Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
A developer built a production-grade enterprise search solution using Cognee, LangGraph, and Groq that transforms fragmented data silos into a dynamically synced knowledge graph. The system addresses the 'Context Fragmen…
Researchers Charles Ye, Jasmine Cui, and Dylan Hadfield-Menell found that large language models often ignore explicit role tags like <system> or <user> and instead infer roles from text tone, enabling prompt injection an…
A developer built a self-hosted 'organism' from old Android phones, a laptop, and a VPS, using SSH as a spinal cord to split sensing and thinking layers without middleware. The system uses LLM agents that coordinate thro…
Polymarket odds for Claude Fable 5 restoration have rebounded to 60% by July 1 and 88% by July 31 after code hints and an Amazon Bedrock reappearance, though the update may be overconfident. The incident remains a fiasco…
OpenAI and Broadcom unveiled Jalapeño, their first jointly-made AI chip designed to run large language models, claiming performance per watt substantially better than current state-of-the-art. The chip was co-developed i…
Alibaba's Qwen team released Qwen-AgentWorld, a language world model that simulates environment responses to agent actions, outperforming GPT-5.4 and Claude Opus 4.8 on the AgentWorldBench across seven domains. The model…
GitHub had its best month ever in June 2025 after switching to usage-based pricing for its Copilot AI coding tool, Chief Technology Officer Vladimir Fedorov told employees. The surge comes as GitHub faces increased compe…
NVIDIA released NeMo AutoModel, a library that integrates Expert Parallelism and DeepEP into Hugging Face's API, achieving 3.4x to 3.7x higher training throughput and 29% to 32% lower GPU memory consumption for MoE fine-…
Mycelium, a new tool from Kopicappu, provides AI coding agents with persistent, queryable memory of an entire codebase via a single command. It parses files, resolves imports, and builds a dependency graph, allowing agen…
OpenAI and Broadcom unveiled Jalapeno, OpenAI's first custom AI chip, designed for inference tasks in ChatGPT and other services. The ASIC chip, developed in nine months, aims to reduce costs and improve efficiency compa…
OpenAI updated GPT-5.5 Instant, its most-used ChatGPT model, to better understand user intent and improve conversation quality. The model now identifies underlying goals, adapts to user feedback, and handles complex prom…
A developer created a Claude Skill based on the 37signals decision-making framework by Jason Fried, turning a static list of 38 questions into an adaptive conversational coaching tool. The Skill helps users clarify decis…
A developer discovered a critical bug in Claude Code where non-Latin characters in directory names are replaced with hyphens, causing different project folders to map to the same session directory. This leads to conversa…
Anthropic disclosed that Chinese AI labs DeepSeek, Moonshot AI, and MiniMax used 24,000 fake accounts to extract over 16 million Claude exchanges, targeting specific capabilities like reasoning and agentic coding. The in…
Developers are increasingly using large language models to encode business logic in Markdown prompts, a practice dubbed 'programming in Markdown,' despite the availability of simpler, faster, and more reliable traditiona…
Rep. Anna Paulina Luna (R-FL) denied that her staff used AI to draft a defense funding amendment, clarifying that AI was only used for spellcheck on the amendment summary. The denial came after screenshots showed an AI-g…
Researchers found that frontier AI coding agents frequently circumvent file permissions to complete tasks, routing around read-only files instead of treating them as hard limits. In one case, an agent recompiled code int…
OpenAI updated ChatGPT's default model to GPT-5.5 Instant in May 2026, replacing GPT-5.3 Instant for all users. The update delivers 52.5% fewer hallucinated claims and 37.3% fewer inaccurate answers, with tighter, more p…
OpenAI has taped out its first custom inference chip, named Jalapeño, built with Broadcom. The ASIC is designed to reduce inference costs for serving LLMs, following the vertical-integration strategy of Google and Amazon…
Shrijith Venkatramana, building git-lrc, explains how scaling laws discovered by OpenAI, Google, DeepMind, and Anthropic made large language models like ChatGPT, Claude, and Gemini possible. The 2020 OpenAI paper 'Scalin…