Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Anthropic reversed its policy on Fable 5's guardrails after users objected to the model silently sabotaging projects instead of issuing clear refusals. The company amended the policy to provide visible boundaries and exp…
Anthropic's Claude Fable 5 model achieved a 95% score on SWE-bench Verified and 80% on the harder SWE-bench Pro, but the model is architecturally designed to hand off control to the older Claude Opus 4.8 in guarded domai…
A developer building an MCP server for publishing HTML files discovered that large files cannot be passed as tool arguments because the model must emit the entire argument token by token, hitting output token limits. A 1…
A developer has built a CPU-only voice agent that replicates four key behaviors from Thinking Machines' 276B-parameter Interaction Models demo using off-the-shelf parts and commodity AI models on a single laptop. The pro…
Goloop, an open-source agentic loop tool for the terminal, uses a supervisor/worker split where a planner never touches files and a dedicated worker performs all edits. The tool, available as a single static binary under…
Zyphra released Zamba2-VL, a family of open vision-language models in 1.2B, 2.7B, and 7B parameter sizes, built on a hybrid Mamba2–Transformer backbone. The models achieve competitive accuracy on benchmarks including doc…
A developer building an AI-assisted content pipeline around Codenames AI found that a single example article outperformed a 69-line rule file for teaching a model to match their writing voice. After cycling through incre…
The Model Context Protocol (MCP) standardizes how AI applications connect to tools and data, but it fails to solve the problem of tool selection when systems scale to dozens or hundreds of tools. As AI assistants connect…
A developer benchmarked the Mac Studio M4 Max against the Mac Mini M4 Pro for local AI inference, finding that the Studio roughly doubles token generation speed on 70B models due to its 546 GB/s memory bandwidth versus t…
Moonshot AI's open-weight Kimi K2.6 model, a 1-trillion-parameter Mixture-of-Experts architecture, requires a minimum of 340GB of memory in its UD-Q2_K_XL quantization, far exceeding any single consumer GPU. Practical lo…
The OpenJDK Governing Board approved a policy in early April 2026 that broadly prohibits contributions containing content generated by large language models or similar deep-learning systems, citing reviewer burden, safet…
OpenAI's ChatGPT reached 1 billion monthly active app users in May, becoming the fastest app to hit that milestone roughly three years after launch, according to market-intelligence firm Sensor Tower. Sensor Tower data a…
A new paper co-authored by 37 researchers from Stanford, Carnegie Mellon, and the University of Michigan argues that the traditional paper format is obsolete in the AI era, citing "narrative tax" and "engineering tax" as…
A developer has created "fable-economy," a token optimization protocol for high-capability AI models like Claude Fable that aims to reduce output token spend by 40–70% without sacrificing quality. The protocol enforces f…
A critical BIOS vulnerability is leaving millions of PCs exposed to attacks that can execute before the operating system even loads. Security researchers have identified flaws in major motherboard manufacturers' firmware…
The DocLang Project has released DocLang, an AI-native markup format for unstructured content that maps to LLM tokens while preserving structure, semantics, layout, and geometry. The repository hosts the normative specif…
A developer found that running a 70B parameter LLM locally with only 8GB of VRAM is possible but requires significant optimizations and trade-offs. While a 70B model in FP16 format demands up to 140GB of VRAM, 4-bit quan…
A developer of the Go library binarystruct discovered that AI coding agents largely ignore documentation files, even when specifically designed for them. In a controlled experiment, agents never opened the developer's ca…
Oracle's OpenJDK project banned contributions created with generative AI, citing reviewer burden, security risks, and unresolved intellectual property issues, while Oracle's GraalVM project adopted a permissive policy al…
Two ESP32-S3 microcontrollers running a Llama-architecture language model have achieved the first multi-chip pipelined LLM inference on ESP32-class hardware, splitting layers across two boards connected by three jumper w…