Paul Davidson
Product Manager | Program Manager
I'm a generalist with a technical lean trying to break into AI Safety as a part of pivoting the 2nd half of my career into more meaningful work. I spent 18 years as an engineer, program and product leader, most recently building the evaluation platform Expedia Group used to decide whether product and ML changes were safe enough to ship. I'm a high-agency executor with years of experience turning ambiguous, high-stakes problems into requirements and strategy, then executing through implementation.
I started this pivot journey off with a serious focus on climate and clean energy, but as I've worked through CEA Career Bootcamp and BlueDot courses, I've been increasingly convinced that the biggest need and impact I can make right now is in AI Safety. Plan B is likely a software product role for a highly tractable solution in clean energy.
Within AI safety, I currently think the cleanest transfer is into evals. In my last role, I owned the internal the experimentation platform and also program at Expedia Group where I built the intake process, the safety guardrails, and the org-wide trust in a decision pipeline from scratch. We supported running 500+ concurrent experiments across all brands, devices and functions where even a small degradation could cost millions of dollars.
I'm currently (August 2026) closing my skills gap through BlueDot's Technical AI Safety course and ongoing self-study. I'm starting to network and write, and soon plan to be working on projects to prove myself within the ecosystem.
> But this works only if the specialist community has not collectively overlooked a foundational issue. If the field has an incomplete model of the object it is trying to make safe, then adding more capable people may accelerate useful work, but it may also accelerate work within an incomplete framing.
Thanks! I largely agree, and would add 2 points onto what you you said here:
1. In my opinion, the goal of adding generalists are to take work off specialists plates that may not be delivering the most value for their capabilities, may not be their largest strengths, etc. That in turn should free up more of their time focus on the toughest questions, like: Are we on the right path?
2. The power of generalists are the diverse careers and passions (outside of work too) they carry. As we look at new problems, we stitch all of that history and context in. Sometimes that generates novel ideas or questions that haven't been asked. In short, bringing in generalists brings diversity and expansion of potential ideas, as well as how to execute work.
A few notes:
1. It looks like https://www.lateralworkshop.org/ might be a partial answer and experiment to your feedback by offering a long weekend of intensive instead.
2. I do think the Bay Area concentration of US opportunities (learning/program, roles, orgs) is a bit of a risk to missing out on great talent to work on AI Safety.
3. The transition into all high-impact work feels like what you described: months of work to credential, prove and network. Typically, mostly on your own time and money. That's not feasible for many for a variety of reasons, meaning more talent is leaking out back into lower impact work so that they can sustain themselves. This might be even more acute for the generalists this post is calling for. Mid- and late-career are going to be less flexible in their situation and have more responsibilities than someone early career.
> What I want to say is this: AI safety is now being discussed, pursued, and championed by many thoughtful people. That's a good thing. But wholehearted commitment to a cause doesn't guarantee sufficient understanding of it.
I don't fully agree with this or that every role needs the same level of depth to be successful.
As generalist with a strong technical background and track record, I believe I can quickly gain a deep-enough understanding of AI Safety and then go further into a niche or specialty. I am already on the road to this a little over a month in through coursework and reading, plan to advance it through projects, and then could get much deeper once inside an org. I already have proof of doing this across my career. The types of roles a technical generalist would support (research manager, technical program or product manager, research operations, etc) don't require maximum depth of knowledge. We're used to first going broad, but shallow, then to a medium or deep depth, as needed. A lot of the work is more about studying what needs to get done, the constraints, the current challenges, then mapping out a plan to move forward. You still lean on the experts when they're needed, but otherwise are taking tasks off their plate that they may be less interested or less adept in. Further, we can usually help with light coding tasks, testing, and other simpler technical work letting the experts focus on the toughest problems.
I think less depth is needed in non-technical roles. Fieldbuilding and the like need to understand a very broad surface area, but not at a ton of depth, as they'll lean on experts when needed. Some aspects of communications, policy and governance also lean heavily on the "art' and "people" side of things to get things done, and when it comes time to write the details, they can call on the experts too.
Yes, there might be people that just aren't well suited for AI Safety overall even if they want to be in the field. I don't know what those reasons might be, or how to exactly assess them. But, I think there's a large group of qualified people out there that are very qualified to contribute and accelerate AI safety work.