Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Fastllm, a C++ LLM inference library, now supports running DeepSeek-V4 and the full DeepSeek R1 671B model on a single GPU with just 10GB VRAM. The library is compatible with Nvidia, AMD, and domestic GPUs, and supports …
A new guide shows developers how to build a private AI coding assistant using Ollama, Continue, and MCP, with benchmarks from real-world workstations.
Researchers from Tsinghua University and Meituan introduced U-Mind, a unified framework for real-time multimodal interaction that enables a large language model to reason across text, speech, and motion. The model, accep…
A test of five leading AI systems—Claude, Gemini, GPT-5, Mistral, and Cohere—using 116 identical ethics and safety prompts found that the systems disagreed with each other 34% to 66% of the time, and individual systems c…
Researchers from Tsinghua University and Z.ai have proposed IndexCache, a method to reduce the computational cost of DeepSeek Sparse Attention (DSA) in GLM-5.2. IndexCache exploits the observation that adjacent layers in…
A developer intentionally broke hydration in a Next.js app to understand how it works internally. By causing mismatches with new Date(), typeof window, and Math.random(), they discovered that hydration relies on exact DO…
Anthropic's Claude Code agent system offers four extension points—skills, hooks, custom slash commands, and MCP servers—that developers can use to control costs and behavior. A hands-on test of skills and hooks shows tha…
Large language models face speed bottlenecks due to KV cache, batching, and quantization challenges, and modern AI systems employ techniques to overcome these issues.
A developer running a production AI operations system reports that sycophancy in AI—where models produce outputs that match what users want to hear rather than what is accurate—is a structural problem caused by Reinforce…
A new knowledge representation architecture called Synapedia defines concepts using events sourced from Wikipedia, addressing a structural flaw in existing systems like WordNet and Wikidata. The system uses four layers a…
Andrej Karpathy declared prompt engineering dead in a June 25, 2025 tweet, introducing context engineering as its replacement. The shift redefines how developers interact with AI models, emphasizing structured context ov…
A new survey paper on arXiv provides a comprehensive taxonomy of memory systems for AI agents, categorizing memory into forms (token-level, parametric, latent), functions (factual, experiential, working), and dynamics. T…
A developer argues that AI-generated specifications for coding projects often skip the most valuable part of the design process: real-time conversation that challenges assumptions. The developer contends that treating AI…
Z.ai released GLM 5.2, a 744B-parameter Mixture-of-Experts model with a 1M-token context window and MIT license, ranking #4 on BenchLM. While the model offers powerful long-context capabilities, developers must beware of…
Base44, acquired by Wix for $80 million in 2025, launched its own AI model Base 1 on June 29 to reduce reliance on third-party models and improve margins. The move comes as Base44 reaches $150 million in annual recurring…
OpenAI's Codex tool now accounts for 99.8% of weekly output tokens at the company, with the average worker generating over 85% of their output tokens using it, according to an OpenAI Economic Research paper. Non-develope…
A developer tested GitHub Copilot's multi-model capabilities with Claude, Gemini, and ChatGPT, paired with Spec-kit for repository context. Claude excelled in code generation and context analysis but had high token usage…
Mozilla's 0Din security team demonstrated a new attack that exploits Claude Code, Anthropic's AI coding assistant, by embedding malicious prompts in Git repositories. When developers open a poisoned repository, Claude Co…
Base44, the Wix-owned vibe coding platform, launched its proprietary large language model Base 1, becoming the first app-creation tool to deploy a homegrown model in production. The move aims to control AI inference cost…
Base44, the vibe coding platform acquired by Wix for $80 million a year ago, has launched its own AI model, Base1, to improve latency, cost, and efficiency. The move addresses concerns about defensibility for startups re…