AI Digest: Technology Theft Accusations, MCP, and Local Models

In today’s roundup: a new turn in the debate over copying AI technologies, practical materials on MCP and team development with agents. We’ve also collected useful community findings for running models locally.

🔥 Hot:

🔹 The US accused Chinese AI companies of large-scale technology copying — The discussion references materials from Reuters and NBC News.

➡️ News:

🔹 minnow — a server for quickly launching LLaDA2.2 — appeared on GitHub

➡️ Useful materials:

🔹 An overview of connecting your own MCP server to Copilot Studio — The article explains why Streamable is needed instead of SSE and how to configure authentication and access through Power Platform. 🔹 An overview of the challenges of team development with AI agents — The author explains why working with agents requires new processes for control and transparency within a team. 🔹 An overview of the limitations of AI routers and MoE models — A small number of activated parameters does not mean that the entire model will fit into a small amount of VRAM: weights, context, caches, and data exchange are also important.

➡️ Discussions and use cases:

🔹 The community compares a pair of RX 6800 cards with two RTX 2060 cards for llama.cpp — The discussion focuses on upgrading from 24 to 32 GB of VRAM and the possible performance gains. 🔹 Users are looking for the best vision models with up to 6B parameters — Qwen 3.5 4B and MiniCPM-V-4.6 are among the candidates being discussed. 🔹 The author built a visual lab for explaining school algebra — The project helps illustrate why the sign of a number changes when it is moved in the equation 2x + 4 = 10.

📝 If you’d like to add other news and materials to the list, write in the comments.