If you're worried about keeping powerful AI under control, start with a practical problem: downloaded copies can outlast the page that supplied them. In my Popeye picture, people can get their own copy of the powerful character. Capability no longer lives exclusively behind a few companies' front doors.
The picture is a metaphor. A screenshot from social media is not a functioning AI. You need the actual model files, compatible software, and somewhere with enough computing capacity to run them.
For a concrete example, the official Qwen3-32B model page supplies model files, a license, and instructions for running the model. This is an example of availability, not a claim that it is today's best model. Open weights also do not automatically mean every ingredient is public or every use is unrestricted. The particular release and license matter.
Once working copies have spread, asking the original developer to take a download page down does not retrieve every copy. That is the part of the policy debate I think we need to face directly.
Consider a hypothetical food pantry using a downloaded model to help volunteers find answers in an approved procedures guide. It might gain more control over how the tool runs. It would also inherit responsibilities: maintaining the software, choosing what information goes in, and checking whether answers match the actual guide. Downloading is not the same as operating a dependable service.
Having money for more computation is not the same as having unlimited intelligence either. Running a model longer may help some tasks. It does not guarantee that missing facts appear, mistaken reasoning repairs itself, or a computer gains access to the physical world. With self-hosting, the resource bill may be hardware and electricity rather than a provider's token charge.
There is a serious counterargument to openness: copying powerful systems can also spread ways to cause harm. Someone may reasonably favor restrictions before a release precisely because recovering copies later is difficult. That argument deserves a serious answer. How should we weigh access against the risks of a release whose copies we cannot reliably retrieve?
But the rules need to fit the world they are meant to govern. What specific harm does a proposal reduce? Does it address a new release, how a system is used, or access to important resources? Does it constrain an actor we can actually influence, or merely assume everyone else will stop too?
My concern is unilateral restraint that leaves us with less useful capacity while copies and development continue elsewhere. That is not proof that every regulation fails. It is a reason to distinguish effective safeguards from the comforting fiction of a worldwide off switch.
Before choosing a rule, ask what it would still accomplish after the model has been downloaded.
AI-assisted editing. Published with Rob Price's approval.