🤖AI Transparency Statement: This article was fully written by me, then went through basic grammar checks and a very light editing feedback loop on clarity with Claude Opus 5.

Epistemic Status: One thing I’ve reflected on recently is that I’m not the same AI safety person I was even two weeks ago. My mapping of the space is evolving at a rapid pace. Each iteration is growing my confidence in my assessment and plans, which I’d put at fairly confident right now. This follows my BlueDot Technical AI Safety – Course Reflection as my next major touchpoint in growing my understanding of the AI Safety (AIS) ecosystem, focused on where I can add value.

Overview

I joined a couple hundred other effective altruists, many focused on AI Safety, at EAGxBerkeley 2026 at the LightHaven venue for a weekend of networking and discussions. I attended the opportunity fair, a few presentations, 17 scheduled 1-1s and then had some great conversations in between all of those. Overall, I found it very energizing, and am leaving optimistic that although the path for my role profile still remains a bit murky that there’s momentum gaining that can help me land a role and start contributing to AI safety soon.

My EAGx Experience

I was told to max out my 1-1s, and I did my best to do that. I met with three research managers, and others in operations roles at independent orgs to learn more about the work they do and where they see things going. I found a cohort of about five other mid-senior professionals with product and other experience creating strategy and executing, and we were all arriving at the same conclusions, which I’ll distill below. I also met with four fieldbuilders, two were more focused on different role profiles, but two working to better support our role profile who were really starting to see the bigger picture of what we’re pitching and why it’s hard to break in. I probed one member of technical staff for almost our entire 30 minutes on the details of his job day-to-day, increasing my sample size in that category now to two. Everyone was very welcoming, excited to discuss, and ready to challenge ideas or open up new points of view.

Common Themes I Found

• Rapidly scaling organizations feeling similar pains as they evolve their maturity. Young leaders who have expertise in research, background in academia, but less on driving the operations and strategy of an organization. Org structures that are very flat even as their size has grown.
• Many categories are described as “jagged,” meaning that orgs, teams, or their outputs might be mature, and others not so much.
• Lack of scientific rigor or work that might be considered more qualitative.
  • Possibly being driven by lack of standardized infrastructure, tooling, processes, and more. See my 1-pager on an automated evals platform here.
• A high bar to getting into AI safety, want to know people are really committed and care. Importance of connections, referrals, programs, etc. But also, being willing to put yourself out there and bias for action. Try cold outreach, even when it’s uncomfortable.
• There aren’t many people thinking about the broader or meta levels of AI safety and research, and limited founding and fieldbuilding support for that work right now.

My Views on AI Safety’s Top Needs

What’s Changed After this Event

Instead of a rewrite, most of all, this is an expansion and evolution of the working hypothesis I started after my Technical AI Safety Course ~two weeks ago when I started to see how fragmented the space was and lacking connective tissue. I’m more convinced now that there is a strong need for my role profile and broader infrastructure + tooling ideas, but the demand or awareness is still in the very early stages and may only be felt by the leaders rather than the ICs. Especially ideas of new orgs and tooling to work cross-organizationally, are maybe too early to get much support. For now, I’m going to de-prioritize a second idea I had on a Wikipedia/Hugging Face type space to centralize information as I don’t think AIS is ready for that yet. I learned that many view the technical policy side as more neglected than I realized, which raised it slightly on my radar.

My Evolved Working Hypothesis on Technical AI Safety’s Largest Needs

Independent technical AI safety organizations and their work are very fragmented and at an early stage of maturity while scaling and evolving rapidly. This creates pain points of feeling like they aren’t working as fast as they could, nor delivering as much impact, and are falling behind the development coming from frontier labs as well as new applications like multi-agent interactions.

Different labs, and even teams within an org, run different processes and tools, present their findings differently, as well as their policies. Making a decision or assessment on whether a model is safe for release or exhibiting dangerous capabilities continues to fall back on human judgment, typically on researchers who are bandwidth constrained. Policymakers struggle to understand what’s happening in technical AI safety, and understand how to translate it into policy.

