Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
GLM 5.2, an open-weight coding model from Z.AI, outperformed Claude Opus in four real-world coding tasks including a codebase audit, UI redesign, and autonomous bug hunting, costing only $3.36 for 6 million tokens. The m…
Alibaba's Qwen team released Qwen-AgentWorld, a language model that simulates complex environments natively, replacing external simulators for training AI agents. The model, trained on over 10 million trajectories across…
A developer debugging an MCP (Model Context Protocol) integration found that an LLM agent's self-report of unavailable tools was misleading. The agent claimed that new stdio MCP servers were not connected, but the real i…
A developer building a local-LLM agent for a boat's ship-computer role encountered two types of confabulation: the agent invented infrastructure topology and data provenance stories. The 8B model fabricated a diagnosis o…
Google's NotebookLM, an AI research assistant that generates insights from user-provided data, has won a CNET Editors' Choice award for its unique ability to organize and transform information into study guides, audio ov…
The most upvoted papers on Hugging Face reveal a trend of AI shifting from answer models to action models, focusing on agents, simulation environments, GUI/mobile interaction, and benchmarks for real-world task capabilit…
Apple has closed a loophole that allowed Mac users to bypass the Siri AI waitlist in macOS 27 beta 2, rendering a Terminal command ineffective. Users who previously enabled the feature via the workaround have been remove…
AI agent developers are warned that large context windows are not equivalent to memory, as models remain stateless between calls and must re-read entire histories each time. The article explains how techniques like retri…
A new analysis argues that brand visibility in the age of AI depends less on SEO and more on organizational data consistency, as large language models reflect internal inconsistencies back to users. With 71% of organizat…
Google's 'Tailor Your Feed' feature, now rebranded as 'Add topics to your feed,' allows users to shape their Discover feed by typing natural language prompts, which are turned into explicit control actions. The feature s…
A growing consensus among AI coding tool users advocates for writing loops instead of prompts, shifting from manual agent steering to designing systems that automate repetitive tasks. The approach aims to free human atte…
The US government is urging Meta to submit its AI models for evaluation, making it the only major AI developer that hasn't voluntarily shared its models for review amid growing safety and security concerns. Meta has expr…
A developer found that Anthropic's Claude Code AI assistant consumed significant storage space on Windows. After running out of disk space, the developer asked Claude Code to generate a cleanup script, which claimed it c…
A new web data infrastructure layer is emerging to enable AI models to access real-time, structured data at scale, overcoming the web's inherent design limitations. Bright Data CEO Or Lenchner emphasizes that stale data …
A developer created an n8n workflow integrating OpenWA with WhatsApp to build an AI bot with three varieties: personal contact, group message, and group message with @agent mention. The workflow uses an Ollama chat model…
The Washington Post tested AI chatbots including ChatGPT for political bias, finding that they exhibit left-leaning tendencies on controversial topics. The study highlights concerns about the neutrality of AI systems use…
Developers are shipping AI agent skills inside NuGet packages so that when a library is installed, the skill automatically teaches coding agents how to use it correctly. This approach, pioneered by Daniel Cazzulino for t…
A developer tested several open-source models on Indian language tasks and found that Qwen 3 handles Devanagari, Tamil, Bengali, and Telugu scripts natively without fine-tuning. The model matches or exceeds GPT-4o on Ind…
A student developer built an open-source AI malware analyzer for cybersecurity education. The tool runs suspicious binaries through static analysis tools like YARA and ClamAV, then uses local LLMs or OpenRouter to genera…
Anthropic released Claude Fable 5, the first public Mythos-class model, on June 9, 2026. The model offers enhanced coherence and autonomy for complex, multi-file, long-horizon systems like Polymarket trading bots, with p…