An AI Safety Career Switch Postmortem

Introduction

When I worked as a dev, whenever some software system blew up we'd do a short presentation and discussion of what went wrong and how to improve. This was a good system, and I've decided to apply it to my recent attempted career switch (unsuccessful, alas). I am not trying to convince anyone of what do to (or not)—if you're looking for AI career switch advice, try and find a good survey of folks who've attempted it and base your decision on a meaningful sample size. The main reason for sharing this publically is, as in the software system failure scenario, to encourage input on what went wrong and on how to do better moving forwards.

Background

I studied Philosophy at Oxford, graduated with First Class Honours in 2014, and after a few years of trying various things and some 80K career coaching, I landed in software development in 2017. I did this full-time for about five years.

I was first properly spooked by AI when GPT-2 released in 2019. I wasn't terrible at Math, but I'd dropped it as soon as I was allowed to at school, and I started trying to reverse this decision. I taught myself stats and calculus on a very part-time basis while working (if you want to do this, I highly recommend MIT OpenCourseWare). I was in Montreal, so I made friends with a few folks at Mila and went to some events.

By early 2023 I was extremely burnt out from coding and extremely fed up with feeling helpless w.r.t. AI. I had always been much stronger in humanities-related fields anyway, and I figured most of the people in that space probably didn't know what backpropagation was. I had also formed the impression that ML researchers and policymakers were frequently talking past each other without realizing it. Improving this communication seemed like a big, dumb win (fine, "impactful, neglected, and tractable"), and I decided to try and do it.

I am based in Canada, and after extensive LinkedIn snooping it seemed like a lot of the folks doing interesting policy work had been through a Master's program called NPSIA. I enrolled and managed to get funded via a few scholarships, partly due to an AI-related research proposal. Faculty members and policymakers I met through the program indeed did not know what backpropagation was. Neither did students—despite not being particularly strong in Math, I was one of the strongest in my cohort on anything quantitative, and I was asked to help teach statistics in my second year.

It was extremely nice to feel good at something again—as a developer I'd felt adequate but not much more. I learnt some interesting stuff about deterrence and collective action theory, and I consumed a vast amount of AI policy literature. I eventually wrote a thesis called "AI Crisis Preparation for States," which I don't think was very good, but I was at least pleased with a few sections (which are available on my personal site). Because nobody in the NPSIA faculty was AI-literate, my supervisor insisted I find someone external to help him assess my work. I found Ben Bucknall through Mila, who was since a Chapter Lead on the 2026 International AI Safety Report (IAISR). While completing my Master's, I worked through Fast AI's curriculum to ensure my technical fundamentals were solid.

I also talked the ear off anyone in government who would listen to me about AI—NPSIA was extremely well-connected and I met quite a few public servants. In particular, I talked pretty extensively with an Associate Deputy Minister (high-ranking public servant) at ISED who was looking for safety-related research. Since I had just read a lot of this, I wrote him summaries of it by email every couple of weeks—adapted versions of these are also available on my personal site. I got a good government summer internship to round off my Master's program, and I graduated last October.

After this, I applied to every government job that sounded related to tech policy—I even had access to internal listings due to my summer placement. I also looked up every single company affiliated with the IAISR, and I applied wherever had job postings. Unfortunately, this is where the good vibes end, because I didn't get a single interview. The closest I got was encouragement from some IAISR authors on LinkedIn and a suggestion from SaferAI that I apply to a related position (which I did, also unsuccessfully).

It's now been a year since the end of my government placement. If I'd gotten anywhere with a single application process, things would likely be different, but I've struggled to stay motivated to read new policy-relevant literature, and I've fallen behind. Also, a year is a long time to put your life on hold, and I've reached a personal breaking point with it. Even if persisting is a good idea, I don't think I have the energy to do so.

Some thoughts on what went wrong

I just didn't find that many jobs. Perhaps this should have been less surprising to me—I am in Canada which is not exactly San Fransisco, and even before Trump's re-election, I could have expected public spending cuts around when I graduated due to a likely incoming Conservative government. But I saw a clear need, and I assumed there'd be positions, if not in some slow-moving government department then at least in think tanks. With hindsight, this looks like a pretty embarrassing assumption to have made.

I have very little hard policy experience. The job SaferAI suggested I apply to had this question in its application: "What is the most impactful area of AI governance that you have worked on?" I didn't have a good answer. I think I have a lot of relevant experience, but I can certainly see how "has actually worked in AI governance" or even just "has worked in government" is a safer bet than me.

My approximate theory of change was "help ML researchers and policymakers not to talk past each other." I should probably have thought more carefully about how hard it might be to convince ML researchers and/or policymakers that they were doing this, especially since it seems a clear prerequisite to me getting hired. I had plans for how to do this in an interview situation, by drawing on concrete examples from government reports and hearings. But I needed a better plan for how to actually get into the interview.

All that said, whenever I've talked to someone who I think should know, whether a PhD student at Mila or a government official, they've basically said "wow, you sound super-qualified." In spite of the above points, I still really struggle to square this feedback with my abject job application performance.

Some things I still stand by

Interdisciplinarity is valuable. I think most people stick to the fields they're best at, and having spent 5 years not doing that, I can see why. Not doing that is hard. You are surrounded by people with different training and different baseline assumptions to you who constantly misunderstand you. Your boss's idea of a sensible career trajectory for you makes no sense, because your starting skillset is completely different to the one they're used to all their other employees having. Your weaknesses are constantly magnified and your strengths are constantly undervalued. And ultimately, you're just worse than you have been and could be somewhere else relative to everyone around you. But the flipside of all this if you do it anyway, you'll be one of the few people bringing your particular skillset and background into that space. This seems like a pretty good way to find neglected opportunities.

ML researchers do not understand policymakers, and policymakers do not understand ML researchers. Fortunately we have a tried and tested solution for groups of people who do not understand each other: translators. Translators typically translate from a language they know well, but which is not their mother tongue, into the language that is their mother tongue. This pattern can be easily applied to AI policy. Whenever a researcher talks to government, they should bring at least one other researcher who's spent meaningful time engaging with government policy, and who can help them see what policymakers care about. Whenever a public official talks to a researcher, they should bring at least one policy analyst who's spent meaningful time engaging with technical ML research and will understand what the researcher is trying to communicate. If either the researcher or the public official does not work with someone who matches this description, they should hire someone who does.

Concluding thoughts

On a personal level, I don't think I will regret doing this (although it's certainly possible I change my mind). I had forgotten what it was like to feel good at something, and I'm grateful that this process reminded me. The best way to feel better about an upsetting situation is to try and change it, which I did. And for some folks to help with AI, it makes sense that a lot of people need to try and fail. Thank you to everyone reading this who is trying to help.