Crossposted on Substack.

TLDR.:

In this post, I challenge the prevalent mindset of “just send people to Bluedot” in the AI Safety community.  Bluedot offers valuable programs, but their work does not justify discouraging local fieldbuilding initiatives, as they are a great way to introduce more people to AI Safety, gain skills and career capital for organizers, promote innovation in fieldbuilding, and result in positive externalities. I also discuss the reasons for the status quo and arguments against my position.

Identifying the problem

This isn't about Bluedot doing anything wrong. I think they are awesome and it’s not easy to find an organisation that has done more for the field of AI safety than them. However, I often see that other people look at what Bluedot does and end up discouraging up-and-coming fieldbuilders from taking initiative. I hope to shed light on this issue and foster a more supportive approach within our community.

Whenever you are in conversations about AIS field-building initiatives, sooner or later you will hear the following phrase: “Why not just send people to Bluedot?”. This response is often given to those considering running courses for their AI Safety group. I think this is one of the most harmful-yet-widespread memes in the community that I have heard many people repeat, including grantmakers and experienced fieldbuilders. I decided to write up some of the reasons why building local and national initiatives can be beneficial, and even encourage you to take up fieldbuilding yourself. 

Why consider local initiatives alongside Bluedot

We want more people to take AI Safety courses

As of March 2025, Bluedot’s acceptance rates for recent courses were:

• Governance Track (12 weeks) - 47%
• Alignment Track (12 weeks) - 28%
• Economics of Transformative AI (9 weeks) -  27%
• Intro to Transformative AI (5 days) = ~60-70%

While Bluedot’s capacity and applicant numbers are growing, this increase means that they will accidentally overlook more and more promising individuals. They are not infallible. I have talked to several people who were rejected from the course, but who I’m excited about.^[1] I also wonder how many people who are now working in AI Safety got rejected at some point early in their career transition journey.

Another thing worth mentioning is that they have only run a given technical or policy course every ~3 months or so. Now they have a shorter intro-course that runs almost weekly, but if you just miss the deadline for their sweet 12-week governance course, you might have to wait several months until you can apply again.

If there were an alternative for these people to be able to take a course from a local/national AI safety group and get involved with the community this way, I think that would be great. Many highly motivated individuals from around the world expressed gratitude for AIS Hungary’s online course, as they either missed Bluedot’s deadline or were rejected.

The AIS community is still incredibly small

For a couple of months, catastrophic risks from AI have gotten a lot more attention than before, but the number of people working on AI Safety is still extremely small. To date, the biggest conference on AI existential risk was the 2024 GCR Bay Area conference with ~1,200 people. This included people from biorisk and nuclear security, so let’s call it a 1,000 participants. For comparison, the biggest neuroscience conferences have around 25,000 attendees.

My favourite intuition pumps about the size of our communities come from this post^[2] that points out that there are approximately 30 times as many genius “swifties” (hardcore Taylor Swift fans) as genius rationalists.

Another example compares the total number of employees in all EA/rationalist organizations to those at the Bill and Melinda Gates Foundation, and the difference is mind-blowing.

If we want to change the picture above, we need to do more outreach.

Gaining skills and career capital

It’s well-known that the AIS community is mentor and management-constrained^[3], and it’s really hard to get a job in an established organisation.

Unless you land a job at one of those orgs, communitybuilding (thanks to being a high-absorbency career opportunity) allows you to:

• upskill in project management
• improve your epistemic understanding
• improve your soft skills
• have an impact
• attend EAGs
• maybe even get paid

Historically, fieldbuilding has been a great way for people to upskill and gain career capital, and many people who started with organising a local community went on to do even more impressive things.

No, you won’t have as many applicants as Bluedot, and your courses probably won’t run as smoothly, but apart from teaching people about AI Safety and gaining new skills, you might end up building something really cool.

