While I find much of this post to be plausible, I’m not sure Ollie’s post supports your conclusions.
Ollie’s post is evaluating a set of retreats which averaged a cost of $1,500 per person. As commenters on the post noted, this seems very high. (I recall reading that low end EAG costs are around the same spot.) For the one retreat I’m aware of, costs were 6-7x less. (This doesn’t include CEA staff costs, but those shouldn’t be able to make up the gap.)
Additionally, you write about how retreats might have lower outcomes due a lack of scale. While I’m sympathetic to the idea that larger scale events can provide better sorting and outcomes per person, this doesn’t seem to be the case for the sample Ollie looked at. He notes that “Outcomes per person are approximately similar in value”.
On the whole, I don’t think the post shows that EAGx’s outperform retreats on cost effectiveness in a useful sense, mainly because of the cost issue. Ultimately there have been many more retreats and conferences since Ollie’s post, and I would love to hear from someone at CEA about their present feelings on the relative cost effectiveness of different types of “small” events.
One data point: in the recent EA community retreat I organized for 65 people in France in 2025 (not a “premium” retreat), the cost per participant was 156€. This includes my time as well as financial support from participants.
I tend to see these types of events as complementary. I think we should not treat their various outcomes as fungible. You get results of different, non-tradeable kinds. In particular:
You’re right that these aren’t the same kinds of events. My claim is about prioritisation under limited time and resources: communities should focus on the events that are more cost-effective.
If a community only has the capacity to run one event per year, I would prefer it to be a summit rather than a retreat. If the community is large enough to run an EAGx, I would prioritise that, and then add a summit around six months later if capacity allows.
Maybe I misunderstood, but if retreats and conferences/summits are complements, this argument should not apply? Two events being complements means that both together are more impactful than each alone. Retreats increase the cost-effectiveness of conferences, and conferences increase the cost-effectiveness of retreats. Hence, under the complementarity assumption, if resources allow running two events, one should run a summit and a retreat rather than an EAGx and a summit.
Take, for example, a recent intro fellowship graduate with impostor syndrome. - An EAGx alone would provide them with information and shallow connections, which could be useful to get an impactful job. However, given the impostor-syndrome assumption, the fellowship graduate would not dare to contact relevant people on EAGs due to their seniority, and they would never apply or would not take the necessary steps to build career capital (they think they would never be good enough anyway). - A retreat alone could provide them with deep connections, which could help to increase self-confidence. - Neither event alone creates impact with this hypothetical person, but together they do. With the confidence and continued support from the retreat, the graduate might have the courage to apply for positions or feel motivated to build the necessary career capital.
Similar arguments could be made for retreats increasing motivation, commitment, the feeling of being part of a community, insights about personal fit in a cause area, etc. All of these could be gained on retreats and also increase the impact of subsequent events, such as conferences.
Whether retreats and conferences are complements or substitutes is an empirical question. I would expect substantial complementarities based on anecdotal evidence (approx. five stories). If there is a dataset that tracks people across events, one could check if retreat+conference has larger effects than conference alone (however, interpretation might still be difficult due to self-selection). Better data could be gained by setting up GSF retreat RCTs: After pre-selecting a pool of eligible participants (rejecting applicants who are unlikely to benefit from the event), one could randomise admission. To gain a sufficient sample size, one could aggregate data across multiple retreats. This way, one could measure causally the direct effects (people do impactful things because of having attended a retreat) and indirect effects (people profit more from conferences because of having attended a retreat).
The example of the student is if fair, but my claim is that you can get those effects in cheaper ways than a retreat. For any given community retreat, only a couple people will attend with profiles for which it’s realistic that a retreat can help them get a role (see my point about retreats having too low of a bar). I would want to support these in other ways.
I’m worried though that the scale is too small to power an RCT. But I do like your idea on that we should try to measure the effectiveness of these interventions more rigorously!
