The arguments against open source AI are bad

The release of Kimi K3 this week has reopened a familiar debate. A cohort of journalists, business leaders, and politicians argue that open source AI is a dangerous threat. Dean Ball of OpenAI wrote that one probable outcome of an open-weight-model-dominant world is “full AI communism” where AI becomes a public good rather than a market product.

This framing is wrong. Tom Bedor, a software engineer and writer, published a thorough takedown of the arguments against open source AI on July 23. His piece is worth reading because it names the specific logical failures that keep recurring in policy debates and media coverage. The arguments are bad, and they keep getting made anyway.

The first error is historical. Bedor points out that open source software is the foundation of all proprietary software. Frontier models are software products built on layers of open source components. To build Uber, you need programming language frameworks, web servers, data analysis tools. Most of these are not differentiators. Commercial actors cooperate on lower layers and compete on higher ones. Frontier labs want AI models to stay in the competitive layer. Whether that happens is not up to them.

The second error is practical. Suppressing open source software is extremely difficult. Bedor cites the history of encryption as a case study. When Phil Zimmermann invented PGP in 1991, the U.S. government considered encryption military technology and opened a criminal investigation. When Netscape created SSL, the government allowed only a weakened international version. Neither effort succeeded. Export controls did not limit encryption availability. They disadvantaged Americans. Courts eventually ruled that releasing encryption source code is protected speech.

The same logic applies to AI models. Narrowing suppression to “Chinese” models raises an impossible definitional question. What makes a model Chinese? Is it Chinese if it was distilled from American models? What about an American fine-tuning a Chinese model? Regulating AI this way will encumber Americans with red tape while the rest of the world moves ahead.

The third error is the assumption that open source AI is only a Chinese government project. Bedor lists four categories of commercial actors with strong incentives to develop open source AI.

Chip makers like Nvidia benefit regardless of who runs the models. Nvidia CEO Jensen Huang has described what the company builds as “token factories.” Nvidia does not care if its chips run frontier models or cheap open source models. It has released a suite of open source models itself.

American startups like Thinking Machines Lab recently released a powerful open source model. They bet that models will be commoditized and that defensible moats can be built around auxiliary services.

Enterprise AI users will want lower-cost models for low-complexity tasks and more fine-grained control over customer-facing features.

Large companies like Google and Meta are watching OpenAI’s new ad product closely. If frontier model ad products gain traction, these behemoths have incentive to commoditize ad-free open source models to squash ad competition.

The “AI race” framing itself is incoherent. Bedor asks what the goal of this race actually is. Is it to develop the best model? To sell the most tokens? To destroy humanity first? Talking about an “AI Race” does not make more sense than talking about an “Internet Race.” The question is not who gets there first. The question is which economies absorb the transition and grow.

Specific bad arguments deserve specific rebuttals.

Scott Galloway has argued that free Chinese AI is “dumping” — the same strategy China used with solar panels, steel, EVs, and batteries. Match quality, cut price by two thirds, own the market. But AI is not a physical good. Solar panels and steel require physical supply chains where each link depends on the others. If no one manufactures solar panels in your country, it is difficult to build a business selling solar-grade silicon wafers. Software is not like that. An open source model coming from China does not prevent a fine-tuning business from succeeding in the U.S. It enables it.

The propaganda argument is equally weak. Chinese models will likely ship with a pro-China point of view. But the models are open source. If any American has an issue with the political slant, they can change and release an “Americanized” version. Within the U.S., a model seen as having a distorted pro-China bias will not outcompete a substantially similar model with a distorted pro-U.S. bias.

The backdoor argument ignores the basic market for vulnerabilities. Responsible actors patch them. Attackers exploit them. Limiting tools for responsible actors only serves attackers. If a bad actor embeds hidden adversarial behavior in a model, the best way to find it is to let anyone inspect it.

”Open source AI is too powerful and too difficult to control. It is coming, and attempts to squash it will not amount to anything more than noise along the way.”

The strongest argument against suppressing open source AI is that it will not work. History shows that suppression of open source software is extremely difficult. Attempting to do so only weakens companies against international competitors. The U.S. government learned this with encryption. The lesson applies to AI.

What this means for AI builders is straightforward. Open source models are not going away. Frontier labs can compete on capability, service, and integration. They cannot compete on suppression. The policy debate should move past whether open source AI should exist and toward how to build on top of it. That is where the actual work is.