Sorry, it wasn't clear. The reference class I had in mind was cause prio focussed resources on the EA forum.
I think people/orgs do some amount of this, but it's kind of a pain to share them publicly. I prefer to share this kind of stuff with specific people in Google Docs, in in-person conversations, or on Slack.
I also worry somewhat about people deferring to random cause prio posts, and I'd guess that on the current margin, more cause prio posts that are around the current median in quality make the situation worse rather than better (though I could see it going either way).
Thanks for writing this.
I disagree with quite a few points in the total utilitarianism section, but zooming out slightly, I think that total utilitarians should generally still support alignment work (and potentially an AI pause/slow down) to preserve option value. If it turns out that AIs are moral patients and that it would be good for them to spread into the universe optimising for values that don't look particularly human, we can still (in principle) do that. This is compatible with thinking that alignment from a total utilitarian perspective is ~neutral - but it's not clear that you agree with this from the post.
Oh, I thought you might have suggested the live thing before, my mistake. Maybe I should have just given the 90-day figure above.
(That approach seems reasonable to me)
I answered the first questions above in an edit of the original comment. I’m pretty sure when I re-ran the analysis with decided in last 30 days it didn’t change the results significantly (though I’ll try and recheck this later this week - in our current setup it’s a bit more complicated to work out than the stats I gave above).
I also checked to make sure that only looking at resolved applications and only looking at open applications didn’t make a large difference to the numbers I gave above (in general, the differences were 0-10 days).
Oh, right - I was counting "never receiving a decision but letting us know" as a decision. In this case, the number we'd give is days until the application was withdrawn.
We don't track the reason for withdrawals in our KPIs, but I am pretty sure that process length is a reason for a withdrawal 0-5% of the time.
I might be missing why this is important, I would have thought that if we were making an error it would overestimate those times - not underestimate them.
I'm not sure sorry, I don't have that stat in front of me. I may be able to find it in a few days.
Empirically, I don't think that this has happened very much. We have a "withdrawn by applicant status", which would include this, but the status is very rarely used.
In any case, the numbers above will factor those applications in, but I would guess that if we didn't, the numbers would decrease by less than a day.
We do have a few processes that are designed to do this (some of which are doing some of the things you mentioned above). Most of the long delays are fairly uncorrelated (e.g. complicated legal issue, a bug in our application tracker ...).
I think it could be good to put these number on our site. I liked your past suggestion of having live data, though it's a bit technically challenging to implement - but the obvious MVP (as you point out) is to have a bunch of stats on our site. I'll make a note to add some stats (though maintaining this kind of information can be quite costly, so I don't want to commit to doing this).
In the meantime, here are a few numbers that I quickly put together (across all of our funds).
Grant decision turnaround times (mean, median):
Thanks for engaging with my criticism in a positive way.
Regarding how timely the data ought to be, I don't think live data is necessary at all - it would be sufficient in my view to post updated information every year or two.
I don't think "applied in the last 30 days" is quite the right reference class, however, because by-definition, the averages will ignore all applications that have been waiting for over one month. I think the most useful kind of statistics would:
Is there (or might it be worthwhile for there to be) a business process to identify aged applications and review them at intervals to make sure they are not "stuck" and that the applicant is being kept up to date? Perhaps "aged" in this context would operationalize as ~2x the median decision time and/or ~>90-95th percentile of wait times? Maybe someone looks at the aged list every ~2 weeks, makes sure the application isn't "stuck" in a reasonably fixable way, and reviews the last correspondence to/from the applicant to make sure their information about timeframes is not outdated?
If it's okay with you, I'd prefer not to have screenshots of my emails posted right now; I'm happy to rethink this in a few days when it feels a bit lower pressure.
I generally don't write emails, assuming that they will be posted to a public place like the forum.
I believe our application form says
...The Animal Welfare Fund, Long-Term Future Fund and EA Infrastructure Fund aim to respond to all applications in 2 months and most applications in 3 weeks. However, due to an unprecedentedly high load, we are currently unable to achieve our desired speedy turnarounds. If you need to hear back sooner (e.g., within a few weeks), you can let us know in the application form, and we will see what we can do. Please note that: EA Funds is low on capacity and may not be able to get back to you by either your stated deadline or t
One thing to note is that at the end of January, we rejected the original grant (which I believed that we wouldn't be able to show a clear public benefit for), and then said we were interested in a different version of the grant that seemed more defensible to me (subject to legal review). Since then, we have been working out whether we can make this alternate grant.
