The Missing Layers in AI Safety: Experienced Builders and Meta Solutions

🤖AI Transparency Statement: This was fully written by me, and has not gone through any review with AI.

Introduction

One of the mantras when I worked at National Instruments (NI) was that if the role you wanted didn’t exist to create it. I did exactly that 3 times in a row and am now applying that lens to AI safety.

As I was building my own mental map of AI safety and how I could help, I started observing repeated problems and challenges around organizations struggling to scale and overall keep up with frontier labs. I went to EAGxBerkeley 2026 and ultimately validated them, solidifying my hypothesis of how I could help, but also observing a large gap in AI safety’s appetite to place me in a role that could achieve that, and the appetite to build those types of solutions.

I wasn’t the only one, we found about 8 people sharing the same viewpoint and frustrations, and created a working group afterwards to help strengthen our case, ideally leading to creating the roles and projects that would allow us to fully leverage our abilities and make the largest impact on AI safety.

This project initially creates various writing artifacts to draw attention to this need, ideally leading to additional conversations, more support, and finally to hiring roles and building projects that fit this profile.

Project Stage

Last Updated: September 14th, 2026

Drafting: We’ve formed a working group, met, aligned on some objectives and initial outputs, and have those nearly to a state that are ready for peer review before posting publicly.

I’ll work to keep this page up to date as the project develops, but wanted to create it as a stub to show some of the projects I’m thinking about for Fall 2026.

Overall Project Idea and Hypothesis

As a quick recap I joined CEA Career Bootcamp in June 2026 where their material + 80,000 Hours AI Advisor steered me towards technical AI safety with a focus on evals due to my background in online A/B experimentation, but likely not as a researcher or engineer, and instead as a “generalist”. I completed the BlueDot Impact AGI Strategy Course, and was thinking something along the lines of research operations. I then completed BlueDot Impact Technical AI Safety Course, had started to observe how fragmented the space was, and wanted to build community while continuing courses and projects to further map the space. Simultaneously, I’ve been reading LessWrong, EA Forums, reading job descriptions, and more to gain context. Each thing I consumed seemed to be confirming my core hypotheses that I left EAGxBerkeley confident in after many discussions:

There was an unrecognized gap between generalist operational roles and researcher/engineer roles. I wasn’t finding anything that I truly fit into that could fully leverage my abilities.
Organizations and funders were very focused on producing more research, while simultaneously struggling to scale up. I wasn’t finding many infrastructure roles or ones that would help an organization define and prioritize their work.
There was a lack of focus on creating meta- solutions that scale across organizations and throughout AI safety and help standardize, centralize and democratize efforts.

So, this project set out to address what we saw as an unaddressed gap in AI safety’s approach to roles and the work they could help build.

Project Plan

In late September or early October 2026, we will share first with peers, then on LessWrong or EA Forums:

Our Main Piece: Arguing why AI safety needs to fieldbuild and hire for our role profile, and the types of projects we would help build – I’m co-authoring on this
Supporting Pieces:
- Generalists can Gain Context Quickly: One fear of hiring outsiders seems to be that we can’t get up to speed quickly and add value. Generalists in industry are used to this, so we hope to assuage those fears by sharing more on this. – I’m co-authoring this.
- Sample Builder Role Profile and Job Description: If the job you want doesn’t exist, create it! In this piece, we’ll show AI Safety organizations what we’d like to see when looking at job postings. – I’m the primary author on this.
- Taking More Risks in Hiring and Funding Orgs – AI safety has more capital than it can deploy. There’s a lot of people trying to break in. Makes the case for taking more risks to grow more quickly. – I’m the primary author on this.
- (TBD if this one moves forward) An example fieldbuilding program: built specifically for this role profile and their constraints around income, timing, location, etc. – I’m just an input on this
- (TBD if this one moves forward) Study of most effective interventions in AI Safety: Maybe more research really is the most effective approach, and we’re wrong on our assessment. This searches for those answers.

What happens or not after this depends on the reception, our ability to land conversations with influential AI safety leaders, and potentially other efforts, like applying for Coefficient Giving’s Project Tailwinds initiatives to create some of what we’re recommending.

My Contribution and What I Want to Learn

I’m authoring and co-authoring several of the efforts, actively participating in our working group, and peer reviewing other’s work. I’ll be doing my best to keep pushing this forward even if momentum looks like it’s going to stall.

I’d be happy to learn where our assumptions are wrong as well as get deeper into the jobs to be done and pains that organizations and funding organizations are experiencing so that I can better describe exactly what I can help build and why.

Outputs

See #project-plan . Nothing has yet been output for public consumption. Once ready, I will update this page with the links.

Expected Outcomes

Near-term: We hope that these both lead to direct discussions with leads in the field, as well as add to the ongoing discourse and discussion on LessWrong and EA Forums.
Medium-Term: This leads to new fieldbuilding efforts to support our role profile and the types of projects that we’d like to work on. This in turn opens up positions matching our role profile in founding organizations and working within existing organizations to build meta level infrastructure and systems.
Long-Term: Structured like the organizations and work that it’s up against, AI safety is able to scale its work to catch up to frontier models and ensure the worst potential harms of AI are prevented.

Appendix

How this Project Work is Unique

If you’ve been around AI safety in 2026, you’ve likely seen the call for generalists, Kairos getting funding to support that, then the call for mid-career on-ramps, and the overall need to scale organizations. There’s more too!

Below will be defined in our writing pieces, but I’ll summarize it below in my own words.

Our goal is to differentiate our roles, what we can contribute and what we’d like to build in at least 3 ways:

Additional Experience: We’ve seen mid-career, senior, etc in AI Safety referring to people with 5-10 years of experience. We’re targeting our role profile to those with at least 10 years of experience which we think brings even more important context and pattern matching.
Generalist Builders, not just Operations: There’s a lot of important operational work to be done in AI safety, but we think there is a gap between that work and the very technical work of researchers and engineers. We see our key differentiator being in how we build something, rather than purely executing. Breaking down tough problems and building for outcomes, not just outputs. Being strategic with limited resources and getting ruthless on prioritization.
Thinking Meta: While there is and has been a lot of great research created, we see a gap in the application of that speed up the impact of AI safety. We’re thinking beyond a research theme or organization to how to stitch all of this siloed work back together.