Open-source communism

Or why I think open-weights models aren't a threat to American Labs

For a long time I thought I had some kind of imposter syndrome for refusing to give opinions on topics I barely understood. It seemed simpler and less embarrassing to say “I’ve never heard of that, let me read about it first.” But that’s a naive way to think. Anyone who has spent more than ten minutes on the internet knows that people love to weigh in on things well outside their wheelhouse.

I enjoy Dean Ball’s writing and read his Substack occasionally. Particularly, his views on AI policy, economics and politics are areas I consider well outside my usual territory. Over time, thanks in part to writers like Dean, I’ve picked up a bit more than the average person might on a few specific topics.

Recently, Dean published some observations about the Kimi model, and a few of them caught me a bit of guard, particularly his points around open-weight models. This is also a topic I’ve spent considerably more time on and heard it many times before. I’m not one of the people saying he’s been captured by OpenAI’s mind-control machine. I just think he’s applying economic intuitions to a space that has historically behaved differently from traditional markets.

Likewise, I want to address the central claim: that open-weight models are inherently “decelerationist”. From what I understand Dean’s argument, roughly summarized, is that if open-source models become as capable as those from the big labs, it would destroy their business model, reduce their ability to raise capital, and ultimately constrain their research. That’s a significant leap. It might hold in a vacuum, but history gives us plenty of reasons to doubt it.

I like to think of model weights as analogous to the source code of a programming language. Google, Microsoft, and Meta all have their own languages that, from time to time, push the boundaries of their predecessors. Did any of them have trouble raising capital to fund those efforts? Of course not. You could argue the language itself is worthless (I’d disagree) and that the surrounding ecosystem is what carries the real value. Dart's value to Google isn’t in its source code. It’s in everything developers built on top of it.

What is C worth to the global economy? Probably a lot. TypeScript too. China can use both. Does that put us on the road to communism?

My bet is that as the gap between models narrows, raw intelligence will matter less to customers, and the surrounding experience will matter more. When Apple releases a faster MacBook, PC users don’t switch because of the so called performance gains. They switch because the Apple ecosystem pulls them in.

American labs won’t win because they have the best models. They’ll win because they have the best developer ecosystems, the best integrations, and the best tools. OpenAI already understands this. Dean should too.