Posts
Comments
Not addressing video recordings specifically; but we might run future iterations of this bootcamp if there's enough interest, it goes well and it continues to seem valuable. So feel free to submit the application form while noting you're only interested in future cohorts.
I was a grantee in 2019 and consent to having my evaluation shared publicly, if Nuno wants to post it.
To really make this update, I'd want some more bins than the ones Nuno provide. That is, there could be an "extremely more successful than expected" bin; and all that matters is whether you manage to get any grant in that bin.
(For example, I think Roam got a grant in 2018-2019, and they might fall in that bin, though I haven't thought a lot about it.)
Counterpoint: yes, Facebook has lots of public image issues. As a result, we have good evidence that they're an org that's unusually resistant to such problems!
They've been having scandals since they were founded. And in spite of all the things you mention, their market cap has almost doubled since the bottom of the Cambridge Analytica fall-out.
They're also one of the world's most valuable companies, and operate in a sector (software) that on an inside view seems well poised to do well in future (unlike, say, Berkshire Hathaway, which has about the same market cap).
You might have concerns about having a non-diversified portfolio in general. But modulo that, I honestly think Facebook seems like a pretty good bet.
Conditional on OpenAI API generating at least $100M in total revenue for OpenAI, by what year will that happen?
(You might also want to combine this with an estimate of the binary variable of whether it will generate $100M in revenue at all.)
I'm also posting a bounty for suggesting good candidates: $1000 for successful leads on a new project manager; $100 for leads on a top 5 candidate
DETAILS
I will pay you $1000 if you:
- Send us the name of a person…
- …who we did not already have on our list…
- …who we contacted because of your recommendation...
- ...who ends up taking on the role
I will pay you $100 if the person ends up among the top 5 candidates (by our evaluation), but does not take the role (given the other above constraints).
There’s no requirement for you to submit more than just a name. Though, of course, providing intros, references, and so forth, would make it more likely that we could actually evaluate the candidate.
NO bounty will be awarded if you...
- Mention the person who actually gets hired, but I never see your message
- Mention a person who does not get hired/become a top 5 candidate
- Nominate yourself and get hired
- If multiple people nominate the same person, bounty goes to the first person whose nomination we actually read and act on
Remaining details will be at our discretion. Feel free to ask questions in comments.
You can private message me here.
Ought (~$5000) and Rethink Priorities (~$500) have both done it, with bounties roughly what I indicated (though I'm a bit uncertain). Don't think either has completed the relevant hiring rounds yet.
In addition, I'll mention:
- Foretold is tracking ~20 questions and is open to anyone adding their own, but doesn't have very many predictions.
- In addition to the one you mentioned, Metaculus is tracking a handful of other questions and has a substantial number of predictions.
- The John Hopkins disease prediction project lists 3 questions. You have to sign up to view them. (I also think you can't see the crowd average before you've made your prediction.)
Here's a list of public forecasting platforms where participants are tracking the situation:
Foretold is tracking ~20 questions and is open to anyone adding their own, but doesn't have very many predictions.
Metaculus is tracking a handful questions and has a substantial number of predictions.
The John Hopkins disease prediction project lists 3 questions. You have to sign up to view them. (I also think you can't see the crowd average before you've made your prediction.)
This set-up does seem like it could be exploitable in an adversarial manner... but my impression from reading the poll results, is that this is weak evidence against that actually being a failure mode -- since it doesn't seem to have happened.
I didn't notice any attempts to frame a particular person multiple times. The cases where there were repeated criticism of some orgs seemed to plausibly come from different accounts, since they often offered different reasons for the criticism or seemed stylistically different.
Moreover, if asked beforehand about the outcomes of something that can be read as "an open invitation to anonymous trolling that will get read by a huge amount of people in the movement"... I would have expected to see things way, way worse than what I actually saw. In fact, I've seen many public and identifiable comments sections on Facebook, YouTube or Twitter that were much worse than this anonymous poll.
(I claim these things weakly based on having read through all the responses in the sheet. I didn't analyse them in-depth with an eye to finding traces of adversarial action, and don't expect my approach here would have caught more sophisticated attempts.)
If there were a way to do this with those opinions laundered out, then I wouldn't have a problem.
