I used Claude to assist with the translation; all arguments were reviewed and revised by me.
I came into AI safety mid-career, after fifteen years of editorial work in Russian-language media – the kind of background that, on paper, doesn’t look like an AI safety resume at all. Russian is my first language; English is one I work in, not one I grew up reasoning in publicly. I’m writing this from inside the US now, a few years into the move – close enough to the hubs this post is about to see the difference between how the pathway looks from outside and how it works once you are standing in it. Trying to find a way in, I kept running into a pattern: the on-ramps weren’t closed to me exactly, but they all seemed designed around a different person.
That person is roughly: young, academic, Anglophone, and flexible enough to relocate, take unpaid opportunities, attend fellowships, and spend concentrated time building field-specific credentials. Many existing programs work well for that profile, and that’s good. But there’s another group that seems structurally under-served: mid-career people from non-Anglophone countries who have relevant skills, but did not grow up inside the US/UK academic, EA, or LessWrong-adjacent pipeline.
This is a diagnosis, not a complaint, and it’s aimed at people designing AI safety and governance pathways. My central claim is that the barriers for this group aren’t merely additive – they’re interlocking, and they differ sharply in how tractable they are. That second part matters most for program design, so I’ll keep returning to it.
A common pathway into AI safety looks roughly like this: learn the basic ideas; attend courses or events; produce visible work; network; volunteer or join a project; eventually become legible enough for paid work or funding.
For a student, this is demanding but plausible. They may already be in an academic environment, have geographic flexibility, carry a lower opportunity cost, and sit closer to the social style of many EA and AI safety spaces.
For a mid-career person from a non-Anglophone background, the same path has a different cost structure. They aren’t missing just one thing. They may face financial constraints, credential mismatch, weak network access, unfamiliar field norms, language-and-culture translation costs, and limited time for portfolio-building all at once. Removing one barrier doesn’t free the person, because the others stay binding. That’s why single-barrier programs often miss this group – not because the programs are bad, but because they’re solving a differently shaped problem. The current pathway asks these entrants to absorb too many transition costs privately.
Mid-career entrants usually already have careers, families, rent or mortgages, dependents, or immigration constraints. They can’t treat a career transition like an extended student project. Unpaid fellowships, speculative volunteering, relocation-heavy opportunities, and long stretches of "just build career capital" are much harder to sustain when you’re already responsible for an adult life.
So portfolio-building has to happen in parallel with paid work – keeping a job, taking freelance work, juggling income streams while also trying to study, attend events, build, write, network, and apply. The result isn’t only slower progress. It systematically crowds out the things that make sustained work possible: rest, health, relationships, recovery.
This is rarely counted as a fieldbuilding cost, but it should be. A pathway that is technically open but practically requires months or years of unpaid parallel labor selects for people who can afford that burden – not necessarily for those who could contribute most.
Credential mismatch is real, but it’s one of the more tractable barriers. A foreign degree may be discounted; a strong career elsewhere may not map onto local hiring signals; experience that was senior in one context can read as ambiguous in another. This hits people coming from journalism, communications, operations, law, policy, education, product, security, or public administration especially hard – serious skills that don’t appear in the expected format.
The good news is that demonstrated skill can partly compensate. A visible portfolio, a concrete project, a clear case study, a well-scoped contribution – these make a person more legible. This is one of the few barriers an individual can meaningfully work around.
But that creates an asymmetry worth holding onto: credential mismatch can be reduced by individual effort. Network access and financial constraints often cannot.
Network access is the hardest barrier, because it’s the least individually solvable. Entry often runs through informal networks: who gives you feedback, who invites you to a Slack or Discord, who suggests a project, who explains which organizations are credible, who says "you should talk to this person." Near the hubs, this feels natural. From outside, it’s opaque.
A concrete illustration is the role of flagship events. Many of the field’s most important gatherings are held in the US or UK. Even when a remote option exists, it usually buys access to the talks, not to the part that converts into opportunities: hallway conversations, demo sessions, informal introductions, weak-tie relationships. For someone mid-career outside those hubs, the cost isn’t just the ticket — it’s travel, visas, time away from paid work, and a price that means something very different once converted into local terms. So "there’s a livestream" can be true but incomplete. The recorded talks are often the most accessible part anyway; the networking layer, the part with real career value, stays bound to a physical room.