Yep, RCT was probably the wrong word. I completely agree with the power concerns. I was more thinking that if one can get some suggestive evidence that might be better than no evidence (even if power is low and the estimated effects aren’t significant at conventional levels).
Maybe I misunderstood, but if retreats and conferences/summits are complements, this argument should not apply?
They might be complementary, but most communities don’t have the resources to do both. (Unless they are a year apart, at which point any momentum you get fades).
Yes, the sample I looked at did have some very expensive retreats, and I think you can run much cheaper ones. Note though that EAGx events / conferences can also be much cheaper so you should adjust on both sides (I think I had a cost-inflated sample due to high spending on community building in 2022)
I still think outcomes-per-person at retreats often don’t seem that different to larger events, so returns to scale are often real. i.e. focus on cost-per-attendee. If your theory of change involves helping lots of people you don’t know well find careers in EA/AIS, I think going bigger is usually a good move.
Retreats are definitely a useful intervention, especially when you have a smaller group whose needs/goals you know well (e.g. more involved community members looking to go deeper on topics).
I think it’s also worth saying: one-day conferences usually require two nights of a hotel, that the attendee pays for, unless they’re in day travel range. You can thereby quite reasonably ask for a higher entry fee for a retreat, as it would be what would otherwise be spent on a hotel.
I’m not sure if CEA gives out travel funding for EA summits, but my take is that they should do so only in very rare cases. One-day conferences, at least in most places, should focus on people who live in the city they are held.
I agree that retreat costs can vary a lot, and it’s quite possible that many retreats are cheaper today than the ones Ollie looked at. That said, the best data we currently have still suggests that retreats-as-they’ve-typically-been-run are relatively expensive per person per connection compared to larger events.
I’m very open to updating on newer or more representative data. My prior, though, is that while retreats may have become more frugal, conferences and summits have as well — EA events overall seem meaningfully leaner than they were a few years ago.
On outcomes: I don’t think I claimed that larger events produce better outcomes per person. As you note, Ollie finds those are roughly similar. My claim is instead about scale: one-day summits produce many more total positive outcomes for a similar or lower cost, largely because they involve more people and avoid subsidised travel and accommodation. (Or they should)
Given that, I would be surprised if a typical 2–3 day, fully subsidised retreat (which is the default in EA) ends up more cost-effective per person per connection than a one-day, local summit — even if some individual retreats are run quite cheaply.
I agree. One more point to highlight in this context is that the type of retreats referenced in this post, namely Group Support-Funded (GSF) Retreats, have a median cost per participant that is very different from the cost per person as evaluated in Ollie’s post. A rough calculation of median costs based on data provided in that same post on GSF retreats presents:
For University EA group retreats: Median attendees = 15, Median funding = $2061 per retreat. This results in a median cost per attendee of $137.
For City & national EA group retreats: Median attendees = 29, Median funding = $6202 per retreat. This results in a median cost per attendee of $214.
It is important to mention that some of these retreats were partly self-funded, but it seems that the relevant financial support from the EA ecosystem is at least in this approximation much smaller than the ~$1,573 figure used in Ollie’s post. The same post on GSF retreats further mentions that “compared to EAGx costs from 2022-2023, the GSF budget per attendee is between 16% and 48% of the EAGx cost per person.”
It therefore seems likely, that the retreats analysed by Ollie are a different kind of retreat or that these retreats have already gotten much less expensive and therefore potentially more cost-effective (assuming that the quality of outcomes didn’t equally drop by a factor of ~7-11).
Thanks for pointing out that we actually have some more data on costs from GSF retreats.
On the other hand, I think the argument of my post is being left unaddressed: that we should prioritise summits over community retreats. As Ollie mentioned above, larger events provide the same benefit per participant. EAGx Conferences may be at a similar cost per participant compared to retreats, but I would be really surprised if the same was true for summits, which are only one day. I would love to get a better sense of how expensive summits are per person thoguh!
