Anthropic and OpenAI have reportedly lobbied for regulations to limit open-source versions of large language models.
David Sacks, the influential investor and technologist who was until recently Donald Trump’s AI and crypto czar, has emphasized that the “frontier labs” (Anthropic and OpenAI) have lobbied for regulations on AI that would make open-source versions of large language models impossible to use. Some “effective altruists” and AI doomsday obsessives have argued for 20-year sentences for the use of open-source AI.
Open-source models are those whose code and “weights” are made public. Once published, other users can download the source and edit and customize the weights. Doing so would presumably fall afoul of the proposed regulation that all models used in America be preapproved.
One phrase that might soon start bubbling in the discourse is “Second Amendment for compute.” I first heard it from James Poulos, an idiosyncratic but valuable voice on these issues. His view is that our ability to wield the tools of the 21st century will be a core software update to our republican freedoms. Citizens can carry arms. Subjects are disarmed. And if a few corporations or the government are the sole overseers of the new technology, we are in some ways made subject to them. I’m still trying to think through the implications, but consider it a little warning of what’s to come.
I would certainly prefer a future in which digital users are more like citizens — and can own and effectively control their personal data. Imagine if data-sharing depended on revokable permission. You could, for example, give permission to the Dunkin’ Donuts app to access your location when you want it to tell you the nearest franchise locations. Or even to keep a list of your favorite drinks. But, when you decide that you need to cut the coffee milkshakes that they call lattes out of your diet, you could, with a single change in settings, stop allowing their servers to pull your data and delete your sensitive data from their files.