The Arguments Against Open Source AI are Very Bad

Kimi K3's release kicked off another round of hand-wringing about open source AI, and honestly, most of it doesn't hold up. Dean Ball of OpenAI has floated the idea that a world dominated by open-weight models leads to something like "full AI communism," treating AI as a public good rather than a product. Strip away the rhetoric and the actual position from frontier labs boils down to this: open models are risky, so access should run through trusted gatekeepers — who, conveniently, tend to be the labs selling the expensive stuff to their best customers.

The framing ignores something basic about how software works. Every commercial product, including frontier AI models, sits on a mountain of open source code underneath it. Companies cooperate on the boring, foundational layers of the stack and compete on the parts that actually differentiate them. Frontier labs would obviously prefer that AI models stay in the differentiated category rather than sliding into commodity territory, but wanting that doesn't make it so.

History isn't kind to the suppression argument either. Encryption tools like PGP and SSL were treated as sensitive technology decades ago — Phil Zimmermann faced a criminal investigation in 1991 just for creating PGP, and the government forced Netscape to ship a weakened international version of SSL. None of it worked. The weakened SSL version ended up more popular even domestically, and courts eventually ruled that publishing encryption code counts as protected speech. Trying to draw a line around "Chinese" models runs into the same wall: nobody can cleanly define what makes a model Chinese once distillation and fine-tuning enter the picture, and red tape mostly just handicaps American developers relative to everyone else.

It's also wrong to treat open source AI as some uniquely Chinese strategy. Nvidia's Jensen Huang talks about building "token factories" and doesn't particularly care whether those tokens come from frontier or cheap open models — Nvidia has released its own open models. Startups like Thinking Machines Lab are betting that models commoditize and the real value shifts to services built around them. Even big incumbents like Google and Meta have incentive to push out free, ad-free open models if it helps them undercut a rival's ad business.

The specific fears people raise about Chinese models mostly collapse on inspection. Scott Galloway compares it to China undercutting solar panels and steel, but AI software doesn't depend on physical supply chains the way manufacturing does — a free model from China doesn't stop an American fine-tuning business from thriving. Worries about propaganda ignore that open weights mean anyone can retrain a model with a different slant. And backdoor fears miss that open inspection is exactly what lets responsible actors find vulnerabilities before attackers do; locking tools away just helps the attackers.

None of this changes the outcome much. Open source AI is too powerful and too diffuse to gatekeeper effectively, and the political noise around it isn't going to change that trajectory.