Executive Summary

Crossposted from the Fieldbuilder’s Blog

My current Generator Project on Routing Mid-Senior (Ops) Talent to High Impact Roles in AIS took shape properly ~3 weeks ago and I want to sketch out some initial findings and hypotheses that I can continue to work on as I proceed.

The Theory of Change was to help AIS orgs scale and generally have more capacity/ function better to have more of an impact and generally mature the field by filling senior ops roles which have historically had a hard time getting filled. The specific mechanism was initially to talk to a bunch of promising senior ops people who are AI Safety curious and match them to roles they are best suited for to ease some of the difficulties in hiring.

However, as I go deeper into the project, I am inclined to think not only about the supply side of talent, but to focus more on the demand side in how we hire and what for.

Overall, I feel optimistic about the mid-senior bottleneck being pretty tractable to fix and like I have found some good low hanging fruit in this area. I would be keen for this work to extend past my efforts alone!

Hiring Managers should fill in my form with roles I may not be aware of so I can send you referrals.

Mid-Senior Talent should fill in my form to add themselves to my process so I can reach out to you.

People with Referrals should fill in my form to refer mid-senior people to my process so I can reach out to them.

Fieldbuilders should reach out if they want to chat about my work and have ideas for how they could help/ get involved.

What I’ve Noticed

Priors

Talking to many hiring managers and leaders across the field, a few priors about mid-senior talent seemed to creep up that inform the organisation’s strategy:

Mid-Senior Career Transitioners lack context
Mid-Senior Career Transitioners will not fit culturally within AI Safety
Mid-Senior Career Transitioners lack mission alignment
Mid-Senior Career Transitioners take too much time to pivot into the field
The field is special and its high context, close-knitness and mission alignment must be protected

Early Data on Senior Ops Hiring

I knew the field had struggled to hire mid-senior ops people for a while, some of the key results from a rough investigation I ran in Mid-July:

Only ⁷⁄₂₇ Senior Ops job searches across the AIS ecosystem got filled over the past 6 months (some of these did get filled in the subsequent month, but this result is still indicative of the struggle to fill these roles)
Many of these searches had been going for at least 4 months
Many of these searches involved many applicants (e.g. high hundreds)
The searches took lots of capacity away from orgs, e.g. in some extreme cases, they took a fte away from single digit teams for multiple months
Of the successful searches, ~half were closed rounds

From this early investigation, I began to suspect that senior ops people interested in AI Safety are just especially rare. I was guessing this is a compounding effect of AIS not having strong talent pipelines for ops/ generalists AND AIS leaning young with most talent efforts focused on students/ recent grads. My impression from talking to the hiring managers in these searches was that they just were not finding ~anyone who exceeded the bar either because of a lack of context, mission alignment or experience.

Null Hypothesis: we lack senior ops talent with context and mission alignment

Updates

Through the course of the project, there’s been lots of interesting data which has made me significantly update away from my hypothesis. Here are some of the most surprising findings:

1. Many senior ops people exist with (some) context AND alignment – I’ll call these people “unicorns”

Most of these people have done several BlueDot courses and regularly engage with AI Safety content and are eager to learn more. Several have impressed me with the depth of their threat modelling and understanding
There is some self selection for more senior people transitioning in to be especially mission aligned, likely moreso than for student/ early career folks since they have a much higher opportunity cost e.g. entering a new field.

2. Many of the “unicorns” fall into one of two buckets with very few of the people I’ve spoken to falling in the middle. Many also swap between buckets without consistency

Bucket 1: Rejected or never heard back at the first stage and left confused

For many of these people applying to roles listed as junior or mid, this is very confusing
I have had a handful of people I’ve spoken to tell me how belittling it is and how it makes them assume they are not a good fit for the field, with two cases even saying they considered quitting the transition altogether

Bucket 2: Rejected at the final stage having wasted lots of time without useful feedback

