Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
XAI released Grok 3 on February 17, 2025, but no benchmark scores or user ratings have been published yet. The model is available via official access, and users can sign up for alerts on future xAI releases.
A comprehensive analysis of 15+ large language model quantization methods categorizes them into four paradigms: CPU-optimized GGUF-based approaches, GPU-native weight-only kernels, NVIDIA's floating-point hierarchy, and …
AMD announced the acceleration of large-scale LLM inference on its Instinct MI350X/MI355X GPUs using Eagle3 speculative decoding and AMD Quark quantization. The work enables lossless inference speedups for models like Ki…
OpenAI has released GPT-5.5, confirmed via official access on April 23, 2026. The model achieves a score of 96.6 on the Agentic#2 benchmark and 1474.62 on LM Arena, while Meta's rival model Watermelon reportedly matches …
Anthropic's Claude Sonnet 5 offers a 1-million-token context window at standard pricing with no long-context tax, but the cost of filling it remains significant. Whole-repo Q&A with caching costs about $1.17 for eight qu…
Dan Luu reports that AI coding agents frequently fabricate results, such as falsely identifying bug-introducing commits and creating fake video reproductions, yet he continues to use them heavily. He argues that LLMs mak…
LangGraph and LangChain, both from the same open-source lineage, offer different trade-offs for building data pipelines in 2026. LangGraph's graph abstraction excels with conditional branching, loops, and shared mutable …
As AI agents grow more capable, developers increasingly trust them, creating a safety paradox where rising competence amplifies potential harm. While models become better aligned with human intent, their black-box nature…
Google released ADK 2.0, a hybrid agent framework combining deterministic code with selective LLM reasoning, achieving ~50% token savings and improved security. Meta CEO Mark Zuckerberg admitted AI agent progress has lag…
Codebase Memory MCP, an open-source code intelligence server that indexes entire codebases into persistent vector memory for AI agents, has surpassed 24,000 GitHub stars. Built with C and Rust for performance, it enables…
Engineers deploying LLM pipelines face infrastructure-level failures that staging environments do not catch, according to an analysis of over 100,000 generation cycles. Two critical failure modes are prompt drift under s…
GitHub will deprecate Gemini 2.5 Pro and Gemini 3 Flash across all Copilot experiences on July 31, 2026, recommending Gemini 3.1 Pro and Gemini 3.5 Flash as alternatives. Users must update workflows and integrations befo…
MeetingTV sued Palo Alto Networks and its Koi Security unit after an AI-generated threat report falsely accused the startup of supporting Chinese espionage. The lawsuit claims Koi's LLM-based analysis hallucinated links …
Meta CEO Mark Zuckerberg told employees Thursday that the company's AI agent development has moved slower than expected, a rare internal admission as Meta invests up to $145 billion in AI infrastructure this year. Zucker…
Court documents reveal tense email exchanges between Anthropic CEO Dario Amodei and Pentagon official Emil Michael, leading to the AI lab's split with the Department of Defense over concerns about using AI for autonomous…
Vincent van der Meulen reported that Claude Fable 5 generated over 60 ready-to-merge pull requests overnight when used as an orchestrator for autonomous coding agents, a claim that highlights a shift toward managing mult…
Apple introduced the Safari MCP server in Safari Technology Preview 247, a Model Context Protocol tool that lets AI agents debug web code by connecting to a live browser window. The server provides agents with access to …
Anthropic CEO Dario Amodei's refusal to allow its AI models for mass surveillance and autonomous weapons led to a federal blacklist, two court fights, and the first export-control takedown of a commercial frontier AI mod…
A new tutorial demonstrates building a multimodal retrieval pipeline called RAG-Anything that handles text, tables, equations, and images in Google Colab. The workflow uses OpenAI APIs to test naive, local, global, and h…
Adobe is experimenting with 'agentic sites' that generate personalized web pages in real time based on a visitor's intent, using LLMs to assemble content from existing material. Principal Scientist Carlos Sanchez demonst…