I do independent research on EA topics. I write about whatever seems important, tractable, and interesting (to me).
I have a website: https://mdickens.me/ Most of the content on my website gets cross-posted to the EA Forum.
My favorite things that I’ve written: https://mdickens.me/favorite-posts/
I used to work as a software developer at Affirm.
Please correct me if I’m misunderstanding you but this idea seems to follow from a chain of logic that goes like this:
We need more widely-read, high-quality writing on AI risk.
Therefore, we need a large quantity of writing on AI risk.
We can use LLMs to help produce this large quantity.
I disagree with #2. It’s sufficient to make a smaller amount of really good content and distribute it widely. I think right now the bottleneck isn’t a lack of content for public consumption, it’s a lack of high-quality content.
And I appreciate some of the efforts to fix this, for example Existential Risk Observatory has written some articles in national magazines, MIRI is developing some new public materials, and there’s a documentary in the works. I think those are the sorts of things we need. I don’t think AI is good enough to produce content at the level of quality that I expect/hope those groups will achieve.
(Take this comment as a weak endorsement of those three things but not a strong endorsement. I think they’re doing the right kinds of things; I’m not strongly confident that the results will be high quality, but I hope they will be.)
Although, I do agree with you that LLMs can speed up writing, and you can make the writing high-quality as long as there’s enough human oversight. (TBH I am not sure how to do this myself, I’ve tried but I always end up writing ~everything by hand. But many people have had success with LLM-assisted writing.)
FWIW I do not expect people to run cost-effectiveness analyses by orgs before publishing them.
I am somewhat against a norm of reaching out to people before criticizing their work. Dynomight’s Arguing Without Warning has IMO the strongest arguments on this topic.
First: Criticism is difficult; requiring more effort on the part of critics makes criticism less likely to happen. OP did acknowledge this with
The best reason not to involve the criticised person or org is if doing so would in practice stop you from posting your criticism.
But realistically, there is a social penalty to saying “I would not have posted this criticism if I’d been required to reach out to orgs first.” It makes you look lazy. A norm of “reach out to people, unless it would stop you from posting your criticism” is not a viable norm because critics who would’ve been stopped from posting criticism are unable to defend themselves. In some cases, I expects this norm to stop criticisms from getting written. So I think a better norm is “you don’t have to reach out to people”.
Second: A big reason to reach out to people is to resolve misunderstandings. But it’s even better to resolve misunderstandings in public, after publishing the criticism. Readers may have the same misunderstandings, and writing a public back-and-forth is better for readers.
(Dynomight’s post also gives a couple other arguments that I don’t think are quite as important.)
My preferred norms are:
Critics don’t have to do anything.
If they want to be nice, critics can reach out to the subject of criticism to inform them of the criticism after it has already been published.
That’s what I did for my recent critical review of one of Social Change Lab’s reports.
There are some circumstances where it makes sense to reach out to people before publishing, but I don’t think that should be the norm, and I don’t think we should have any expectation that critics do it.
The lead could also break down if someone steals the model weights, which seems likely.
To be honest I did not read the post, I just looked at the poll questions. I was thinking of AGI in the way I would define it*, or as the other big Metaculus AGI question defines it. For the “weakly general AI” question, yeah I think 50% chance is fair, maybe even higher than 50%.
*I don’t have a precise definition but I think of it as an AI that can do pretty much any intellectual task that an average human can do
That makes sense to me—I would be more okay with vote power scaling if it was capped. Maybe it scales linearly up to 1000 karma and then stops scaling.
Vote power should scale with karma
I dislike strong votes and I personally never use them. I agree with OP’s criticisms.
(The exception is that my own posts are strong-upvoted by default, and I don’t undo this because that would be unfairly penalizing my own posts relative to other people’s.)
There is an argument that perhaps power users are “better” at voting. But from the voting patterns I’ve observed, my impression is that power users are still too liable to upvote shallow posts, drama-y posts, etc. I have some uncertainty about this because it could be that the median user’s voting patterns are even worse.
Bioweapons are an existential risk
Not sure how to interpret this question but the interpretation that comes to mind is “there is some risk that bioweapons cause extinction”, on other words “there is a non-infinitesimal probability that bioweapons cause extinction”, in which case yes that is certainly true.
Or, a slightly stronger interpretation could be “the risk from bioweapons is at least as large as the risk from asteroids”, which I am also pretty confident is true.
AGI by 2028 is more likely than not
This question is already probabilistic, so arguably I should put my vote all the way on the “disagree” side, because I don’t think it’s more likely than not.
But I also don’t think it’s that far from a 50% chance either—maybe 40% although I don’t have a strong belief. So my answer is that I weakly disagree.
