Temporary Aligned Zones

Used by more than 18 million developers, researchers, and creators, Hugging Face is the closest thing the AI world has to an open commons for sharing models, datasets, and applications. Yesterday, Axios reported that Nvidia will acquire the platform for $12.93 billion.¹ That’s a lot of money. Nvidia says Hugging Face will remain open and that developers can continue to choose their models, frameworks, and providers.²

»The possibility space creates the conditions that eventually close the possibility space«

The obvious question on everybody’s minds is whether Nvidia – the dominant supplier of the silicon chips the entire AI industry runs on – will eventually kill the thing it just bought. I find that unlikely. Nearly $1 billion of that price is a retention pool, set aside to keep Hugging Face’s own staff in place: an odd thing to pay for if the plan were to gut it. Nvidia has spent nearly thirteen billion dollars buying a machine for making people want compute. Its own SEC filing (a public financial statement) says that demand for open-source models promotes the use of its products.³

Nvidia isn’t allowing Hugging Face to stay open. It needs Hugging Face to stay open.

There’s a peculiar kind of freedom that appears when its protection is crucial to somebody else’s interests. Call it a temporary aligned zone. A space doesn’t have to be independent of power to be autonomous. It can be autonomous because the interests of the people inside it happen, for a while, to align with the power on the outside who controls the constraints around it.

Hakim Bey was the pen name of philosopher-poet Peter Lamborn Wilson, who shared a New York townhouse with writer William S Burroughs. In 1991, he dedicated a book to his concept of the temporary autonomous zone, a space such as a rave, which escapes the surrounding order before it can be captured.⁴ The Hugging Face situation is almost the inverse: the temporary aligned zone survives because power has a reason to protect it. Artists need somewhere to experiment, patrons need artists; scientists need somewhere to investigate, corporations need discoveries. Bell Labs is the famous example, an extraordinary research culture and a temporary aligned zone inside a telephone monopoly.⁵

The interesting part is what happens when those interests change, as with gentrification. A cheap neighbourhood becomes interesting to artists and creatives, the interesting neighbourhood becomes valuable, rents rise, and the people who made it interesting leave. The possibility space creates the conditions that eventually close the possibility space: a temporary aligned zone until the zone itself becomes valuable enough to end it.

GitHub is a more direct precedent. Microsoft bought it for $7.5 billion in 2018, without much reason to destroy its openness: openness was part of what made the platform valuable.⁶ Then AI changed what GitHub was. In March this year, GitHub announced that from 24 April, interactions from consumer Copilot users would be used to train and improve AI models unless they opted out.⁷ The platform had acquired a new value, and the calculation around it changed. Nothing necessarily went wrong: the object changed, and with it, the constraints around the object. The zone was no longer aligned.

That’s what makes the Nvidia deal interesting. Nvidia may be buying the open AI ecosystem precisely because it is open. The more people can download, modify, compare, and deploy models, the more experiments happen. The more experiments happen, the more demand there is for the machinery on which those experiments run.

The temporary aligned zone and the product are, for once, almost the same object.

The alignment is productive, but it isn’t necessarily benign. Every new model trained, experiment run, and inference served requires physical infrastructure: electricity, cooling, data centres, chips, water, and raw materials. The openness that makes more things possible is part of what makes more computation profitable. And the same incentive that protects the possibility space accelerates the consumption required to explore it.⁸

Creativity is to some extent a property of the geometry of the reachable set around us. Give people more things they can try, and more things become possible. But exploring those possibilities changes the constraints around what can happen next.

While the world asks itself whether Nvidia will eventually betray Hugging Face, the more interesting question is this: what would have to change for Nvidia’s interests to stop requiring Hugging Face to be free?

Perhaps compute becomes cheap enough that experimentation no longer drives hardware demand. Perhaps models become efficient enough to reverse the economics. Perhaps another technology becomes the bottleneck. Perhaps owning the place where everyone experiments becomes more valuable than having everyone experiment freely.

At that point, the alignment changes. The strange thing about these zones is that they don’t necessarily depend on the absence of power. Sometimes they depend on power having a reason to leave them alone. A temporary aligned zone doesn’t need power to disappear. It only needs power to stop needing it.

References