I interpret [1] you here as saying "if you press the button of 'make people search for all their offensive and socially disapproved beliefs, and collect the responses in a single place' you will inevitably have a bad time. There are complex reasons lots of beliefs have evolved to be socially punished, and tearing down those fences might be really terrible. Even worse, there are externalities such that one person saying something crazy is going to negatively effect *everyone* in the community, and one must be very careful when setting up systems that create such externalities. Importantly though, these costs aren't intrinsically tied up with the benefits of this poll -- you *can* have good ways of dispelling bubbles and encouraging important whistle-blowing, without opening a Pandora's box of reputational hazards."
1) Curious if this seems right to you?
2) More importantly, I'm curious about what concrete versions of this you would be fine with, or support?
Someone suggested:
a version with Forum users with >100 karma
Would that address your concerns? Is there anything else that would?
[1] This is to a large extent: "the most plausible version of something similar to what you're saying, that I understand from my own position", rather than than "something I'm very confident you actually belief".
Why do you think this is better than encouraging people to join foretold.io as individuals? Do you think that we are lacking an institution or platform which helps individuals to get up to speed and interested in forecasting (so that they are good enough that foretold.io provides a positive experience)?
I'm not sure if the group should fully run the tournaments, as opposed to just training a local team, or having the group leader stay in some contact with tournament organisers.
Though I have an intuition that some support from a local group might make things better. A similar case might be sports. Even though young children might start skiing with their parents, they often eventually join local clubs. There they practice with a trainer and older children, and occasionally travel together to tournaments. Eventually some of the best skiers move on to more intense clubs with more dedicated training regimes.
Trying to cache out the intuition more concretely, some of the things the local group might provide are:
- Teammates. For motivation, accountability and learning.
- A lower threshold for entering.
- Team leaders. Someone to organise the effort and who can take lots of schleps out of the process (e.g. when I did math competitions in high school I met some kids from the more successful schools, and they would have teachers who were more clued in to when the competitions happened and who would pitch it to new students, book rooms for participants and provide them with writing utensils, point them to resources to learn more, etc)
I don't think this list is exhaustive.
do you think that these tournaments would be good signaling for students applying for future EA jobs?
Yes, I think they would be.
A while back me and habryka put up a bounty for people to compile a systematic list of social movements and their fates, with some interesting results. You can find it here.
How would this be an "internal practice"? The only way this would work would be to have people publically post their earn addresses.
"Internal" in the sense of being primarily intended to solve internal coordination purposes and primarily used in messaging within the community.
I think you underrate the cost of weirdness.
You gave a particular example of a causal pathway by which weirdness leads to bad stuff, but it doesn't really cause me to change my mind because I was already aware of it as a failure mode. What makes you think I underrate the cost in comparison to the benefits of coordination?
While the kind of his status EA that might be contacted this way might get more emails then they prefer, it's important for them to be easily contacted by outsiders because that allows for valuable interactions to happen.
They'd still have a normal email. Though there is a risk of moving to an equilibrium for non-paid emails get no attention, and I haven't thought that through in detail.
It's not clear to me that we are in a mess.
Well, that's why I'm posting this -- to get some data and find out :)
(I guess the title seemed to have turned a few people off, though)
In hindsight, I should have made the intended use-cases clearer in the post. I optimised for shipping it fast rather than not at all, but that had its costs.
The reason I wrote this was basically entirely motivated by problems I've encountered myself.
For example, I’ve spent this year trying to build an AI forecasting community, and faced the awkward problem of needing a critical mass of users, but at the same time recruiting from a base with high opportunity costs and attention value (largely EA). This usually involves a pain-staking process of thinking carefully about who we message and how much, and being quite risk-averse and rather not messaging people at all when we're uncertain. I would have loved the ability to send paid emails, such that if we did happen to spam people, they could just claim some compensation. Moreover, this is a scalable strategy which would avoid the failure mode where project's like ours which think a lot about attention costs get deprioritised in favour of projects which don't.
As another example, I've considered unilaterally launching initiatives that seemed important and that no one was doing (like this!), but that very busy people might have reservations/opinions about. This put me in a spot of making awkward trade-offs along the lines analysed above.
In addition to that, I added on some problem that I've not personally experienced but which seemed like they should happen due to basic microeconomics.