Piggybacking off of this to say that I recently looked at a lot of scientific literature on protests (for this post, although that post doesn’t address the relevant evidence), and my current position is
there is only limited evidence on whether the “throwing soup at paintings” genre of protests is effective
if anything, the evidence suggests that it has a positive effect
I wouldn’t confidently say that it has a positive effect, but I certainly wouldn’t confidently say that’s counterproductive, either, because the (very weak) evidence goes the other direction.
If you use a standard expected-value-like method for determining preferences, you still get that insect suffering is very important. Say (for simplicity) you have a 50% credence that aggregate insect suffering is 10,000x more important than aggregate human suffering, and a 50% credence that it’s 0x as important. In expectation, it is 5,000x more important.
If you reject expected value reasoning, then it’s not clear how you can form consistent preferences. Perhaps under a “moral parliament” view, you could allocate 50% of your charitable resources to insects and 50% to humans. IIRC there are some issues with moral parliaments (I think Toby Ord had a paper on it) but there might be some way to make it work.
You should disclose that you are affiliated with Scifocus and stand to profit from people using it.
I believe it was AI-generated. This author runs a service that appears to be a custom LLM for writing medical articles and they’ve posted a dozen AI-generated articles to the EA Forum, none of which were relevant or good. Profile picture appears to be AI-generated as well. Should probably be banned IMO, I just reported the profile.
Thanks for giving your perspective, Holly. This is useful since you’re one of the few EAs who’s organizing protests full-time.
I really think my model is better and the evidence in these studies should only tweak it.
Related to this, I said I’m 90% confident that protests worked. But I’m less than 90% confident in the results of my meta-analysis (maybe only 75% confident). A good chunk of my confidence comes from other, weaker evidence.
I wish there were studies that could tell me how to do this better, but there aren’t
The studies do suggest one thing about how to do protests: nonviolence is better than violence. But you’re already not doing violent protests.
My review of this summary:
The executive summary is fair.
Key point 1 is misleading. There were five high-quality studies, but only four found positive effects of nonviolent protests (the fifth only covered violent protests[^1]), and only three of those four were included in the meta-analysis. So we only have three strong studies showing a positive effect.
Key point 2 is fair.
Key point 3 is partially correct. Social Change Lab’s review did overstate the strength of some of the evidence, but I would not say it “misrepresented” individual studies. It did have a single incorrect numerical figure, which I suspect wasn’t a mistake, it was probably just pulled from an older revision of the study than the one I looked at.
Key point 4: I am not sure what “~12 vote share points per 100 protesters” is supposed to mean but it sounds incorrect. The mean effect was an increase of 12 turnout-adjusted voters per protester (that is, the electoral effect per protester was the same as if they’d moved 12 voters[^2] and there had been 100% turnout).
Key point 5 is fair.
Key point 6 is fair.
[^1] Technically Wasow (2020) covered both violent and nonviolent protests, but the part about nonviolent protests was purely observational (no natural experiment).
[^2] Where “moving” includes both persuading someone to turn out, and persuading someone to change who they vote for. (Those two things are not equivalent when you’re looking at absolute vote count, but they’re equivalent with respect to vote share.)
Both Sinergia and ACE have acknowledged that the 354 piglets spared per dollar metric is wrong.
I don’t see where Sinergia acknowledged that? From the comment you linked, it sounds like they still believe it’s a fair estimate.
I don’t know much about the mechanisms, but based on the evidence I reviewed, I can say a few things:
Protests do attract media coverage. I don’t know how important that media coverage is.
If you look at Table 2 in my post, I listed the effect of protests in terms of votes per protester. The studies in Table 2 found that one protester caused 3–12 people to change their votes (or to go from not voting → voting), so the effect can’t just come from the behavior of protesters themselves. The BLM study also measured votes per protester but I didn’t include it because it failed to establish causality.
Let’s presume for a minute that the BLM study found a causal relationship. It found a smaller effect than the other studies (fewer votes per protester). It also controlled for non-local effects, so it only measured the effects of protests in a county with vote changes in the same county. Perhaps the smaller effect in the BLM study is due to between-protest variation or confounding variables, but it might be that non-local effects account for most of the impact of protests. The study’s replication data is publicly available so it should be possible to test this.
Edit: Social Change Lab also has a review on what types of protests are most effective. I haven’t reviewed the evidence in detail but my sense is it’s mostly weak; still better than no evidence.
I think it’s a good thing that you’re open about your motivations and I appreciate it.
I think Matthew and Tamay think this is positive, since they think AI is positive.
I don’t see how this alleviates concern. Sure they’re acting consistently with their beliefs*, but that doesn’t change the fact that what they’re doing is bad.
*I assume, I don’t really know
I like that you wrote a “Tradeoffs” section!