Open, With an Asterisk

"Open source AI" is one of the most repeated phrases in the industry. It is also one of the most misused. Here is what "open" really means, and what it doesn't.

The Trouble With "Open"

Back in Issue 15, we defined the difference between proprietary and open source AI, and flagged that the line gets blurry. This week we walk into the blur, because it matters.

When a company says its AI is "open", most people hear "free, transparent, and yours to inspect." Sometimes that is true. Often it is not. The word has become a marketing asset, and the gap between what it implies and what it delivers is where the confusion lives.

The clearest example is the biggest one. Meta's Llama family is the most downloaded, most talked-about "open" AI in the world. It is also not open source in the way most people assume.

Open Weights Is Not Open Source

Here is the distinction that clears up almost everything.

Open source, in traditional software, means you can see the source code, understand exactly how it was built, modify it, and reuse it with very few restrictions. The recipe is public.

Open weights means something narrower. You get the finished model, the trained parameters, the thing you can download and run. What you usually do not get is the training data, the full training process, or the freedom to do absolutely anything you want with it. You get the cake, not the recipe, and sometimes with a note about who is allowed to eat it.

Most "open" AI models, including Llama, are open weights. You can download them, run them on your own hardware, fine-tune them, and build products on top of them. You cannot see exactly what they were trained on, and the license may restrict how large a company can be before it has to pay.

That is a useful kind of open. It is just not the same as open source, and the difference is important.

Why a Company Gives Its AI Away

If training a frontier model costs tens of millions of dollars, why release it for free? It is not charity. Open weights is a strategy.

Releasing a capable model widely means developers build on your platform instead of a competitor's, your model becomes a default other tools are compared against, and a global community finds improvements and use cases you never would have staffed for. It also applies pressure to rivals who charge for access. When a strong model is free to run, the price of the paid ones gets harder to defend.

Understanding the motive helps you read the landscape. "Open" is often a competitive move, not a philosophical one, and that is fine. It just means you should not mistake the generosity for the whole story.

Safe Harbor: Three Things You Can Do This Week

• Research one tool that is "open source" and one that is "open weights". Pick two AI models and look up how each is actually released. Meta's Llama is the classic open-weights example. For a true open-source model, search for one that publishes its training data and code, not just the weights. Comparing two real cases side by side makes the distinction stick far better than a definition does.
• Read the license, not the headline. Many "open" models restrict commercial use above a certain size or forbid specific applications. If your organization is considering building on one, the license is the part that matters, and it is usually one search away.
• Download a small open model and run it once. If you or someone technical on your team spends ten minutes running a small Llama or Gemma locally, the whole "you get the weights, not the recipe" idea stops being abstract.

Next week: why AI states wrong things with total confidence. Hallucinations are the single most misunderstood behavior in AI, and knowing why they happen is the difference between trusting a tool and getting burned by it.