开源不等于开放权重：Gary Marcus 剖析两者本质差异

Putting scares quotes around open-source does not make an open-weight model more open. And although one can post-train1 Meta’s new model, readapting it to some degree, for all the reasons I just described, it can’t be freely modified, nor used for proper science. You can’t do any of the modifications I described above. And you can’t figure out how the model got to be the way it is in the first place.

It’s no exaggeration to say that open-source software — in which the complete source code is available to anyone — has changed the world and accelerated software development.

Meta just released an open-weight model. To his credit, Zuckerberg himself got the terminology right here:

For developers, let’s say you have a hypothesis that you could make a safer AI if you delete all of the Reddit data from training. Sorry, out of luck: in an open-weight system (as opposed to a truly open-source system) you can’t change the training data at all, and in fact you can’t even see the training data. Or let’s say you would like to train a model on legal reasoning and recognize that common law and statutory law are different. To reduce confusion in the system you would like to filter out the common law stuff before training; again you are out of luck. Or suppose you would like to use Big Company’s training process purely on your own data, rather than on whatever random stuff they scraped from the web. Nope, can’t do that either. (Nor can you share the results of any of those experiments with the community, since you can’t run them in the first place.)

If pigeons are “just rats with better PR”, open-weights is open-source with all the good press but far fewer advantages.

And then today he released a model that is NOT open-source, hoping you wouldn’t notice the difference between the two. And, sadly, the Times didn’t.

Understanding all this requires a basic familiarity with the life cycle of a machine learning model. Let’s start with a simplified sketch I have snarfed (and modified) from geeksforgeeks.org. The highlighting, which I will explain in a moment, is mine.

P.S. Bonus analogy; pick your favorite:

But also played bait and switch with another tweet here:

Are you a scientist? Want to investigate how much of what the model does is regurgitation and how much extends beyond the training data, which is arguably the central question in AI? Sorry, out of luck there, too. Those who released the open-weight model could do that, since they are privy to all these details that are actually closed, but you can’t. Or maybe you just want to know whether the answers to your favorite benchmarks are actually already in the training data? Nope, can’t look at that, either.

I implore them, and everyone else, for the love of Turing, please don’t confuse open-weight with open-source.

Open-weight piggybacks onto all the great press that open-source has but offers few of the advantages. Open-weight models are NOT fully customizable, and not transparent. They are the product of machine learning (the weights of a trained neural network), but not the whole thing, from soup to nuts. That means you can run them (or “posttrain” them), but, crucially, users can’t customize or improve them to remotely the same degree (more about that in a second), and outsiders can’t investigate them to nearly the same degree. You also can’t build community in the same way.

The essence of open-source is twofold: transparency and customizability. If you want to know how something works, you look at the code. You can change literally anything about that code (“forking” a new version of that code, for example, if you want to make a different version). All of this is fantastic for the community, and that is why open-source has been such a dynamic force. It’s called open-source because you release the source code of the final product. Open-weight doesn’t do that.

Or say you are a regulator, and you would like to know how much bias there is towards white males in the training data. Sorry, can’t look at that, either. What kinds of instructions are there relevant to bioweapons? Again, nope.

But the Times did not:

Furthermore, if you wanted to know how much of the training data was copyrighted, you can’t look at that either — which is perhaps part of why none of the big companies want to release true open-source, and instead settle for open-weight, getting nearly all of the positive press without actually helping the community nearly as much.

(Not shown is “post training”, in which people try to align and bend trained “base” models to their will. That’s all you can do with an open-weight model; I will give you examples of what you can’t do in a moment.)

Zuckerberg knows perfectly well the difference between open-weight and open-source. A few weeks ago he was extolling the value of open-source:

In contrast, when people release open-weight models, they are not revealing the Full Monty. They are not releasing the raw data; they are not revealing how they preprocessed that data; and they are not revealing the exact algorithms and parameters that they used in training. (For those who can stand yet one more loose metaphor, it’s like they are releasing a cake, but not the recipe; you can add icing on top, and maybe some fruit or hot fudge too, but you can’t change the raw ingredients that went into the cake itself.)

When someone releases an open-weight model they are releasing the (highlighted) candidate model [also sometimes known as a base model]—which is the output of a complex process—not the full pipeline that generated that model. You can adapt that candidate model, but you can’t rebuild it or investigate it to your liking, the way you could with a true open-source model, such as AllenAI’s Olmo or Nvidia’s Nemotron, which truly is what it says on the tin: “open weights, training data, and recipes.”

Post-training is a tiny bit like adopting an adult stray dog and trying to retrain it; it’s not that you can’t do anything, but it’s simply not the same as training it from day one.

But open-source and open-weight are two very different things, and too many people —this morning it was editors and writers at The New York Times among others—are failing to distinguish the two.

What prompted me to finally write about this?

All this makes a world of difference, to developers who might use the systems, to regulators, to scientists, and ultimately to the world.

All of this is unfortunate. It means that developers often can’t really use the open-weight models (despite the hype) to develop what they want; scientists can’t use them to test a lot of key hypotheses, and regulators can’t peer inside the process.