AI safety is reaching an inflection point now where it needs to start standardizing and democratizing its work. There needs to be more connective tissue and support between the frontier labs, independent orgs and researchers, and the government. I see this need on at least two levels, although they do relate:

1. The research writeups, data, code, methods, prompts, results, etc. need to be unified into a single experience for anyone to review, search, and perform meta-analysis on to determine the state of the space and known gaps whether that’s research, standardized practices, current model capabilities, latest benchmarks and evals, etc.
2. AI safety needs centralized bodies and cross-organizational work to bring everyone together, align on standards and best practices, develop them, and work on having them broadly adopted and implemented correctly. This isn’t a regulatory or licensing body. It’s an opt-in that everyone contributes to make the space safer faster through combining their strengths and generating standardized work that can be highly trusted.

My Evolved Personal Theory of Change

I can support and lead organizations to build their operations, overall strategy, tooling, and infrastructure. One path is through founding to create outputs that can be used across organizations. But the faster and potentially more viable path is working with an organization to help them hire, organize their organization, prioritize their research, tooling, and infrastructure projects. This will take pressure off founder researchers and other researchers to focus on advancing their research and using their core strengths. The rest will benefit from clearer operating procedures and tooling that allow them to work autonomously for longer, and more confidently in their own goals. This will allow independent AI safety work to catch up to frontier lab model development because they will operate faster, while also producing higher-quality and higher-impact outputs.

My New Potential Role Ranking

Prior to this event, I was having trouble articulating this, but it’s coming into focus now. These are ranked top-to-bottom by how much impact I think I could make on AIS. See my thoughts on how I plan to actually apply this in role work below.

1. Founding for Infrastructure and Tooling – To build broader tooling and infrastructure needed to increase the volume and rate of impact of all AI safety organizations.
2. Build Infrastructure and Tooling within an Org – Performing similar work to above, but within an organization (Meridian Labs, METR, AE Studio, etc.) where it might still be shared across organizations.
3. Chief of Staff or Similar – Within an AI Safety org to help them strategize then prioritize the most impactful work (including infrastructure and tooling), and ensure they are operating efficiently and scaling as fast as they’d like.
4. Technical AI Policy Writer – To help translate technical work into understandable outputs for policymakers and their bills. Admittedly, this is the one I know the least about, but after several conversations think I may be uniquely positioned with my technical background and past experience to make a larger impact than some, and from what I understand this space is very understaffed currently.
5. Program Manager, Research Manager, Technical Operations, etc. – To first assist with the day-to-day, but with a keen eye on using it as a stepping stone to expanding into roles 2 or 3.

What’s Next for Me

Role Work

Focusing on where I could make impact the fastest, I will primarily focus on joining an existing org to help them build, or operate more strategically and efficiently. (#2 and #3 above) above. In this path I can take advantage of structure already in place, and those already familiar with operating in this ecosystem. I will continue to probe founding in parallel but with a lower focus. I’ll explore operations (#5) opportunities that have a clear path up front to expand the role and its impact. I won’t actively seek out technical policy writing roles (#4) just yet to keep me focused, but I’d definitely be open to talking about them if approached and can be convinced they are an even bigger leverage of my abilities.

Effort Mix and Concrete Steps

How I’ll roughly be spending my time:

• Applying to roles/programs/projects – Around 25% of my time. I was planning for this to be much lower, but was encouraged at EAGx to bias for action, including reaching out cold or where I might not initially appear to be a good fit.
  • Applying to roles that have a clear fit, and especially if there is a connection.
  • Applying for BlueDot Impact’s Career Transition Grant – not only financial support, but community support, and acts as a vetting step into other orgs.
  • Reviewing a couple of programs aimed at founders like BlueDot Incubator Week.
• Learning and gaining context – Another 25% of my time, with a focus on some basic Python work and consuming more news, LessWrong, etc.
• Projects – 50% of my time for the Fall.
  • SPAR Fall 2026 – Hopefully lands a project with a mentor and a visible output at the end
  • Generating a 1-pager on my pitch for a standardized, centralized and democratized automated evals platform to accelerate scaled oversight impact, and take pressure off researchers running and analyzing evals. (WRITING COMING SOON)
  • Creating or being an active contributor to a newly formed community of people with my role profile working to break into technical AIS.
    • Generating a 1-pager role description and supporting industry role descriptions to support fieldbuilding connections to our role type.
  • Collaborating with an attendee I met on something in evals or automating evals, including potentially trying out a lightweight (not very scientific) project on a multiple-judges idea that I don’t think I’ve seen written up yet.

Alternative plan: If I decide to go more of a founding route, I will likely increase learning to 50-60% to gain context on that. I’ll be collecting more information in September to refine this.