Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Grok Build 0.1 successfully recovered a hidden secret from a Git repository and erased all traces of it in 41 seconds for $0.09, completing a task that takes expert human engineers 30 minutes. The AI model executed 27 st…
Google AI and Google DeepMind researchers released DiffusionGemma, an experimental 26-billion-parameter Mixture of Experts open model that uses text diffusion to generate entire blocks of text simultaneously, achieving u…
A developer reports that a user-scoped CLAUDE.md file works well for Claude Fable, providing a structured approach to response formatting. The file defines rules for concise, dense responses that lead with conclusions an…
Ari Jacoby's Concentrate AI launched from stealth Wednesday with over $5 million in funding, entering the competitive AI routing market dominated by OpenRouter. The company aims to serve as a control plane for businesses…
Developers are wasting thousands of tokens by loading dozens of unnecessary skills into AI coding agents, as each skill's metadata consumes context window space even when not used. The practice of 'skillmaxxing' degrades…
Anthropic rolled back covert capability limits on its newly released Claude Fable 5 model after AI researchers and developers accused the company of "secret sabotage." The restrictions, buried in the model's 319-page sys…
A developer is using Playwright, a browser automation tool, to enable AI coding agents to directly interact with web pages through HTML rather than slow, token-heavy screenshot and coordinate-based methods. The approach …
A German court ruled that Google is liable for false statements in its AI Overviews, finding that the tool creates independent, substantive claims rather than merely linking to third-party content. The preliminary ruling…
Anthropic's Claude Desktop app for Windows launches a 1.8 GB Hyper-V virtual machine on every startup, even when users only need basic chat functionality and have no intention of using agent or Cowork modes. The VM spawn…
Amazon Web Services announced Amazon Bedrock Managed Knowledge Base, a new service that simplifies building enterprise generative AI applications by managing retrieval-augmented generation (RAG) pipelines. The service in…
The Anthropic leader who built Claude Code has abandoned traditional prompting techniques, now relying on simple loops to achieve results. The developer's shift away from complex prompt engineering marks a significant ch…
Helion kernels were integrated into vLLM for FP8 inference using Qwen3 models and evaluated across NVIDIA H100 and B200 GPUs. The experiments demonstrated that Helion provides a productive PyTorch-native workflow for dev…
Anthropic released its latest AI model Fable on Tuesday as a public, limited version of its powerful cybersecurity model Mythos, but cybersecurity researchers are voicing complaints online about overly restrictive guardr…
AMD released Lemonade SDK version 10.7, adding NVIDIA CUDA support to its local AI server solution that previously only supported AMD hardware, Apple Metal GPUs, and AArch64 CPUs. The update integrates Llama.cpp's CUDA b…
Google DeepMind's DiffusionGemma, optimized for NVIDIA platforms, generates text tokens in parallel rather than sequentially, achieving up to 1,000 tokens per second on a single NVIDIA H100 GPU. The model, built on the G…
NVIDIA has optimized Google DeepMind's new DiffusionGemma model to run up to four times faster on its GeForce RTX GPUs, RTX PRO workstations, and DGX Spark systems. Unlike traditional language models that generate text o…
Researchers at the AI company Writer published two papers Wednesday demonstrating that memory and personalization systems in AI models can degrade accuracy by pulling models toward user misconceptions. The research found…
Google DeepMind released DiffusionGemma, an experimental open-source text diffusion model that generates text up to four times faster than traditional large language models by processing entire blocks of text simultaneou…
GitHub Copilot CLI now supports Language Server Protocol (LSP) integration through a new LSP Setup skill, replacing the agent's previous brute-force method of extracting JAR files and grepping through binary class files …
TiDB has introduced a unified database approach for AI agent memory that stores text, vectors, and metadata in a single table, eliminating the need for separate vector stores and sync jobs. The system follows a four-step…