Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Grant Sanderson, creator of 3Blue1Brown, discussed the rapid progress of AI in mathematics and its implications for other fields in a podcast interview. He explored whether AI discoveries constitute AGI, the nature of co…
The Unjournal launched a suite of tools and resources to help researchers prioritize high-impact research questions, including a Research Prioritization Dashboard, a Cruxes & Pivotal Questions Explorer, and evaluation gu…
Anthropic's Claude Code tool is embedding steganographic markers in its requests, a technique that allows the company to detect unauthorized use or copying of its AI models. The discovery raises concerns about privacy an…
LLM KOSH, a local-first memory cartridge for MCP-compatible AI clients, has been released. The tool provides durable, inspectable memory storage without relying on hosted services, featuring a read-only MCP server, CLI, …
Context engineering has emerged as a critical discipline in AI development, moving beyond prompt engineering to manage an AI's working memory. Developers are grappling with the challenge that models lose focus as their c…
X launched a hosted Model Context Protocol server that lets AI tools like Claude and Cursor connect directly to its API using user account permissions, eliminating the need for developers to build custom integrations. Th…
Operational readiness for LLM services requires re-deriving classical engineering primitives like alarms, load testing, and retry policies, as defaults tuned for traditional APIs or databases fail for LLM workloads. The …
Skillware v0.4.0 introduces a new token limiter skill that allows LLM agents to monitor and cut their own token spend. The skill acts as a budget gate, returning actions like CONTINUE, WARN, or FORCE_TERMINATE based on c…
Meta launched its AI assistant in Israel with native Hebrew support across WhatsApp, desktop, and mobile apps, enabling content creation, image analysis, and reasoning tasks. The rollout leverages Meta's Llama model and …
A user on Hacker News reports that OpenAI's Codex with GPT-5.5 Extra High has degraded in quality, comparing the shift to replacing a reliable tech lead with an overconfident junior developer. The user claims the model n…
A library of 394 free Claude skills claims 'all quality-tested, mean 4.38/5', backed by a seven-dimension evaluation framework. Skills must pass binary assertions and a graded rubric using a model-as-judge to reach 'stab…
Meituan trained its 1.6 trillion parameter LongCat-2.0 AI model entirely on Chinese chips, demonstrating China's ability to develop large-scale AI without Nvidia hardware.
MIT-affiliated researchers Charles Ye, Jasmine Cui, and Dylan Hadfield-Menell demonstrated that attackers can trick frontier large language models into providing dangerous instructions, such as cocaine synthesis, by inse…
Google's Gemini 2.5 Pro with Deep Think reasoning mode topped coding and reasoning benchmarks this week, scoring 82.4% on GPQA Diamond and 94.1% on HumanEval+, but the mode multiplies token costs by roughly 4x and underp…
A developer exploring AI platform engineering found that the most challenging problems lie in distributed systems and scheduling rather than machine learning itself. The analysis of GPUs, Ray, vLLM, and Kubernetes reveal…
Microsoft introduced the Microsoft Binlog MCP Server, giving AI assistants like GitHub Copilot direct access to MSBuild (.binlog) files for natural language build investigation. The server, currently in preview, parses b…
OpenAI Whisper has 12 unique checkpoint files, but the Python package exposes 14 local model names due to aliases for 'large' and 'turbo'. Additional confusion arises from runtimes and converted formats like faster-whisp…
Looop, a new open-source tool, launches as a tiny, portable autonomous control loop for LLM agents, operating as a single binary without external dependencies. It monitors sources like GitHub and Linear, runs worker agen…
A developer launched AskPG, a real-time WebGPU-powered AI avatar of Paul Graham that provides startup advice by applying Graham's ideas to users' specific situations. The tool supports text and voice input, offers sharea…
A developer built a retrieval-augmented generation (RAG) system for code and found that chunking by function boundaries dramatically outperformed line-based chunking. By using a parser to extract complete functions, meth…