This Isn't the Open Source You Were Promised

A senior engineer at a mid-sized analytics firm pulls the latest Llama 3 weights to a local server. The terminal progress bar fills up. For the first time in months, the finance department isn't looking over her shoulder at the spiraling OpenAI invoice. This feels like freedom. It feels like control. It is neither.

We are watching a strategic sleight of hand play out in real time. The term "open source" is being co-opted to mean "source available, with conditions." The ethos of collaborative, unrestricted development that gave us Linux and Apache is being used as a branding exercise for assets that are anything but open. These new models are not a commons; they are gated communities.

Look at the licenses. Meta’s Llama 2 and 3 licenses forbid you from using the model to improve any other large language model. They stipulate that companies with over 700 million monthly active users must request a special license. This isn't a bug; it's a moat. It’s a legal barrier designed to prevent a potential competitor from emerging from the very community the model purports to empower. It uses the language of openness to enforce a new kind of platform control.

Then there is the data. The true source code of an AI model isn't just its architecture; it's the unimaginably vast corpus of text and images it was trained on. That corpus is almost always a secret. We get vague descriptions—"a mix of publicly available online sources"—but the specific composition, the cleaning process, the filtering, the copyrighted material that was inevitably scraped? That’s a trade secret. A company building its flagship product on such a foundation is not standing on solid ground. It's building on a black box, hoping no one ever comes knocking with a lawsuit about the training data.

The illusion of control shatters further when you consider the hardware. Running inference on a powerful open model is one thing. Fine-tuning it or, heaven forbid, pre-training a competitor from scratch requires a server farm full of H100s that only a handful of corporations and nation-states can afford. This isn't like compiling a new Linux kernel on your laptop. The capital expenditure required to meaningfully participate in the model's evolution creates a stark dividing line between users and creators. You can use the tool, but you can’t build the factory.

This new wave of "open" models is a brilliant gambit. It allows companies like Meta and Mistral to rapidly build ecosystems around their technology, offloading the costs of discovery and integration to the wider community. It starves truly community-driven, permissively licensed efforts of oxygen. And it gives enterprises a comforting, but ultimately false, sense of security. They believe they are escaping vendor lock-in from the big API providers, but they are merely swapping one form of dependency for another.

The choice is no longer between an open field and a walled garden. The choice is between different landlords, each with their own set of rules and their own right to change the terms of the lease. The engineer watching her download complete hasn't escaped the system. She's just signed up for a new one.