry point ac…
A developer created cc-reflection, a framework that uses Claude Code's EDITOR hook to enable structured reflection on coding work. The system captures observations as seeds, which can later be expanded into codebase-grou…
Anthropic analyzed ~400,000 Claude Code sessions from October 2025 to April 2026, finding that domain expertise—not coding proficiency—drives successful agentic coding outcomes. Users with greater expertise achieve highe…
A developer optimized TypeScript type inference in the Sury schema validation library, achieving up to 15.7x faster type extraction by reading a pre-resolved type field instead of matching the entire schema type. The fix…
A developer recommends five categories of MCP servers to connect first for maximum utility: filesystem, code platform, web-fetching, document/notes, and messaging. The filesystem server is highlighted as the most impactf…
GLM-5.2 (Max) ranked second in Code Arena: Frontend, scoring 29 points higher than Claude Opus 4.7 (Thinking) and trailing only Fable 5. The model also achieved top positions in React and HTML sub-leaderboards and led in…
Mistral AI announced plans to release a new, larger family of sparse models this summer, starting with a model expected to delight and surprise in certain capabilities. The company will open an early access program in Ju…
MiniMax released MSA (MiniMax Sparse Attention), a block-sparse attention method that reduces the quadratic cost of softmax attention at long contexts. The method was trained on a 109B-parameter Mixture-of-Experts model …
Appwrite shipped new Auth controls for password strength, made BAA acceptance self-serve for Pro organizations, and covered Anthropic's launch of Claude Fable 5 and Claude Mythos 5. The company also released a May produc…
A developer argues that the first AI rule should be the most frequent correction made during sessions, which for React developers often involves component boundaries, state placement, or naming conventions. Writing down …
Mlx-Code, a new coding agent, natively integrates Git as its core state machine, enabling developers to branch, diff, and resume from any checkpoint. By mapping the reasoning loop onto Git mechanisms, it allows isolated …
A developer draws parallels between coding with Claude Code and playing video games like Civilization, Dota 2, and StarCraft, noting that the loop of prompting and responding in LLM conversations mirrors game mechanics. …
Software engineers at a recent conference expressed fear and grief over the spread of agentic AI systems, and discussed practical responses including token-cost management and cheaper diffusion models for text generation…
The cofounders of Mechanize, Tamay Besiroglu, Matthew Barnett, and Ege Erdil, rebut the argument that current AI progress will likely lead to a misaligned superintelligence causing extinction, claiming the AI safety comm…