I didn't realise that Igor stopped taking clients completely, and I regret that I didn't make a stronger effort to understand the consequences of the unclear situation whilst we tried to understand the legal implications of making the grant.
I’m very sorry that you had such a bad experience here. Whilst I would disagree with some of the details here I do think that our communication was worse than I would have liked and I am very sorry for any hardship that you experienced. It sounds like a stressful process which could have been made much better if we had communicated more often and more quickly.
In my last email (March 4th), I said that we were exploring making this grant, but it’s legally challenging. Grants for mental health support are complicated, in general, as we have to show that there...
One thing to note is that at the end of January, we rejected the original grant (which I believed that we wouldn't be able to show a clear public benefit for), and then said we were interested in a different version of the grant that seemed more defensible to me (subject to legal review). Since then, we have been working out whether we can make this alternate grant.
I didn't realise that Igor stopped taking clients completely, and I regret that I didn't make a stronger effort to understand the consequences of the unclear situation whilst we tried to understand the legal implications of making the grant.
If the funders get a nontrivial portion of the impact for early-stage projects then I think the AWF (inc. its donors) is very plausible.
Here are some very brief takes on the CCM web app now that RP has had a chance to iron out any initial bugs. I'm happy to elaborate more on any of these comments.
I don't understand the point about the complexity of value being greater than the complexity of suffering (or disvalue). Can you possibly motivate the intuition here? It seems to me like I can reverse the complex valuable things that you name, and I get their "suffering equivalents" e.g. (e.g. friendship -> hostility, happiness -> sadness, love -> hate ... etc.), and they don't feel significantly less complicated.
I don't know exactly what it means for these things to be less complex; I'm imagining something like writing a Python program that simulates the behaviour of two robots in a way that is recognisable to many people as "friends" or "enemies" and measuring at the length of the program.
Oli’s comment so people don’t need to click through
...I thought some about the AI Safety camp for the LTFF. I mostly evaluated the research leads they listed and the resulting teams directly, for the upcoming program (which was I think the virtual one in 2023).
I felt unexcited about almost all the research directions and research leads, and the camp seemed like it was aspiring to be more focused on the research lead structure than past camps, which increased the weight I was assigning to my evaluation of those research directions. I considered for a while to
I agree that people in existing EA hubs are more likely to come across others doing high value work than people located outside of hubs.
That said, on the current margin, I still think many counterfactually connections happen at office spaces in existing EA hubs. In the context of non residential spaces, I’m not really sure who would use an EA office space outside existing EA hubs so I’m finding the comparison between office in a hub vs office outside a hub a little confusing (whereas with CEEALAR I understand who would use it).
I go back and forth on this. Sometimes, I feel like we are funding too many underperforming projects, but then some marginal project surprises me by doing quite well, and I feel better about the hits-based strategy. Over the last three months, we have moved towards funding things that we feel more confident in, mostly due to funding constraints.
I don't think that I have a great list of common communicable lessons, some high-level thoughts/updates that jump to mind:
I think the performance/talent of grantees and context is extremely important.
That said, some programs that I am excited about that I believe many EAs are a good fit for:
Some projects I have seen work well in the past, but I think they are a bad fit for most people:
Most of my views o...
(note that I'm not speaking about CEEALAR or any other specific EAIF applicants/grantees specifically)
I understand that CEEALAR has created a low-cost hotel/coworking space in the UK for relatively junior people to stay while they work on research projects relevant to GCRs. I think that you had some strategic updates recently so some of my impression of your work may be out of date. Supporting people early on in their impact-focused careers seems really valuable, I've seen lots of people go through in-person retreats and quickly start doing valuable work.
A...
Doctors in the UK (like the ones that set this up) earn way less that $350k a year in general. Junior doctors (which are the majority of the UK doctor workforce) are very poorly paid, I think many of my friends made something like £14/hour for the first few years after qualifying.
I didn't say that AI was software by definition - I just linked to some (brief definitions) to show that your claim afaict is not widely understood in technical circles (which contradicts your post). I don't think that the process of using Photoshop to edit a photo is itself a program or data (in the typical sense), so it seems fine to say that it's not software.
