Top story

TOP STORY DeepSeek Reasonix: Native Coding Agent with High Caching. DeepSeek releases Reasonix, a cost-efficient native coding agent built for high-throughput caching and low-latency inference. DeepSeek Reasonix

Research

Memory Now Drives Two-Thirds of AI Chip Component Costs. Epoch AI analysis shows memory has become the dominant cost factor in AI chip design, reshaping hardware economics. Epoch AI

Constraint Decay: LLM Agent Fragility in Back-End Code Generation. New arxiv paper identifies how LLM agents progressively lose adherence to constraints during complex backend code generation tasks. arXiv

Good Token Hunting: Token Selection for Visual Geometry Transformers. Researchers propose a principled framework for selecting tokens in vision transformers to improve geometric understanding. Hugging Face

PhotoFlow: Agentic 3D Virtual Photography Missions. A new agentic system automates 3D virtual photography by decomposing complex scene-capture tasks into sequential missions. Hugging Face

Tools

Ollama Adds Support for Kimi-K2.5, GLM-5, MiniMax & More. The popular local model runner expands its model library with several major new releases including DeepSeek and Qwen variants. GitHub

Datasette 1.0a30 Released. Simon Willison ships a new alpha of Datasette alongside companion packages datasette-agent and datasette-fixtures. Simon Willison

Dify Reaches 142K Stars as Production Agentic Workflow Platform. Dify continues momentum as a leading open-source platform for building and deploying production-ready agentic workflows. GitHub

Industry

OpenAI Releases GPT-5.5, GPT-5.4, and Codex Frontier Models. OpenAI expands its model lineup with a cybersecurity-focused GPT-5.5, a general frontier GPT-5.4, and an agentic coding model GPT-5.3-Codex. OpenAI Release Notes

OpenAI Open-Weights gpt-oss-120b & gpt-oss-20b Now Available. OpenAI releases two open-weight reasoning models for community use and self-hosting. OpenAI

Anthropic Ships Claude Opus 4.7. Anthropic releases the latest iteration of its flagship Opus model with updated capabilities. Anthropic Docs

Community

Is NVIDIA Still the Default for Local LLMs in 2026?. The LocalLLaMA community debates whether NVIDIA remains the go-to hardware choice as AMD and Apple Silicon mature. Reddit / LocalLLaMA

Testing a Cold War-Era AI on Modern Satellite Image Datasets. A community experiment evaluates legacy AI architectures against contemporary satellite imagery benchmarks with surprising results. Reddit / artificial