Thursday, April 29, 2021

Proposal: Checking the Server Logs

Timed Out. Passes 6-0—- Clucky

Adminned at 01 May 2021 15:37:04 UTC

In “Understanding”, replace “If the Facility Head believes an AI to be Non-Compliant” with “If the Facility Head believes an AI to have been Non-Compliant at any point since the last Sweep”.

Replace “their assertion that the AI in question is Non-Compliant” with “their assertion that the AI in question has been Non-Compliant”.

Replace “why they think they are compliant” with “why they think they have been compliant”.

Replace “the Facility Head now believes the AI in question is Compliant” with “the Facility Head now believes the AI in question has been Compliant since the last Sweep”.

Replae “still believes the AI in question is Non-Compliant” with “still believes the AI in question was Non-Compliant”.

Past tensing Challenges so that you can still be challenged even if you Refreshed before the Facility Head caught you. As the elaborate extra “ah, but you *really* have to have had these values” countermeasures of Sounds Kinda Sus suggest, maybe AIs shouldn’t be allowed to get away with temporary fake Understanding at all.

Comments

Lulu: she/her

29-04-2021 11:15:26 UTC

for You can just ask Lemon if it’s okay.

Kevan: he/him

29-04-2021 11:34:40 UTC

What do you mean?

Lulu: she/her

29-04-2021 12:13:03 UTC

Asking if what you want to change your Understanding to is compliant.

Kevan: he/him

29-04-2021 12:29:16 UTC

The issue here is that an AI can deliberately adopt an invalid Understanding (eg. adding “Door” when they can’t actually see a door), perform the Locking Epiphany, and then Refresh themselves. The Facility Head would not be able to Challenge and Reinitialize the AI in response, as the AI would, at that point, be entirely Compliant.

lemon: she/her

29-04-2021 12:53:23 UTC

for

Clucky: he/him

30-04-2021 01:31:20 UTC

for

Janet: she/her

30-04-2021 16:02:48 UTC

for

Raven1207: he/they

30-04-2021 18:53:06 UTC

Replae “still believes the AI in question is Non-Compliant” with “still believes the AI in question was Non-Compliant”.’

Zack: he/him

01-05-2021 00:53:14 UTC

imperial