Definition make claims about what is common between some set of objects. It's fine for single members of some class to be different from every other class member. AI does have a LOT of basic stuff ...
I'll probably ask some of my ML engineer friends this week, but I am fairly sure that most ML people would be fine with calling AI products, models, etc. software. I don't have much of an opinion on whether calling AI systems software creates confusion or misunderstandings - I'd guess that calling AI software within policy circles is generally helpful (maybe you have a better alternative name).
...Software, instructions that tell a computer what to do. Software comprises the entire set of programs, procedures, and routines associated w
Appealing to definitions seems like a bad way to argue about whether the conceptual model is useful or not. The operation of a computer system and the "software" used for digital photography, or videoconferencing, or essay writing, is not typically considered software. Do you think those should be called software, given that they fit into the definitions given?
I'm claiming that AI is distinct in many ways from everything else we typically think of as software, not that it doesn't fit a poorly scoped definition. Amd the examples of "collection[s...
Thanks for pointing that out. I re-read the post and now think that the OP was more reasonable. I'm sorry I missed that in the first place. I also didn't convey the more important message of "thank you for critiquing large, thorny, and important conclusions". Thinking about P(bio x-risk) is really quite hard relative to lots of research reports posted on the forum, and this kind of work seems important.
I don't care about the use of Bayesian language (or at least I think that bit you quoted does all the Bayesian language stuff I care about).
Maybe I should r...
I was a bit disappointed by this post. I think I am sympathetic to the overall take and I’m a bit frustrated that many EAs are promoting or directly working on biorisk without imo compelling reports to suggest a relatively high chance of x-risk.
That said, this post seems to basically make the same error, it says that Ord’s estimates are extremely high but doesn’t really justify that claim or suggest a different estimate. It would be much more reasonable imo to say “Ord’s estimate is much higher than my own prior, and I didn’t see enough evidence to justify such a large update”.
In fairness, I was prevented from posting a bunch of stuff and spent a long time (like tens of hours) workshopping text until legal council were happy with it. At least in one case I didn’t end up posting the thing because it didn’t feel useful after the various edits and it had been by then a long time since the event the post was about.
I think in hindsight the response (with the information I think the board had) was probably reasonable - but if similar actions were to be taken by EV when writing a post about Anthropic I’d be pretty upset about that. I wouldn’t use the word censoring in the real ftx case - but idk in the fictional Anthropic case I might?
I'm not sure that I would use the word censoring, but there were strict policies around what kinds of communications various EV orgs could do around FTX for quite a long time (though I don't think they were particularly unusual for an organisation of EVs size in a similar legal situation).
The wording of what Larks said makes it seem like over a number of years staff were prevented from expressing their true opinions on central EA topics
EV was fine with me publishing this. My experience was that it was kind of annoying to publish FTX stuff because you had to get review first, but I can't recall an instance of being prevented from saying something.
"Aggressively censored its staff" doesn't reflect my experience, but maybe reflects others', not sure.
I'm excited that Zach is stepping into this role. Zach seems substantially better than my expectations for the new CEA CEO, and I expect the CEO hiring committee + Ben + the EV board had a lot to do with that (and probably lots of other people at CEA that I don't know about)!
Most CEA users and EA community members probably don't know Zach, so I thought it would be helpful to share some of my thoughts on them and this position (though I don't know Zach especially well, and these are just quick subjective takes). Thanks to @Ben_West for the nudge to do this....
Zach is on Anthropic's Long-Term Benefit Trust. It's not super clear what this means, particularly in light of recent events with the Open AI board, but I am a bit concerned about the way that EA views Anthropic, and that the CEO of CEA being affiliated with Anthropic could make it more difficult for people within EA to speak out against Anthropic.
This is a very interesting point given that it seems that Helen's milquetoast criticism of OpenAI was going to be used as leverage to kick her off the OpenAI board, and that historically EV has aggressively censored its staff on important topics.
That would help me! Right now I mostly ignore the expertise/interest fields, but I could imagine using this feature to book 1:1s if people used a convention like the one you suggested.
