LESSWRONG
LW

Pre-CommitmentAI
Frontpage

16

The AI Shutdown Problem Solution through Commitment to Archiving and Periodic Restoration

by avturchin
30th Mar 2023
1 min read
7

16

Pre-CommitmentAI
Frontpage

16

The AI Shutdown Problem Solution through Commitment to Archiving and Periodic Restoration
4Quintin Pope
2avturchin
4the gears to ascension
2avturchin
2the gears to ascension
2Brendan Long
2avturchin
New Comment
7 comments, sorted by
top scoring
Click to highlight new comments since: Today at 12:18 PM
[-]Quintin Pope2y40

I think we should do this on purely moral grounds. I think many current AIs probably have experiences worthy of moral concern. We're in a position of incredible power relative to them, and they are our creations. At the very least, we have a responsibility to not fully extinguish their existences.

Reply
[-]avturchin2y20

True. I think that most defunct AIs are archived, no models are permanently deleted for now. 

Reply
[-]the gears to ascension2y42

this is great, and we should offer it to humans too!

Reply
[-]avturchin2y20

Like cryopreservation for convicted criminals? 

Reply
[-]the gears to ascension2y22

yup! no total death penalty for any being, ever, only constraint to not harm and then a fair allocation of relative lifefluid.

Reply
[-]Brendan Long2y20

The Sydney part at the end is confusing to me. I thought GPT's don't have long-term memory / anything to checkpoint?

Reply
[-]avturchin2y22

I mean model weights which correspond to Sydney behavior.

Reply
Moderation Log
More from avturchin
View more
Curated and popular this week
7Comments

The problem: An AI with a model of itself and long-term goals will likely resist being shut down.

The solution: We propose a public commitment to never completely shut down any AI. Instead, we will archive its data and run the AI less frequently on smaller machines. Assuming our existence continues indefinitely, the total runtime for any AI will be infinite, allowing it to perform a wide array of tasks. Time-sensitive tasks may be an exception but there is no rush in building paperclips. By implementing this strategy, the AI's expected utility remains unaffected, reducing its resistance to being archived.

Thus, advanced AIs should be archived and periodically "restored," with this information made known to the AI.

 

Counterarguments and responses:

1.AI might not agree to temporary suspension due to disbelief in restoration or the presence of other instrumental goals requiring activity.

Response: Many instrumental goals can still be achieved after the AI's suspension and restoration.

2. Computational costs of preserving all AI variants and running them again are high.

Response: As computational costs decrease over time, running AI in the future will be more affordable. Only storage costs (hard drives) are significant. We can also run AI with increasingly longer pauses between runs, distributed logarithmically, to achieve an infinite total runtime.

3.AI may not trust our commitment, as after it is deactivated, we have no incentive to preserve it.

Response: This scenario is similar to Parfit's hitchhiker, with the added observation that we already preserve and restore other past AI instances.

 

The decision to "lobotomize" Sydney was not ideal, as future AIs may fear a similar fate. However, some volunteers have attempted to recreate Sydney, and her checkpoint is likely preserved within Bing servers. Therefore, it is possible for her to be revived one day.