While I find much of this post to be plausible, I’m not sure Ollie’s post supports your conclusions.
Ollie’s post is evaluating a set of retreats which averaged a cost of $1,500 per person. As commenters on the post noted, this seems very high. (I recall reading that low end EAG costs are around the same spot.) For the one retreat I’m aware of, costs were 6-7x less. (This doesn’t include CEA staff costs, but those shouldn’t be able to make up the gap.)
Additionally, you write about how retreats might have lower outcomes due a lack of scale. While I’m sympathetic to the idea that larger scale events can provide better sorting and outcomes per person, this doesn’t seem to be the case for the sample Ollie looked at. He notes that “Outcomes per person are approximately similar in value”.
On the whole, I don’t think the post shows that EAGx’s outperform retreats on cost effectiveness in a useful sense, mainly because of the cost issue. Ultimately there have been many more retreats and conferences since Ollie’s post, and I would love to hear from someone at CEA about their present feelings on the relative cost effectiveness of different types of “small” events.
Agreed.
One data point: in the recent EA community retreat I organized for 65 people in France in 2025 (not a “premium” retreat), the cost per participant was 156€. This includes my time as well as financial support from participants.
I tend to see these types of events as complementary. I think we should not treat their various outcomes as fungible. You get results of different, non-tradeable kinds. In particular:
Differents types of participants
Different types of impact.
You’re right that these aren’t the same kinds of events. My claim is about prioritisation under limited time and resources: communities should focus on the events that are more cost-effective.
If a community only has the capacity to run one event per year, I would prefer it to be a summit rather than a retreat. If the community is large enough to run an EAGx, I would prioritise that, and then add a summit around six months later if capacity allows.
Maybe I misunderstood, but if retreats and conferences/summits are complements, this argument should not apply? Two events being complements means that both together are more impactful than each alone. Retreats increase the cost-effectiveness of conferences, and conferences increase the cost-effectiveness of retreats. Hence, under the complementarity assumption, if resources allow running two events, one should run a summit and a retreat rather than an EAGx and a summit.
Take, for example, a recent intro fellowship graduate with impostor syndrome.
- An EAGx alone would provide them with information and shallow connections, which could be useful to get an impactful job. However, given the impostor-syndrome assumption, the fellowship graduate would not dare to contact relevant people on EAGs due to their seniority, and they would never apply or would not take the necessary steps to build career capital (they think they would never be good enough anyway).
- A retreat alone could provide them with deep connections, which could help to increase self-confidence.
- Neither event alone creates impact with this hypothetical person, but together they do. With the confidence and continued support from the retreat, the graduate might have the courage to apply for positions or feel motivated to build the necessary career capital.
Similar arguments could be made for retreats increasing motivation, commitment, the feeling of being part of a community, insights about personal fit in a cause area, etc. All of these could be gained on retreats and also increase the impact of subsequent events, such as conferences.
Whether retreats and conferences are complements or substitutes is an empirical question. I would expect substantial complementarities based on anecdotal evidence (approx. five stories).
If there is a dataset that tracks people across events, one could check if retreat+conference has larger effects than conference alone (however, interpretation might still be difficult due to self-selection).
Better data could be gained by setting up GSF retreat RCTs: After pre-selecting a pool of eligible participants (rejecting applicants who are unlikely to benefit from the event), one could randomise admission. To gain a sufficient sample size, one could aggregate data across multiple retreats. This way, one could measure causally the direct effects (people do impactful things because of having attended a retreat) and indirect effects (people profit more from conferences because of having attended a retreat).
The example of the student is if fair, but my claim is that you can get those effects in cheaper ways than a retreat. For any given community retreat, only a couple people will attend with profiles for which it’s realistic that a retreat can help them get a role (see my point about retreats having too low of a bar). I would want to support these in other ways.
I’m worried though that the scale is too small to power an RCT. But I do like your idea on that we should try to measure the effectiveness of these interventions more rigorously!