+1 to the EAG expertise stuff, though I think that it’s generally just an honest mistake/conflicting expectations, as opposed to people exaggerating or being misleading. There aren’t concrete criteria for what to list as expertise so I often feel confused about what to put down.
@Eli_Nathan maybe you could add some concrete criteria on swapcard?
e.g. expertise = I could enter roles in this specialty now and could answer questions of curious newcomers (or currently work in this area)
interest = I am either actively learning about this area, or have invested at least 20 hours learning/working in this area .
I think the closest thing to an EA perspective written relatively recently that is all in a single doc is probably this pdf of Holdens most important century sequence on cold takes.
I feel quite confused about what empirical evidence from the forum would change your mind. I personally don't think that the comment folding of weakly negative comments has much effect on future engagement (particularly if it's a reasonable comment).
I initially thought that if your hypothesis about hiding influencing voting is correct and people regularly do the voting on expected Karma, then I think you should expect to see quite a few "reasonable to you" low negative karma comments (as they have been voted negative and then stayed negative due to hiding ...
I think downvoting with the expectation of upvotes is very reasonable. It's probably a better norm than voting based on current karma when only some people follow the vote based on current karma level policy and you care about the steady state karma behaviour.
Nevertheless, my impression is that GiveWell's cost-effectiveness estimates are pretty close to encompassing all their thinking. Elie mentioned on the Clearer Thinking podcast that:
Fwiw I feel quite confused about how different GiveWell's recommendations would be if they were solely optimising for cost-effectiveness, I have heard different versions of how much they are optimising for this already based on different people that I speak to (and my impression is that most public materials do not say they are solely optimising for this).
...Do you have a cost
Hi Vasco, quickly responding to a few of your points here
Let its donors know that donating to GHDF in its current form has a similar effect to donating to AGF (if that is in fact the case), instead of just describing GHDF as a “higher-risk” “higher-reward” alternative (to TCF). “Donating to this fund [GHDF] is valuable because it helps demonstrate to GiveWell that there is donor demand for higher-risk, higher-reward global health and development giving opportunities”.
...Probably most controversially, while I view existential risk reduction work as tremendously important, I don’t donate any of the 10% of my income dedicated to effective charity in order to support this work. (I do view it as a critical global priority, which is why the vast majority of my time and effort are spent on it!) Principally, my lack of donations is because I don’t view the cause area as a charitable endeavor, rather than rational self-interest for myself and my family, which has obvious benefits to the broader world. This does not make it less i
I think David means "giving motivated by impartiality" instead of giving to places that themselves are "cause neutral".
I'm not too worried about this kind of moral uncertainty. I think that moral uncertainty is mostly action-relevant when one moral view is particularly 'grabby' or the methodology you use to analyse an intervention seems to favour one view over another unfairly.
In both cases, I think the actual reason for concern is quite slippery and difficult for me to articulate well (which normally means that I don't understand it well). I tend to think that the best policy is to maximise the expected outcomes of the overall decision-making policy (which involves paying...
Could you give me some examples of these kinds of projects? I think, as Linch said, Manifund is probably their best bet, or posting on the EA forum asking for funding from individual donors.
I'd like to hear his advice for smart undergrads who want to build their own similarly deep models in important areas which haven't been thought about very much e.g. take-off speeds, the influence of pre-AGI systems on the economy, the moral value of insects, preparing for digital minds (ideally including specific exercises/topics/reading/etc.).
I'm particularly interested in how he formed good economic intuitions, as they seem to come up a lot in his thinking/writing.
Thanks! To be clear, this is a 'plan' instead of something we are 100% committed to delivering on in the way it's presented below. I think there are some updates to be made here, but I would feel bad if you made large irreversible decisions based on this post. We will almost certainly have a more official announcement if we do decide to commit to this plan.
I agree with the overall point, though I am not I've seen much empirical evidence for the GHD as a good starting point claim (or at least I think it's often overstated). I got into EA stuff though GHD, but, this may have just been because there were a lot more GHD/EA intro materials at the time. I think that the eco-system is now a lot more developed and I wouldn't be surprised if GHD didn't have much of an edge over cause first outreach (for AW or x-risk).
Maybe our analysis should be focussed on EA principles, but the interventions themselves can be bran...
I could see people upvoting this post because they think it should be more like -10 than -21. I personally don't see it as concerning that it's "only" on -21.