Why the AI 'Open or Closed' Debate Is Missing a Third Option

Right now, the fight over powerful AI is treated like a light switch: either a company hands out the model's brain files for free ("open weights"), or it keeps them locked in its own servers ("closed"). Both sides have a real point. Free downloads fuel university research and tiny startups, but anyone can strip the safety training off a downloadable model and use it for scams, hacking, or worse. And once those files are out, you cannot take them back.

A new essay on the discussion site LessWrong says we should stop asking "allow or ban" and start asking "who gets what, and when?" The author compares it to how offices handle keys. Some visitors get the lobby. Some get a lab badge. A few get the master key. An AI model can be sliced the same way: the full recipe (code, data, and training methods), just the weights (the brain files), or only the text answers it types out for you.

The second half of the idea is about who actually runs the model. You could run it on your own laptop, a trusted partner like Hugging Face could run it for you, or the company could keep it on its own cloud. Pair "how open" with "who runs it" and you get a grid of options, each with different risks. Fully open plus locally run means anyone with enough computing power can rebuild it and remove its guardrails. Highly restricted means the company can track every user.

So what does this mean for you? Probably tiers. A free, downloadable AI might still exist for hobby projects, students, and small businesses. The most dangerous abilities — say, expert-level cyberattack help — could sit behind an application, an ID check, or a research license, like a controlled API. The catch: more gatekeeping means more paperwork, and whoever draws the lines holds real power. The author admits these are early thoughts from a governance course, not a finished policy.