Open Source AI Wins: Modularity, Protocols, and AI Potluck Back a Composable Future

A Tim O’Reilly essay argues that open source AI matters less for “open weights” licensing and more for architecture: modular layers, clean interfaces, and easy swap-ability. Using Apache vs. Netscape/Microsoft as the core analogy, O’Reilly says open source AI keeps ecosystems open because developers can extend components without permission and upgrade parts when better options appear. The piece warns that frontier closed models increasingly push behavior into model weights, making them appliances to rent rather than infrastructure to modify. It frames this as a trade of “trading diversity for reliability,” where reliability rises but innovation diversity shrinks. Key open standards and projects cited include Anthropic’s Model Context Protocol (MCP), now hosted beyond Anthropic at the Agentic AI Foundation (a Linux Foundation subproject). The article also highlights “Current AI” and its AI Potluck initiative—aimed at building a vertically integrated AI product entirely from open source components as an alternative not owned by any one company or country. Funding mentioned: about $400 million of a five-year, $2.5 billion French government commitment, supported by partners such as DeepMind and Salesforce and major philanthropies. O’Reilly concludes that future competition should move from “models” to “context” and “harnesses,” enabling developers to build “weird” workflows outside lab roadmaps. Overall, the article positions open source AI as a market-opening strategy through composability and protocol-centric access.