Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Databricks announced at its Data + AI Summit 2026 that Agent Bricks is now generally available, with over 100,000 agents built on the platform processing one quadrillion tokens annually. The company introduced managed ag…
A developer implemented a full Transformer with KV cache on an FPGA, achieving over 56,000 tokens per second at only 80 MHz, without using a GPU or CPU. The design was created gate by gate as a custom digital integrated …
ZeroGPU released a Batch API that enables overnight tagging of customer reviews using a small, efficient model (LFM2.5-1.2B-Instruct) instead of expensive frontier LLMs. The API processes thousands of rows asynchronously…
Georgi Gerganov, creator of llama.cpp, reported that the Qwen3.6-27B model is highly capable for local coding tasks, using it daily on his M2 Ultra and RTX 5090 systems. He noted the model's utility for mundane maintenan…
Meta launched AI Mode on Facebook in June 2026, replacing traditional search with conversational answers synthesized from public posts, Groups, and Reels using its Llama 3 model. The feature lacks a system-wide off switc…
A developer built a NestJS-based inference layer using DeepSeek's open API after receiving a $4,200 invoice from a proprietary AI vendor. The setup provides access to 184 models at prices ranging from $0.01 to $3.50 per …
A writer outlines a method for training AI to replicate an individual's unique writing style, called voice cloning, by conducting a voice audit and creating a Voice Profile Document. The process involves analyzing five t…
NVIDIA released a guide showing how to optimize transformer-based models for low-precision training using Hopper and Blackwell GPUs, focusing on FP8 and NVFP4 formats. The method translates model configurations into GEMM…
Red Planet Labs reports progress in teaching LLMs to one-shot complex backends at scale, achieving consistent one-shot generation of the fanout challenge—a core Twitter-like service—using Fable and Opus models. The team …
Google's Threat Intelligence Group confirmed the first AI-generated zero-day exploit used in a real attack in May 2026, with forensics revealing its AI origin through telltale signs like educational docstrings and halluc…
Microsoft released a preview of new AI performance reports in Bing Webmaster Tools, adding Intents, Topics, Citation Share, and Compare features to help publishers understand why their content is surfaced in AI responses…
A developer's analysis of agentic AI frameworks with 370,000 combined GitHub stars reveals that production reliability remains elusive despite impressive demos. The gap between 'can complete task' and 'reliably completes…
A 2026 analysis reveals that 92% of developers use AI coding assistants monthly, yet only 7% of organizations have deployed autonomous agents in production, creating a 13x gap. The report warns that organizations failing…
Anthropic's Claude models now support mid-conversation system messages, allowing developers to inject new instructions into a long-running agent session without invalidating the entire prompt cache. By placing a system-r…
BlitzGraph, a new graph database built for LLM agents, launches in beta on Hacker News. The database supports polymorphic records and relations, time-evolving entities, and a JSON query language designed for AI agents, a…
Vercel's Skills.sh, an open ecosystem for sharing reusable AI agent capabilities, has surpassed 91,000 total installs with top skills reaching millions of downloads. The platform, built on an open standard from Anthropic…
A user on Hacker News reports that large language models consistently fail to understand board game rules, raising concerns about fundamental limitations in AI reasoning. Despite rulebooks being self-contained, LLMs cann…
Subquadratic released the model card and technical report for SubQ 1.1 Small, a long-context model using Subquadratic Sparse Attention, on June 16. The company claims the model achieves 100% retrieval accuracy at 1 milli…
A developer describes a workflow where AI agents handle implementation while humans focus on planning, architecture, and system design. The key insight is that the quality of AI-generated code depends on the quality of t…
Mistral AI rebrands Le Chat as Vibe, a unified AI agent for work and coding tasks. Vibe features Work Mode for multi-step enterprise workflows and Code Mode for remote coding, with a new VS Code extension. The agent runs…