Store a copy of the Accelerating Alignment AI Safety Research Dataset in the Arctic World Archive

David Staley
Not accepting new donations
0 donations
Support score: 0AI governanceCivilizational resilience

In the event that a global catastrophic event leads to a temporary (on the scale of generations to centuries) loss of technological capacity it is likely that our progress in cutting edge technologies such as AI (and therefore alignment) would be lost and thus the long term danger of misaligned AI would remain.

If we preserve alignment research through this disruption, alignment may get a significant lead in the post-disruption development of AI. It may even be that AI safety becomes the dominant paradigm around which AI development occurs.

The accelerating alignment team from ElutherAI have already assembled an archive of alignment research, and are actively updating/refining it. The Arctic World Archive seems like a good place to preserve such research.

After storing the data we would publish the unique reference code to the data reel on the EA/Alignment Forum with the goal of ensuring knowledge of the archive is widely spread.


After communicating with the AWA team, I believe the optimal method of implementing this project is to purchase 5-10GB (the current dataset size is 1GB) of storage on a 25 year contract at a cost (in euros) of 399(5GB) - 699(10GB) plus membership fees of 108/year.

We would then ideally upload updated snapshots of the archive on a yearly basis.

Thus the total yearly cost would be 508-808 euros.


We can publish screenshots of the AWA data portal, as well as share access to the AWA user account with a trusted 3rd party for verification.

If further proof is required, we can also opt for a public ceremony at the vault (for significant additional costs) where photo/video coverage is provided, along with a shiny certificate.

Potential Harms

The dataset may end up containing infohazards.
The users of the dataset may lack the contextual/meta understanding to utilize it.
The dataset could simply result in an awareness of AI tech that restarts the 'arms race' without significant focus on alignment.


By far most of the actual work done has been by the Accelerating Alignment team from the EleutherAI community, the dataset project's github page here.

The project was suggested by Plex, and research and implementation has been carried out by myself.