Yep, RCT was probably the wrong word. I completely agree with the power concerns. I was more thinking that if one can get some suggestive evidence that might be better than no evidence (even if power is low and the estimated effects aren’t significant at conventional levels).
They might be complementary, but most communities don’t have the resources to do both. (Unless they are a year apart, at which point any momentum you get fades).
Quick take:
Yes, the sample I looked at did have some very expensive retreats, and I think you can run much cheaper ones. Note though that EAGx events / conferences can also be much cheaper so you should adjust on both sides (I think I had a cost-inflated sample due to high spending on community building in 2022)
I still think outcomes-per-person at retreats often don’t seem that different to larger events, so returns to scale are often real. i.e. focus on cost-per-attendee. If your theory of change involves helping lots of people you don’t know well find careers in EA/AIS, I think going bigger is usually a good move.
Retreats are definitely a useful intervention, especially when you have a smaller group whose needs/goals you know well (e.g. more involved community members looking to go deeper on topics).
I think it’s also worth saying: one-day conferences usually require two nights of a hotel, that the attendee pays for, unless they’re in day travel range. You can thereby quite reasonably ask for a higher entry fee for a retreat, as it would be what would otherwise be spent on a hotel.
I’m not sure if CEA gives out travel funding for EA summits, but my take is that they should do so only in very rare cases. One-day conferences, at least in most places, should focus on people who live in the city they are held.
Thanks for these comments!
I agree that retreat costs can vary a lot, and it’s quite possible that many retreats are cheaper today than the ones Ollie looked at. That said, the best data we currently have still suggests that retreats-as-they’ve-typically-been-run are relatively expensive per person per connection compared to larger events.
I’m very open to updating on newer or more representative data. My prior, though, is that while retreats may have become more frugal, conferences and summits have as well — EA events overall seem meaningfully leaner than they were a few years ago.
On outcomes: I don’t think I claimed that larger events produce better outcomes per person. As you note, Ollie finds those are roughly similar. My claim is instead about scale: one-day summits produce many more total positive outcomes for a similar or lower cost, largely because they involve more people and avoid subsidised travel and accommodation. (Or they should)
Given that, I would be surprised if a typical 2–3 day, fully subsidised retreat (which is the default in EA) ends up more cost-effective per person per connection than a one-day, local summit — even if some individual retreats are run quite cheaply.
I agree. One more point to highlight in this context is that the type of retreats referenced in this post, namely Group Support-Funded (GSF) Retreats, have a median cost per participant that is very different from the cost per person as evaluated in Ollie’s post. A rough calculation of median costs based on data provided in that same post on GSF retreats presents:
For University EA group retreats: Median attendees = 15, Median funding = $2061 per retreat. This results in a median cost per attendee of $137.
For City & national EA group retreats: Median attendees = 29, Median funding = $6202 per retreat. This results in a median cost per attendee of $214.
It is important to mention that some of these retreats were partly self-funded, but it seems that the relevant financial support from the EA ecosystem is at least in this approximation much smaller than the ~$1,573 figure used in Ollie’s post. The same post on GSF retreats further mentions that “compared to EAGx costs from 2022-2023, the GSF budget per attendee is between 16% and 48% of the EAGx cost per person.”
It therefore seems likely, that the retreats analysed by Ollie are a different kind of retreat or that these retreats have already gotten much less expensive and therefore potentially more cost-effective (assuming that the quality of outcomes didn’t equally drop by a factor of ~7-11).
Thanks for pointing out that we actually have some more data on costs from GSF retreats.
On the other hand, I think the argument of my post is being left unaddressed: that we should prioritise summits over community retreats. As Ollie mentioned above, larger events provide the same benefit per participant. EAGx Conferences may be at a similar cost per participant compared to retreats, but I would be really surprised if the same was true for summits, which are only one day. I would love to get a better sense of how expensive summits are per person thoguh!