We want more innovation in fieldbuilding

With more communitybuilders, we will have more learnings on what works and doesn't work (assuming you share your learnings, which people don’t do enough of!)

Experienced fieldbuilders will tell you that their work is far from being a hard science, and there are many uncertainties about strategy we should explore, and plenty of low-hanging fruit to pick (see here for some examples and potentially get involved).

Organising local/national groups has positive externalities

I got into the community by organising an EA university group, which eventually expanded to a city and national group. Since then, we have spun out or are in the process of spinning out several initiatives that are now doing their own things:

• AIS Hungary
  • The ops team of AIS Hungary is now helping out with AIS Collab
  • In collaboration with ENAIS, it helped seed AIS Dublin
  • Our technical lead is helping a Bluedot teaching fellow create an online ARENA, which will fundraise to be an independent organisation
  • Our policy lead just finished creating a 201 governance curriculum (to be published soon!)
• Amplify
  • Started with me experimenting with social media ads in Hungary
  • Now employs a marketing professional with 15+ years of experience, helping a wide range of fieldbuilding organisations with their digital presence
  • We are working to expand to broader AIS communications work
• Scaling Altruism
  • Helps less established EA groups run intro fellowships, and more established ones save operations time on running their courses

Without being able to first organise a national group in a professional capacity, our team wouldn’t have been able to start any of the projects above. Starting a group allows you to get more context, and you can eventually “graduate” to more difficult projects.

International initiatives I know of that started from local/national groups:

• SPAR was originally organised by Stanford AI Safety, Berkeley AI Safety Initiative, and Georgia Tech AI Safety. It eventually expanded its scope beyond university students
• LARS Labs came out of AI Safety Hub Labs, which grew out of OxAI Safety & Governance Labs (Oxford)
• AIS Collab grew out of AIS Stockholm
• High Impact Medicine grew out of EA Germany’s ex-director connecting two people
• EA Netherlands seeded 10 EA university groups in 6 months
• The Simon Institute for Longterm Governance seems to have been downstream of EA Geneva/EA Switzerland
• EffiSciences was originally a group at Écoles Normales Supérieures (ENS) university. Since then, they have spun out additional projects like ML4Good and CeSIA that are independent.
• The Swiss Existential Risk Initiative (CHERI), now called Pivotal, was the brainchild of EA Switzerland’s ex-director
• MATS grew out of the Stanford Existential Risk Initiative
• ERA grew out of CERI
• WhiteBox Research was co-founded by people from EA Philippines

I guess it’s also worth mentioning that Bluedot grew out of EA Cambridge.

These examples came to mind with minimal research, but I wouldn’t be surprised if most fieldbuilding infrastructure we have came out from local initiatives that initially focused on “top of the funnel” outreach.

Building local infrastructure:

To start with an excerpt from Naomi’s post on “Examples of success from city and national groups”:

[...] many city and national groups engage in building local infrastructure, from launching translation projects, to hosting local co-working offices and providing operational support, for example employer of record services and fiscal sponsorship. These initiatives are often highly appreciated, as it allows local grantees (for example from OP, LTTF, SFF, etc.) to have formal employment and to focus on their work.

Two examples of impact stories:

EA Sweden has during the past 12 months supported 6 people through their employer of record service and fiscal sponsorship. One of the recipients is Stefan Schubert, a researcher in philosophy and psychology who studies EA topics:“Thanks to EA Sweden's operational support, I was able to secure grants for research on epistemic rationality.”

Another example, this time about a coworking office that is hosted by EA Netherlands.

Stan van Wingerden (co-founder of Timaeus, a new AI safety organisation) said:“The office played a major role in starting the collaboration with Jesse and Alexander [two other core team members]. I met Jesse Hoogland at EAG London, after which he came to Amsterdam. The coworking office made it a natural place to cowork together and meet up in the evenings to talk more about collaborating and founding an AI safety organisation.”