Some events offer travel grants or diversity stipends, which genuinely helps — but these are limited, competitive, and often easier for students or early-career participants to use than for mid-career people changing fields. The gap isn’t mainly access to information. It’s access to the relationship-driven layer where opportunities circulate.
There’s also a translation problem that isn’t only linguistic. AI safety material often assumes familiarity with EA culture, LessWrong norms, US/UK career signaling, and a particular style of public reasoning: how uncertainty is expressed, how disagreement is framed, how criticism is handled, how people signal seriousness, which institutions are trusted, which concepts count as basic, which career paths count as legitimate.
A version of this problem shows up even in the way people talk about AI tools. There is a line that gets repeated in AI circles – Andrej Karpathy’s "the hottest new programming language is English," later echoed in similar ways by prominent AI and tech executives. It is usually presented as a story of democratization: you no longer need formal technical training; you just need to describe clearly what you want.
But that framing quietly assumes that the describing happens in English, and in a particular register of English. If English becomes the interface to the field’s tools and conversations, then a non-native speaker is not just learning the ideas. They are operating through a second language in which unfamiliar phrasing, calques, or a different argumentative style can be read not as an accent, but as imprecise thinking.
The advice to "just describe what you want clearly" hides a cost that is distributed very unevenly.
There is a subtler version of the same cost. Fields reward a particular texture of reasoning – in many EA and AI safety spaces, that means making reasoning explicit, stating uncertainty, exposing assumptions, separating claims by confidence level, and producing artifacts others can inspect. These norms have real value. But they are easy to mistake for neutral rationality when they are also a local dialect of it. A person trained in a different intellectual tradition – where, say, authority is signaled through synthesis rather than through visible hedging – may not be read as reasoning differently. They may be read as reasoning worse. The penalty falls on the style, but it is scored as if it fell on the thinking.
For an outsider, then, the difficulty isn’t just "learning AI safety." It’s learning the implicit social operating system around it. This is harder for mid-career entrants precisely because they already carry professional norms from another field or country – they know how to be effective in their original context, but not yet how to be legible in this one.
So they must learn the technical ideas, the governance debates, the career landscape, and the local communication style all at the same time, before they can become visibly useful to the field.
Many programs are optimized for a student-shaped problem: someone early-career, motivated, flexible, mostly needing context, mentorship, and first opportunities. That’s a real and important use case. But mid-career, non-Anglophone entrants need a different bundle: part-time entry rather than full-time immersion; paid or low-cost opportunities; remote-first participation; explicit explanation of field norms; recognition of demonstrated skill over local credentials; practical sequencing instead of vague advice; and bridges into networks they can’t reach geographically.
When those are absent, the pathway is open in theory and inaccessible in practice. This isn’t anyone’s fault – it’s a design-scope issue. Programs built for one population will under-serve another unless the difference is made explicit.
Some existing programs already help with parts of this problem, especially by making introductory material available online or offering grants and fellowships. My claim is not that nothing exists. It is that access to information does not automatically solve access to networks, credibility, or sustainable entry.
One way to describe this is as a form of fieldbuilding infrastructure debt. AI safety grew partly through small, high-context communities, informal trust networks, and shared intellectual norms. That helped the field move quickly while it was small. But as the field tries to absorb more people from outside those original networks, the missing infrastructure becomes more visible: explicit pathways, low-cost entry points, track-specific artifacts, and ways to build trust without already being inside the network.
1. Create paid, part-time entry paths. Scoped projects, microgrants, part-time fellowships, paid trial tasks, research-assistant work, evaluation or translation projects, short fieldbuilding contracts. The point is not to require abandoning income before becoming legible. Worth naming the obvious objection: funding is finite, and fieldbuilders should be careful about spending scarce resources on people who are not yet proven. But that is exactly why the work should be scoped. A paid task with a tangible output is both a filter and a signal; the money buys verifiable work, not a bet on hope. “Volunteer until someone notices” is a far more expensive strategy than it looks, and it quietly excludes exactly this group.
