My rough impression is that there are indeed some "AI safety" orgs that operate in the way you describe, where they are focused more on promoting US hegemony and less on preventing AI from killing everyone.* But CAIS is more on the notkilleveryoneism side of things.

*from what I've seen, the biggest offenders are CSET, Horizon Institute, and Fathom

I think that still leaves the question of why didn't Open Philanthropy (or any other big grantmakers besides SFF) fund CAIP. The original post identifies some missteps CAIP made but I also think most grantmakers' aversion to x-risk advocacy played a big role.

OP's ranking had Doom Debates at 3rd-from-bottom; I re-calculated the rankings in 3 different ways and Doom Debates came last in all of them. But I think this under-rates the expected value of Doom Debates because most of the value comes from the possibility that the channel blows up in the future.

Nice analysis, this is the sort of thing I like to see. I have some ideas for potential improvements that don't require significant effort:

• Alternative methods of evaluating videos, e.g. by view count rather than by view-minutes. I already did a bit of this in my other comment.
  • In that comment, I calculated channel rankings by views-per-dollar, and by an average of views-per-dollar + view-minutes-per-dollar.
  • You could also treat the value of a view as the logarithm of view-minutes, which feels about right to me. I couldn't calculate that from just the spreadsheet, I'd need to modify the Python script, but that still shouldn't be hard to do.
    • [ETA 3] Actually I'm not sure this is possible since I don't think there's a way to see view time for individual users. But maybe it's possible if you can see bucketed view times, e.g. "X people watched the video for 1 to 2 minutes"? Then you can take the logarithm of each bucket and get the average.
• How many views are from new viewers, and how much is it "preaching to the converted"?
  • An easy heuristic: assign more value to views-above-expectation. For example if a channel routinely gets 10,000 views, assume most of those are coming from the same people, and discount them. But if one video gets 100,000 views, treat the extra 90,000 viewers as unique.
  • I believe YouTubers get data on how many views come from subscribers and how many new subscribers they get per video. You could discount views from pre-existing subscribers, and give full credit to views from non-subscribers and maybe extra credit to views that convert to a subscription.
• [ETA] I would do a Bayesian adjustment based on the number of videos a channel has published. For example, AI in Context only has two videos and only one video with significant views, so it's hard to predict whether they will be able to reproduce that success. Whereas Rob Miles has tons of videos, so I have a pretty good idea of how many views his next video is going to get. You could use a Bayesian algorithm to assign greater confidence to channels with many videos. You could even use the variance on a channel's video views to calculate the posterior expected views (or view-minutes) on a new video.
• [ETA 2] Look at channel growth over time, not just the average. Maybe fit an exponential curve to the view count and use that to predict the views on the next video. (Or combine this method with a Bayesian adjustment.)

Before doing this, we might have guessed that it'd be most cost-effective to make many cheap, low effort-videos. AI in Context belies this; they've spent the most per video (>$100k/video vs $10-$20kthatothers)butstillachievestrongQAVM/$

I think this is an artifact of the way video views are weighted, for two reasons:

1. I suspect that impact is strongly sub-linear with minutes watched. The 100th minute watched for a particular viewer probably matters far less than the 1st minute watched. So judging impact by viewer-minutes over-rates long videos and under-rates short videos.
2. If I'm reading the code correctly, the actual viewer-minutes number is unknown, and was just assumed to be 33% for all videos. I strongly suspect that shorter videos have a higher % watch time. Sounds like the 33% number comes from talking to long-form video creators, so my guess is 33% is about accurate for long videos, but an underestimate for short videos.

Here are the results on views per dollar, rather than view-minutes per dollar:

(EDIT: apparently Markdown tables don't work and I can't upload screenshots to comments so here's my best attempt at a table. you can view the spreadsheet if this is too ugly)

1. Michael Trazzi (TikTok): $217
2. AI In Context (80k): $36
3. AI Species (Drew Spartz): $35
4. FLI podcast: $35
5. Cognitive Revolution: $23
6. The AI Risk Network: $8
7. Robert Miles AI Safety: $7
8. The Inside View: $7
9. Rational Animations: $5
10. Doom Debates: $1

For lack of any better way to do a weighting, I also tried ranking channels by the average of "views per dollar relative to average" and "view-minutes per dollar relative to average", i.e.: [(views per dollar) / (average views per dollar) + (view-minutes per dollar) / (average view-minutes per dollar)] / 2

1. Michael Trazzi (TikTok): 2.95
2. AI In Context (80k): 1.74
3. AI Species (Drew Spartz): 1.69
4. Cognitive Revolution: 1.16
5. FLI podcast: 1.07
6. The AI Risk Network: 0.60
7. The Inside View: 0.30
8. Robert Miles AI Safety: 0.21
9. Rational Animations: 0.17
10. Doom Debates: 0.08

I think this isn't a great weighting system because it ends up basically the same as ranking by views-per-dollar. That's because views-per-dollar are right-skewed with a few big positive outliers; whereas view-minutes-per-dollar are left-skewed with a few negative outliers. Ranking by geometric mean might make more sense.

These are the results ranked by geometric mean:

1. AI In Context (80k): 1.55
2. AI Species (Drew Spartz): 1.52
3. FLI podcast: 1.06
4. Cognitive Revolution: 1.02
5. Michael Trazzi (TikTok): 0.85
6. The AI Risk Network: 0.46
7. The Inside View: 0.27
8. Robert Miles AI Safety: 0.21
9. Rational Animations: 0.16
10. Doom Debates: 0.06

(here is my spreadsheet)

That's fair. I was imagining you wrote an outline and then fed the outline into an LLM. I usually prefer reading outlines over long posts, and I think it's good practice to have a summary at the top of a post that's basically your outline.

You're correct that I used AI as an editor - with limited time, it was that or no post at all.

What if you took whatever input you fed to the AI and posted that instead?

This may end up solving an upcoming problem of mine in which I wrote an org-mode doc, converted it to a Google Doc, made some changes, and might need to convert it to Markdown to publish it.

Apparently emojis don't render properly on Firefox. I didn't see any emojis so I tried opening this page on Chrome and indeed they are there, but they don't show up in my normal browser.

This comment currently has 7 agree-votes and 0 disagree-votes. Which makes the think the median EA's intuitions on protest effectiveness aren't as pessimistic as I thought.

(Perhaps people who are critical of a strategy are more likely to comment on it, which creates a skewed perception when reading comments?)