In AI Safety, I think it’s especially useful to start building up infrastructure in advance. As AI Safety gains more attention, it’s important to have a team of experienced fieldbuilders in key locations to help the incoming talent orient the field or to be able to organise high-quality conferences on short notice after major events, such as a warning shot.

Location-specific career advice and networking

I don’t think it’s necessary to expand too much on this point. If you have a competent fieldbuilder in a given location such as Brussels, they will be in a good position to give the lay of the land and connect you with the right people. Currently, there isn’t anyone whose job is to help new people enter the Brussels policy bubble.^[4] That’s not to say that the same person wouldn’t also benefit from talking to 80,000 Hours or Successif etc. It’s really hard to land a role in AI Safety, people need all the help they can get.

Similarly, while international conferences are valuable, attending local events can be especially useful in terms of networking and motivation to get more involved. Seeing other AIS people “in the flesh” reminds one that they are not alone on this daunting journey they have taken up.

Does Bluedot want you to send people to them?

Currently, the answer is yes, as they have begun offering intensive courses weekly, increasing their capacity to accept applicants. Still, Bluedot is most interested in specific talent profiles, primarily experienced professionals outside the current AIS bubble. If you have a young friend interested in AI Safety, they might just be fine with taking a local group’s course if they have the opportunity.

It won’t be run as professionally as Bluedot’s course, but they are more likely to give AI Safety the benefit of the doubt. Bluedot is well-suited to impress those who require a more formal experience.

A stronger version of this claim would be that Bluedot shouldn’t rely on the AIS community to fill in the slots for their courses and try to reach an audience that would otherwise not be exposed to AI safety. They do this already, but perhaps more can be done. It is going to cost a lot of money, and funders should pay for it.

How did we get here?

The “just send people to Bluedot” meme spreads wide and deep in the community. A similar dynamic occurs, to a lesser extent, when people consider starting other fieldbuilding initiatives. Why is that?

Lack of location-specific fieldbuilding infrastructure

First, it’s intriguing that the meme “just send people to CEA’s virtual programs” is almost nonexistent in the EA community. I found one person asking the following question on the EA forum: “What's the value of creating my own fellowship program when I can direct people to the virtual programs?”

People told him that it’s useful so that he can get a sense of “who are the people [they] can support best and how to support them” and that “facilitating yourself gets you to engage (critically) with a lot of EA material”. After this, the EA community called it a day.

The main reason for the difference between the two movements is likely that EA city and national groups already have a significant presence in a handful of locations. The CBG program has been around for years, and no one would direct Swedes to do an intro course with Virtual Programs instead of the one with EA Sweden^[5]. Virtual programs can focus on individuals without access to local or national groups for courses.

Bluedot had ~2,500 course participants in 2024, and their stretch goal for 2025 is 10,000 people^[6]. With ideal fieldbuilding infrastructure, Texas alone could produce this many graduates. There is currently no AI Safety Texas. There are almost no local/national AIS groups that are run in a professional capacity.

There's a saying: “if you have a hammer, everything looks like a nail.” You want to have a wide toolkit and choose between tools based on the problems you want to solve. However, if you have a hammer and see many nails, it’s fine to use it, particularly if you have yet to acquire other tools. There are more than enough nails for everyone.

To put it in EA-lingo terms, if you have a neglected problem and a tractable solution, it makes sense to use those solutions, even if others are doing something similar. The idea that we'll run out of skilled and motivated people interested in AI Safety is likely a misconception that contributes to the 'just send people to Bluedot' mindset.

Lack of marketing (and awareness of it as a possibility)

Historically, marketing has been really underutilized in the community and still is extremely so in the fieldbuilding space, which is why we founded Amplify^[7]. Because of this, people mistakenly believe that only a small number of individuals who are interested in joining the space, and a few organizations can handle everyone.

A more elitist take on this is that there might be a lot of people interested in joining, but it’s only worth supporting 1% of them. Can’t we do at least 2%?

