Tier 1 Longtermist

Karma: 21

Tier 1 Longtermist 31 Jan 2023 7:25 UTC
17 points
2 ∶ 0
in reply to: Leo Gao’s comment on: Squiggle: Why and how to use it
https://github.com/rethinkpriorities/squigglepy

Tier 1 Longtermist 31 Jan 2023 3:02 UTC
1 point
0 ∶ 8
in reply to: Jason’s comment on: Announcing Interim CEOs of EVF
It seems unlikely the particular legal issues you mentioned are a great reason to have 2 CEOs?

I would find a functional/structural reason more convincing—e.g. if there were separate orgs under EVF with ops/admin/funding needs that have require distinct attention. But this seems unlikely?
My guess is that the choice behind the duo CEO involves bandwidth and resilience if things go poorly. Other factors might be really mundane, such as idiosyncratic personal considerations that are benign and normal (e.g. Howie has duties at 80,000 hours still).

Tier 1 Longtermist 31 Jan 2023 2:27 UTC
6 points
0 ∶ 0
on: Time-stamping: An urgent, neglected AI safety measure
This post was well written and well structured. You are talented and put a lot of effort into this!
Related to what others have commented, I think the cost of this is much higher than what you suggest, e.g. your guess of cost using HDD storage cost and a multiplier seems crude.
- In addition to dev time as someone else mentioned, I think the compute/network for the hash (going through literally all content) seems large, possibly multiple orders of magnitude more than the cost implied here.
Also, I’m unsure, because I haven’t fully thought through your post, but some other thoughts:
- Isn’t there really some sort of giant coordination problem here, it seems like you need to have a large fraction of all content timestamped in this way?
  - If 20% of content remain not timestamped, do issues about credibility of content remain?
- Something beyond hashing is valuable and important? (I probably need to think about this more, this could be completely wrong)
  - There’s probably content with many tiny variations and it’s better to group that content together?
    E.g., changing a few pixels or frames on a video alters the hash, instead you want some marker maps closer to the “content”
  - Finding the algorithm/implementation to do this seems important but also orders of magnitude more costly?