Top story
TOP STORY: Qwen3.5-27B Rivals 397B Models in Coding Benchmarks. The 27B parameter model matches the performance of models nearly 15x its size in the Game Agent Coding League, coming close to GPT-5 mini. reddit/LocalLLaMA
Research
What Is Agentic Engineering?. Simon Willison's comprehensive guide defines and maps the emerging patterns behind building autonomous AI agent systems. Simon Willison
Detecting Self-Preservation in Autonomous Agents. Researchers propose a unified protocol for identifying both intrinsic and instrumental self-preservation drives in AI agents. HuggingFace
Can Vision-Language Models Solve the Shell Game?. A new study probes whether VLMs can track objects under occlusion, testing a fundamental aspect of physical reasoning. HuggingFace
Multimodal OCR: Parse Anything from Documents. A new approach enables robust extraction of text and structure from complex, heterogeneous document types. HuggingFace
Tools
LLM Architecture Gallery. Sebastian Raschka's visual reference compiling diagrams and breakdowns of major large language model architectures. Hacker News
GreenBoost: Expanding NVIDIA vRAM with System RAM and NVMe. An open-source driver project aims to let users run larger LLMs by augmenting GPU memory with system RAM and NVMe storage. reddit/LocalLLaMA
LookaheadKV: Faster KV Cache Eviction. This method improves inference speed and accuracy by peeking ahead in generation to make smarter KV cache decisions. HuggingFace
Quillx: Open Standard for Disclosing AI in Software Projects. A proposed open standard gives developers a structured way to document AI involvement in their codebases. Hacker News
Industry
Consultants Are Cashing In on the AI Boom. WSJ reports on the surge in demand for AI consultants as enterprises scramble to implement AI strategies. WSJ / reddit/artificial
Coding Benchmark That's Hard to Fake, Best Score: 11%. A new adversarial coding benchmark proves resilient to prompt engineering tricks, with top models from multiple labs scoring in the single digits. reddit/LocalLLaMA
Gig Workers Paid to Film Daily Chores for Robot Training. Companies are recruiting gig workers to capture everyday household activity data to train embodied AI systems. reddit/artificial
Community
A Visual Introduction to Machine Learning (2015). This evergreen r2d3 interactive essay on decision trees resurfaces as a perennial favorite for intuitive ML education. Hacker News
How I Write Software with LLMs. A practitioner shares their concrete, opinionated workflow for integrating LLMs into the day-to-day software development process. Hacker News
LLMs Can Be Exhausting. An honest reflection on the cognitive and emotional friction that comes with relying heavily on LLMs in creative and technical work. Hacker News
Homelab Has Paid for Itself. A LocalLLaMA community member shares how running local AI infrastructure eventually justified the hardware investment. reddit/LocalLLaMA