Fair points — I should have been more careful with the “free access” framing. Here’s a quick revision:
Yes, UBI is rationed per-person. These commons should be too. In both UBI and the commons, we abolish evaluative rationing — rationing based on perceived deservingness.
I’ll split the “projects you would not approve of” concern in two parts. The first is a harm worry (dangerous use); the second is a dilution worry (most compute goes to non-safety projects).
Amount-rationing could address harm by not giving enough compute to do harm, but it would not address dilution concerns, which may be as much, if not more, of a problem for funders.
Which is why, upon further reflection, I’d like to introduce a second gate: a gate by type. Shevlane (2022) and Bucknall & Trager (GovAI, 2023) have argued basically this — that selective access can preserve safety-positive work while limiting misuse.
So yes, I think my model needs some gates. But I would still push hard to keep the gates non-evaluative. That’s how, I think, to inspire safety research that would never have been done, by people who’d never have applied for a grant. And that’s what the commons is for.
Fair points — I should have been more careful with the “free access” framing. Here’s a quick revision:
Yes, UBI is rationed per-person. These commons should be too. In both UBI and the commons, we abolish evaluative rationing — rationing based on perceived deservingness.
I’ll split the “projects you would not approve of” concern in two parts. The first is a harm worry (dangerous use); the second is a dilution worry (most compute goes to non-safety projects).
Amount-rationing could address harm by not giving enough compute to do harm, but it would not address dilution concerns, which may be as much, if not more, of a problem for funders.
Which is why, upon further reflection, I’d like to introduce a second gate: a gate by type. Shevlane (2022) and Bucknall & Trager (GovAI, 2023) have argued basically this — that selective access can preserve safety-positive work while limiting misuse.
So yes, I think my model needs some gates. But I would still push hard to keep the gates non-evaluative. That’s how, I think, to inspire safety research that would never have been done, by people who’d never have applied for a grant. And that’s what the commons is for.