This is super helpful, thanks (and that's a really awesome list of email hygiene tips, I've saved it).
I wonder whether educating and encouraging good email hygiene could be an easier solution (at least initially).
I think it would improve things on the margin, and also has a much smaller risk of landing us in a worse equilibrium, so it seems robustly good for people to do.
Still, I'm not super excited because if you believe that the initial mess is a coordination problem, the solution is not for individuals to put in lots of effort to be helpful; but for everyone to jointly move to another game where the low-effort/incentivised action is to cooperate rather than defect.
On the topic of weirdness: I expect that if what I'm pointing to is a real problem, and paid emails would help the situation, then the benefits from becoming more effective at coordinating internally would massively outweigh reputational risks from increased weirdness.
I find it somewhat hard to elucidate the reasons I believe this (though could try if you'd want me to), but some hand-wavy examples are Paul Graham's thoughts that it's almost always a mistake for startups to worry about competitors as opposed to focusing on building a good product (see paragraph 4); as well as extremely succesful organisations with pretty weird internal practices (e.g. Bridgewater, Amazon).
I think the way to answer the question is: "given the distribution of equilibria we expect following this change, what are the expected costs and benefits, and how does that compare with the costs and benefits under the current equilibrium?" (as well as considering strategic heuristics like avoiding irreversible actions and unilateralist action.)
I don't update much on your comment since it feels like it's just pointing out a bunch of costs under a particular new equilibrium, without engaging enough with how likely this is or what the benefits would be. [1]
For example:
-If Julia Wise were prioritising paid emails in her role regarding community health, is she more likely to miss emails from people on the periphery of EA or who have less money, who are potentially very vulnerable?
Here, by assumption, Julia Wise already gets so many emails that she misses some/has to prioritise. So the question is: what gets prioritised currently, and would get prioritised under the new system? There would likely be a shift towards people with more money being more able to get their issues heard -- but I'd expect it to be very small (e.g. initial email costs of $5-$25 might be enough). It might also allow her to find out about stuff she otherwise wouldn't ("I don't know if this is worth your time, though it might be, and if it wasn't, here's $10 to offset the attention cost").
Though to be clear, I've not thought a lot about community health matters, and it's not the area where I would pilot this.
[1] To be clear, I'm not claiming you should do the entire analysis, this would be an isolated demand for rigor. Just to engage more with opposing points and say why they're not convincing.
This was crossposted to LessWrong, replacing all the mentions of "EA" with "rationality", mutatis mutandis.
I'm posting this as a first step towards collecting data. Poll is a good idea, thanks!
Thanks, fixed!
I'm unfortunately only publishing the transcript at this time. The audio contains some sections that were edited out for privacy reasons.
Thanks, that's great to hear.
The prize has been going on for a while, which seems important, and I think the transparency of the Prize post is really important for making common knowledge of what kind of work there is demand for. So overall it's pretty great.
The structure of feedback looks to me like: "here's the object-level content of the post, and here are 2-3 reasons we liked it". I think you could be more clear about what you want to incentivise. More precisely, the current structure doesn't answer:
- How strong were the reasons relative to each other? (e.g. maybe removing Reason A would make the person win 2nd prize instead of 1st, but removing Reason B might make them win no prize)
- Were the reasons only jointly sufficient to merit the prize, or might accomplishing only one of them have worked?
- What other properties did the post display, which did not merit the prize? For example, maybe prize-meriting posts tend to be quite long -- even though length is not something you want to incentivise on the margin.
- Why did the posts end up ordered the way they did? Beyond "the black-box voting process gave that verdict" :) Currently I don't know why SHOW was judged as deserving 4x the prize money of "The Case for the Hotel", for example.
[Note: I double-checked with the moderators before posting this to ensure it was not too "marketingy".]
- When I and Tom came up with that, I don't think we meant "belief" to be imbued with the usual philosophical connotations. Rather, we intended it to mean something like "action-guiding, introspectively accessible representation of a state of affairs existing independently of whether it is queried".
When people ask me what I think about the world, I can often come up with lots of intelligent sounding answers -- but it is unfortunately more rare that my actual actions, plans and normative evaluations are somehow suitably hooked up to, and crucially depend upon, those answers.