Open Source AI Must Be Infrastructure, Not Appliances

The article draws a historical parallel between the dominance battles of the 1990s and today's AI landscape. Netscape and Microsoft both tried to lock users into monolithic products, but Apache succeeded by staying modular—offering a clean extension layer so anyone could add features without permission or waiting for a release cycle. The author, who studied this pattern in 2004 as the "architecture of participation," argues that open source victory was never about source code availability alone; it was about architectural freedom. Red Hat founder Bob Young put it simply: "What we really sell to our customers is control." That control enabled companies like Google and Amazon to build without bowing to Microsoft's dominance. Today's AI labs are making the same mistake Netscape and Microsoft did: locking desired behavior into model weights rather than keeping it in editable layers. The stakes are not merely about fairness or license philosophy—they are about innovation velocity. As Drew Breunig notes in the article, labs optimize for "distribution convergence," making models reliably good for lazy prompts, which creates a monoculture of outputs. Open protocols and modular harnesses, by contrast, let developers push models deliberately out of distribution to ship genuinely novel work. Current AI's AI Potluck and Anthropic's Model Context Protocol represent real progress on this front, but success requires that models remain infrastructure—tools developers can extend and modify—rather than appliances locked behind corporate guardrails.