Worse, some are actively discouraging otherwise good quality outreach^[8], believing that all capable individuals will discover AI Safety on their own. This likely stems from a desire to believe the EA/rationality clubs are inherently special, and those meant to join will find their way independently.^[9] This mindset is misguided and potentially quite harmful. You can get a lot of smart and motivated people through active outreach, even TikTok ads.

The median EA group (speculative)

I’ll focus on EA groups since AIS groups are relatively scarce, and most of them have started post-FTX.

Securing group support funding (e.g. for venue booking) has historically been relatively easy. Perhaps so was paid community building, particularly during the FTX era. The talent level required for securing a salary for local groups has been much lower than for international fieldbuilding initiatives. This in part left organising a city or national groups a low-status effort, perhaps even putting off people from taking up these responsibilities in the first place.

One way to improve on this would be expanding the CBG grant program (and creating an AIS equivalent), which has been limited for many years to a small number of countries with larger EA communities. Efforts could be made to scout for promising community builders and create a stronger pipeline for the field.

There are promising signs of establishing the right criteria for receiving funding, as CEA recently started a certification program for organisers to be able to apply for group support funding. What I’m worried about is that funders are now overcorrecting based on the picture of the past, and raising the bar too high for fieldbuilding grants.

Why you *should* just send people to Bluedot

I could be wrong about this. Before going into why, one thing that is worth clarifying is that most people most of the time should still just direct people to Bluedot, unless there is a group that can adequately^[10] serve the given person locally. This is a post to encourage you to engage in communitybuilding yourself.

With that aside, here are some reasonable objections as to why you shouldn’t take up communitybuilding yourself by starting a local or national group.

Personal fit

Alignment with broader career goals

As mentioned earlier, doing community building can help you get a lot of skills and provide career capital that can be useful for a wide range of roles in the future. Having said that, if you are confident that you want to go down a research path, you will benefit less from doing this kind of work compared to others.

You have better alternatives

If you can secure a prestigious job at an established organization, you should probably do that, though note that there are a lot of competent people applying to such roles. Perhaps you could have a higher counterfactual impact if you start your own initiative. However, note that guessing the impact of you passing a role to the runner-up is much more complicated than it sounds. Should you take the role, the 2nd place candidate won’t just sit around doing nothing, and neither will the 3rd, 4th…)

Having this problem would be a luxury in today’s AIS job market, as the competition is tight.

My current best guess is that highly entrepreneurial people should start their own projects instead of competing for low-absorbancy roles. Securing funding for an early-stage initiative is likely easier than landing a competitive job (I’m mainly deferring here to Non-Linear’s founder, Kat Woods, as I remember reading something like this from her). If you are in a position to do earning to give that could be a worthwhile path too.

We don’t have time to start building more infrastructure

Let’s practice a bit of rationalist lingo again. One might argue that on the margin, building local infrastructure has less counterfactual value due to short timelines. There is insufficient time to start pulling up new fieldbuilding organisations from scratch that try to recruit people into the movement. We should wing it with the infrastructure we already have and hope for the best.

This is an argument I’m sympathetic to, but it mostly falls short, as it assumes that fieldbuilders will sit around doing the same thing until the singularity arrives. This topic is worth its own post, but as we are getting closer and closer to AGI, I expect more fieldbuilders to pivot from recruitment to:

• communications
• help naturally incoming people orient
• help existing organisations to coordinate

They won’t be able to do these things (or get funding for them in the first place) if they don’t build some track record and a deep understanding of the lay of the land by doing a bit of the “tried and tested” activities groups do first. This is why Bluedot also supports their alumni in starting local initiatives.

Increasing returns with scale

With more participants in your program, accommodating them requires more effort, but increased scale can reduce the organizational time needed per attendee. Bluedot is clearly going to have the edge in this regard over smaller organisations, as they have one central place for a lot of people. However, that’s not to say that smaller local groups can’t benefit from centralisation. Two projects, AIS Collab and Scaling Altruism assist local and national groups with operational tasks for running courses, allowing them to focus on non-outsourceable activities^[11].