There seem to be lots of people qualified on paper who make it to the final stage of a process only to be told they were not a good cultural fit. Although they passed the bar for competence, they missed the elusive barrier since they didn’t have good insights into the culture of that org
Often these processes are very arduous, especially for people sacrificing a lot, e.g. taking a week off work to fly out for multiple days for an in person work test

3. These people are transitioning fast and going all in!

A prior I’ve heard as a crux for organisational strategy that senior people take much longer to pivot (12-18 months) than junior people (6-9 months) seems to be untrue in many of the cases I have found. The majority of the transitioners I’ve spoken to only discovered AI Safety in the past 6 months and are trying to pivot now, with many only coming across the concepts THIS summer
Some of the cases have been pretty drastic with people quitting comfortable jobs and moving across the world, even without secured next steps

4. These people are easy to miss in a hiring round that isn’t selecting for them

Many are left without ever hearing back
Many are given unreasonable/ inconsiderate deadlines to do tasks, e.g. can you do this writing task by end of day/ can you do this multi hour work test in the next few days, but then have to wait ages to hear anything back
Some have been given unclear/ confusing signals that are meant to be rationally useful, but come off as condescending e.g. percentage chance they get the role
Rejection emails can be insensitive/ a missed opportunity to be helpful, this is a great forum post about how to design better rejection emails

So the position I’ve been left in after talking to dozens of these quickly transitioning, committed “unicorns” is one of confusion. It seems strange to me that these people aren’t being hired. As much as I can individually try and route them, I want to figure out the systemic reasons for the disconnect of great talent applying to roles that purport to not find great talent.

Some suspected structural reasons for the disconnect:

1. There aren’t always jobs clearly suited to someone’s skillset

Some of the people I’ve spoken to complain about not having their relevant skill sets represented anywhere, e.g. there are not many jobs looking for product managers or community health professionals
Some of these people think the orgs actually need these roles without realising, e.g. many balls get dropped or a fte actually has 2 or 3 fte’s worth of responsibility rather than hiring separate people for the role
Role titles and job descriptions are sometimes just pretty badly written that act as rough suggestions for a role rather than clear outlines of what to expect, making it easy for senior people to self select out of rounds that don’t look well suited to them

2. Because of the field being young, we often don’t know what we’re looking for

Many of the fairly junior recruiters/ Hiring managers leading the searches have never done senior ops and don’t have a great sense of what competencies are needed
It’s easy to goodhart on jargon and someone matching the AI Safety mould, e.g. there are people with decades of LessWrong reading or AI Safety blogs who don’t naturally talk about “counterfactuals” because they are used to talking like real people, so their context is never elicited because in real world interviews people don’t talk about their reading habits, even though that’s exactly what lots of AIS interviewers would love to hear about
It also seems like many orgs want ops people (even those on the hard ops side) to have excellent AIS knowledge and fluency, even though this is not required for the role

3. I suspect hiring/ talent strategy anchors too much on historical norms and misses out because of it

While it may be true that mid-senior people in the space have taken longer to pivot historically, I suspect people thinking about talent in AI Safety anchor too much of their strategy around this datapoint.
I would push these people to consider whether increased legibility of AI Safety due to recent warning shots and political salience could unlock much faster transitions that we aren’t capitalising on because hiring signals lend themselves more to junior people who have been taught the jargon

4. Mid-Senior people don’t have an “in” or someone who will vouch for them.

Referrals go a long way and without a referral, people don’t take disconnected nodes seriously or the person doesn’t know about the closed hiring round

5. Some AI Safety people seem biased against senior folks because of bad experiences due to lack of mission alignment/ cultural fit that they extrapolate from

A few hiring managers I’ve spoken to are very wary of hiring senior folks who haven’t come from EA/ rat circles citing individual bad experiences with such people. This feels pretty uncharitable given how many bad students there are compared to the few good students who get hired
(Less sure on this point than others) While it is true a bad senior hire can be more expensive than a bad junior hire, we can get far more information about the track record of senior people from their experience. Often junior people are hired based on potential with an assumption they will improve over time, leading to greater uncertainty about whether the hire was “worth it”. However, for more senior people, they likely have less competencies to learn off the bat, meaning their potential in the role can be evaluated more easily.

