1. Yoav Artzi
2. Log
3. Open models, narrow interests, hard questions

Open models, narrow interests, hard questions

The discussion around open models has been largely co-opted by self interests, from all sides, masking over critical complexities. This drains this debate from impact, beyond opportunistic signaling displays, and relinquishes any chance of steering the development of AI.

Posted on: July 28, 2026

The¹ discussion around open models (i.e., open weights) has been extremely opportunistic. At times, it seems to be kept intentionally shallow by everyone involved, as if not to expose complexities that do not serve one’s goals. This is ongoing for probably 2-3 years now, but has been charged with urgency and intensity given two recent episodes: the OpenAI model hack of Hugging Face and what so far appears to be a major capability-gap-closing² step by Chinese open models.

This is an important debate. Maybe even with significant implications. Not engaging with its complexities simply means abandoning an opportunity for meaningful impact.

The opposition of frontier lab startups (OpenAI and Anthropic) is easy to attribute to their financial model, and how open models likely undermine their stratospheric valuations. It is also not hard to decipher the reason behind the possible opposition of the US government. Current open models undermine American industrial lead (but this is not necessary), and also violate the assumptions that underlie current policies.

Security and safety reasons are not invalid. Abstention harnesses are not foolproof, but there is a difference between leaving the pistol on the table and locking it in the cabinet. It is easy to imagine that regulation and certification [e.g., 1, 2] are easier with a stable cohort of large actors. One can imagine the value of having a liable party to turn to when a model hacks into your system. Hugging Face did just that.

Academics are cheering on the open side because open models are a lifeline for academic research (including mine!). Nvidia’s open source manifesto hits all the right notes, but one can imagine the financial and defensive benefits for Nvidia, when model training and deployment are distributed across many actors, none strong enough to develop their own chips [e.g., 3, 4, 5]. Hugging Face’s business model is a platform for sharing models and data.³ Other joiners to sign open manifestos seem to reflect perceptions of their own technological position or strategies of their PR departments [6]. The cynicism is at times too naked to bear. Finally, there is also plenty of (not misplaced) schadenfreude about the labs’ cry for IP theft.

There are important arguments to support the open cause. A pluralism of models may overcome risks of degeneration into a single mode of thinking, and diminish concentration of power. Open models can more easily become edge models, thereby slowing the erosion of autonomy and privacy. An ecosystem of open development is more likely to foster a creative re-thinking of both models and how they are used, maybe overcoming many of the risks and deficiencies of current paradigms. Academic research will more likely continue to prosper.

But, it is the interests that seem to largely drive this important debate, masking over critical complexities. There are two big, tightly connected factors that are at times ignored.

The first is money. Open source software prospers on contributed labor.⁴ This labor of talented engineers is costly, but it is theirs to provide, and carries no cost beyond that of their time and opportunity. Training models differs in carrying significant hardware and energy costs at a ratio unlike traditional open source projects. Where will this funding come from? The startups at the heart of the debate may have been sub-optimal in their deployment of compute, but the exploration and the process itself remain costly. This is an especially critical factor if claims of distillation hold, because it means the open models are largely hitchhiking on bills paid by the closed camp.⁵

The second factor is China’s likely industrial policy. I say “likely” because deductions about policy, intention, and implementation are always tinged with speculation, especially in non-transparent systems. It is likely impossible to fully estimate the financial autonomy of the Chinese labs, versus how much they are benefactors of various forms of state subsidies, and thereby policy instruments. The appearance of private enterprise is not as simple in the complex state-capitalism economy of China. Western economies are maybe not wholly different, but transparency is much higher. It is easy to see in China the industrial policy that underlies the rise of open-model labs. It is well-aligned with a history of nascent-industries and strategically placed subsidies. There is also potentially a broader set of reasons that make this particular industry important.

Bottom line: it is complicated. But, maybe it matters little.

That second factor might just carry the day. China is a disciplined and resource-rich actor. If this happens, the frontier lab startups should be worried. It is hard to justify their valuation based on the stickiness of the end-user apps and serving infra alone. That is unless they hit their exponentially-multiplying RSI wonderland real soon. The US industrial policy, under this scenario, is stacked on shaky scaffoldings.

---