Generally speaking, as you increase the scale of your courses you should also expect to improve your cost-effectiveness per participant too, although it’s worth noting that AIS Collab has been volunteer-run, meaning that up to date they accommodated ~400 participants for a cost of a premium zoom account^[12]. This makes it more cost-effective than Bluedot but with a much smaller impact.

Conclusion

There are reasonable objections to what I’m proposing, and one should think deeply about whether they want to commit to building a local organisation. What I’m worried about is people spreading the “just send people to Bluedot” meme don’t think the various arguments through before discouraging others from taking initiative. I hope this post fosters a healthier and more balanced dialogue on this topic.

1. ^{^}

    “Adam from Bluedot told me the following: I have also spoken to people we've rejected and been excited about them. Many times this occurs where people aren't good at written applications, and this causes us to miss them. Also maybe a related point, but often we reject people multiple times and then accept them on say their 3rd/4th application after they have got better.”
   
   I think that’s super interesting, and I wonder what we could do to help people write better applications. If they have gotten it the 2nd or 3rd time, in my opining it would have been good if they make it for the first time so as that they can get involved as early as possible.

2. ^{^}

    The author posts a lot of cool AI Safety criticism, which you should check out!

3. ^{^}

    I forgot where I read this originally, but a good explanation for this is that it’s a very new field that has recently gotten, in relative terms, a lot more attention from within the EA community. This means that there are a limited number of people working in the area for 5+ years to mentor and manage people, and you would also want the same people to be doing direct work.

^{^}

 Talos is doing excellent work as a fieldbuilding organisation focused on Brussels. However, they are an initiative aiming to bring early-career talent to Brussels from various locations. Here I’m talking about a person to whom an experienced professional in Brussels could turn to after googling AI Safety.

^{^}

 Assuming both are run at the same time.

^{^}

 Note that as of 2025 March they are in the process of making a strategic pivot, so the number of course graduates might become less important to them.

^{^}

 Our initial plan was to help small local groups. However, we quickly realized that marketing in fieldbuilding is more neglected than we initially thought. In addition to numerous requests from small groups, several established organizations also contacted us. For context, GoodImpression’s and User-friendly does not specifically focus on fieldbuilding initiatives, although the former did help out Bluedot and Non-Trivial in the past. To the best of my knowledge, all orgs in the marketing space are capacity-constrained.

^{^}

I have come across two experienced fieldbuilders who have told me this before. See also this highly upvoted post on LessWrong that blames our failure to solve AI alignment on recruitment done by EA university clubs. (And this response buried in the comment section on why that’s not the case.)

^{^}

 That’s not to say there isn’t any truth to this claim. Local communitybuilders (as well as Bluedot) will tell you that the people who contact their group who join after stumbling across 80,000 Hours etc. online will be much more engaged than the median fellowship graduate who applied through active outreach. It’s also worth noting that these “natural followers” wish that someone exposed them to these ideas much earlier.

^{^}

In short:

-the local organiser is experienced enough to give good career advice

-courses are run well enough to not put off professionals

-rest of the community has a similar profile to the potential applicant

Basically if you are a student organiser new to AI Safety you shouldn't try to get someone over 30 to join your course and be the only professional in the cohort. That’s not to say you shouldn’t keep in touch, but send them to Bluedot to take the course.

^{^}

 If you are wondering why isn’t Bluedot doing this, see Adam’s comment here. If I had to choose between just having one of these orgs, I would go with Bluedot due to their advantage in scale. However as I outline in this post as well as here, there are good reasons to have both.

^{^}

 The 2025 Spring iteration has a bit of support from AIS Hungary’s paid staff, but my best guess is that this costs at most ~1000 USD.

Show all footnotes