2. Build remote-first, free or low-cost networks. This is the barrier individuals can’t solve alone, which makes it the natural target for organizations. Free online entry points, remote-first reading groups, low-cost office hours, structured feedback, project-matching, regional or language-specific cohorts, mentorship pools for people outside major hubs. The goal isn’t to replace in-person hubs — they’re valuable — but to stop making hub proximity a hidden prerequisite. If becoming legible requires proximity to London, Berkeley, Oxford, Boston, or DC, the field is leaving talent unused.
3. Recognize demonstrated skill over local credentials. Make it clearer what artifacts count as evidence of fit: an evaluation harness, a model-behavior case study, a governance memo, a red-teaming report, a localization project, a fieldbuilding experiment, a clear writeup of a failed but well-designed attempt. This beats generic "write publicly" advice. Public writing matters for some tracks – research, policy, communications – but for many operator or generalist profiles the strongest signal isn’t an essay, it’s demonstrated execution.
A more useful instruction: produce one visible artifact that shows the kind of contribution you want to make, with different examples for different tracks. Concretely: someone interested in evals might publish a small, reproducible benchmark with a clear threat model, dataset limitations, and failure analysis. Someone with a communications or translation background might localize an important safety resource and add notes on which concepts do not transfer cleanly across languages. Someone with operations experience might document a process improvement for an AI safety project, including the problem, intervention, result, and remaining uncertainty.
I’ll admit the bias here – this is the strategy I bet on myself, building evaluation artifacts rather than waiting to be credentialed. In my case, this has meant building small public artifacts around AI evaluation and agent safety: reproducible eval sets, CLI tools, benchmark-style reports, and GitHub projects aimed at making failure modes inspectable rather than merely describing my interest in the field.
4. Make the pathway legible as a sequence. "Learn, network, write, volunteer, apply" is directionally fine but operationally vague. Better: understand the landscape, choose a tentative track, identify the relevant skill gap, produce one credible artifact, and then seek feedback, scoped work, funding, or applications. Don’t assume student-level flexibility on timing; alongside paid work the same sequence just takes longer. But even a slow path is easier to walk when the order of operations is visible.
5. Don’t reward only those who can go full-tilt. Someone who can study, attend, write, volunteer, network, and build for six months without meaningful income is receiving a subsidy – from savings, family, a partner’s income, or unusually low obligations. That’s not bad, but it isn’t neutral. If the field unintentionally selects for people who can afford long uncertainty, it misses capable people with stronger obligations and equally valuable skills. Sustainability isn’t a luxury concern; it shapes who can enter at all.
AI safety and governance need more than researchers. They need people who can evaluate systems, run programs, build institutions, communicate across cultures, manage operations, translate concepts, and notice how technical ideas meet real-world institutions. Many of those people already exist – mid-career, outside the Anglophone world, immigrants, people who won’t look like traditional candidates at first glance.
This is especially relevant for AI governance and policy. If AI governance is partly about international coordination, regulation, and public legitimacy, then the field cannot afford pathways that mostly recognize people already fluent in Anglophone professional and epistemic norms.
The question is whether the field has pathways that can recognize and absorb these people. If not, the problem isn’t individual. It’s infrastructural.
This is grounded in personal experience, not a representative study – one vantage point, not a survey of the group I’m describing. I’m offering it because the pathway seems under-designed for people outside the usual Anglophone and early-career pipelines. I’d be glad to hear from others with similar or different experiences, and especially from people designing these programs. If this diagnosis is wrong, incomplete, or already being addressed somewhere, I’d like to know.
I don’t mean this as a criticism of existing fellowships, courses, or events. Many of them seem well-designed for the audience they primarily serve. The claim I’m trying to make is narrower: mid-career, non-Anglophone entrants face a different bundle of constraints, and some of those constraints are much less individually solvable than others. I’d be especially interested in examples of programs that already handle this well.