Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
A developer has published a guide for running Google's Gemma-4 12B instruction-tuned model on Windows Subsystem for Linux 2 (WSL2) using the llama.cpp framework. The process involves installing build tools, the NVIDIA CU…
Ollama, an open-source runtime for large language models, enables users to run models locally on Mac, Windows, or Linux with a single command, eliminating the need for cloud dependencies or complex environment setup. The…
A new Chrome extension called Perplexity to Obsidian enables users to batch export and download their PerplexityAI conversations as Markdown files for use in Obsidian and other knowledge management tools. The tool suppor…
A new file format called `skill.md` provides agents with concise, actionable instructions for using a product's documentation, telling them where to start and what to avoid. Unlike `llms.txt`, which serves as a map of do…
A developer has demonstrated how pairing open-weight LLMs like DeepSeek V4, Qwen, and GLM with the self-hosted, open-source Oinone metadata-driven low-code framework can produce maintainable enterprise business applicati…
A new architectural pattern called the Deterministic Core Architecture provides a transferable design pattern and production reference implementation for AI-augmented applications. The architecture addresses the fundamen…
Apple's current success remains driven by iPhone sales, with artificial intelligence not yet significantly impacting the company's business, according to Ranjan Roy, cofounder of Margins. Roy said the tech industry curre…
Thinking Machines Lab released two open-source libraries, `tinker` and `tinker-cookbook`, to help developers and researchers fine-tune large language models. The `tinker` SDK handles distributed training complexities via…
A new study analyzing six large language models on the SWE-bench Verified benchmark found that agent-written tests do not significantly improve issue resolution rates, despite being a common practice. Researchers observe…
A software engineering manager at a big tech company, who is preparing for future interviews at Google and Meta by practicing LeetCode-style coding problems, questions whether these data structures and algorithms intervi…
A developer has identified a critical flaw in LLM-generated code for the Japanese Rokuyo calendar: while the lunar date calculation formula is correct, the underlying lunar date conversion is frequently wrong, causing si…
Researchers introduced Tangram, a serving system that enables non-uniform Key-Value cache compression for multi-turn large language model inference. The system uses deterministic budget allocation, head group page cluste…
OpenAI on February 13, 2026 launched Lockdown Mode, an optional advanced security setting for ChatGPT that deterministically disables tools attackers could use to exfiltrate data through prompt injection. The feature, in…
A developer named Shrijal Shrestha has identified and named "AI-smell," the recognizable residue left by AI-assisted work across writing, design, and code. The developer behind a content pipeline and dev tools reports th…
A developer deploying a 70B Llama-3 model on 8x H100s found that scaling from 8k to 32k context windows causes the KV cache to balloon to 10.7 GB per request, forcing memory paging to CPU at 200 concurrent users. KV cach…
A game developer has built FreeUltraCode, a local desktop application that provides access to over 20 free and low-cost large language model channels for AI-assisted coding. The tool, built with Tauri 2 and Rust, allows …
An engineer has identified three distinct patterns of "agent drift"—gradual degradation in AI agent output quality that occurs without hard failures or schema violations, often going unnoticed for days until a customer c…
Apple's Mac Studio with M3 Ultra, offering up to 512 GB of unified memory at 819 GB/s, is the most practical desktop solution for running large local AI models like 70B to 400B-class quants without multi-GPU setups. Owne…
A developer has published Throughline, a hook plugin for Claude Code that offloads tool I/O to SQLite to reduce token consumption. In a 50-turn session, the plugin reduced context from 125,000 tokens to within 13,000 tok…
A developer has reverse-engineered the TR2 game file format for the obscure Japanese-exclusive Wii game And-Kensaku, disabling signature checks and enabling edits to the game's phrase files. The project, which uses the A…