6. Frustratingly, lots of these candidates do well/ are silver medalists in hiring rounds and are never passed on anywhere else.

Although some orgs are better at referrals, supporting transition grants and advising, this is not a priority for all, meaning that many great applicants are immediately back to square one with lots of lost information

I think these points indicate overarching structural reasons that the field as a whole struggles to absorb the mid-senior talent past that talent not existing. It extends far past one or two orgs being bad at hiring and points more at many well intentioned hiring managers who haven’t noticed or haven’t been able to address these structural issues.

What’s Next?

I plan to continue talking to more mid-senior ops people and trying to route them to roles. I also want to continue thinking about the wider structural reasons these people aren’t being absorbed and doing whatever I can to try and fix these barriers to impact.

To try and speed up hiring excellent people, I am looking in a few different directions:

Continued coordination through my ongoing routing efforts
Investing in and advising new talent pipelines to round out senior people’s profiles to be more legible to AIS hiring rounds, e.g. GenStream, BlueDot’s Context Programme, etc.
Raising awareness of the field’s current blockers through writing like this
Focusing on some more targeted fixes, e.g. making conditional commitments for roles that don’t currently exist which harness senior people’s expertise
Thinking about how mid-senior transitioners can receive more support through more AI Safety hubs, online communities and more visibility about their journeys

I’m excited to see how much my referral efforts result in counterfactual talent being routed into the field as hiring rounds move along and more datapoints are added.

How can you help:

Hiring Managers should fill in my form with roles I may not be aware of so I can send you referrals.

Mid-Senior Talent should fill in my form to add themselves to my process so I can reach out to you.

People with Referrals should fill in my form to refer mid-senior people to my process so I can reach out to them.

Fieldbuilders should reach out if they want to chat about my work and have ideas for how they could help/ get involved.

Appendix: Stats

People

So far, I have spoken to ~100 people, made up of:

~25 Generator Residents
~25 EAGxBerkeley
~20 from my Form
~5 GenStream Fellows
~5 Leaf/Non-Trivial Alum
~5 Serendipitous Meetings
~5 Personal Network
~5 Referrals

Of these, the seniority is skewed slightly younger, mainly because of the Generator Residency and my role in Leaf/ Non-Trivial:

~20% students
~15% Junior (0-3yrs)
~25% Mid-Career (3-10yrs)
~40% Senior (>10yrs)

I have ~60 more people I am working through who have reached out/ been referred and potentially more than 100 more through upcoming conferences and workshops (e.g. EAG, Longitude, Lateral, PRAXIS, etc.)

I also expect many more people to send referrals since there is a ballooning effect of people I route sending their friends/ personal networks my way, plus having other headhunting efforts from orgs like 80K, C*, SuccessIf, SteadRise, etc. sending people my way.

While it may be tempting to point to the survivorship bias in my sample, e.g. “this sample is unrepresentative of most mid-senior ops people because they’re self selected”, I would argue that the magnitude of this group alone suggests there is work to be done. Even if it is only a very small fraction of the senior-ops people in the world who could be great for AI Safety, that is still a large absolute number of people who could be absorbed, especially when the field itself is only in the low thousands of ftes.

Roles

So far, I have tracked ~300 roles across the field, currently ~200 are live with only ~110 of those being fully accessible to the public.

I have sent ~20 batches of referrals to different roles so far, totalling ~60 individual referrals, and received positive feedback on their usefulness and calibration from several hiring managers.

While few of these have materialised in offers yet (the earliest was 2 weeks ago), candidates counterfactual to me have now received offers or are medalists in rounds from Coefficient Giving, Eleos AI, BlueDot and OAISI and I expect many more successful cases in the coming weeks. I have also helped a handful transition out of unhappy jobs in AIS from orgs I won’t mention here, and got candidates into GCP’s upcoming Bio Workshop, BlueDot’s Context Week and GenStream.

Good Senior People Exist and We’re Wasting Them