I think the strongest part of this post is the emphasis on shutdown as a signal, not only as direct risk reduction.
That said, I wonder whether the relevant move has to be full organizational shutdown, rather than a public frontier-development freeze: no new frontier training runs, no capability-scaling race, continued service of existing models, and a major reallocation toward safety research, evaluations, external auditing, and coordination.
That might send a stronger signal than 'we are racing, but responsibly,' while avoiding some obvious objections: loss of internal safety capacity, investor/employee backlash, and the public interpreting shutdown as financial failure.
Maybe the key distinction is: do we need an exit from the company, or an exit from the race?
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.
One fieldbuilding gap I’d like to see discussed more is how people from non‑English or non‑US backgrounds can enter the AI safety space. A lot of AI safety material assumes not just the technical ideas, but also familiarity with EA or LW culture, US/UK career norms, and a specific style of public reasoning. For people with relevant skills from other countries or languages, the challenge isn’t only translation - it’s also making the field’s norms, expectations, and opportunities understandable. I wonder if localized ‘on‑ramps’ for skilled mid‑career people could be especially valuable.
I get the idea of a 'generalist gap,' but I think it helps to separate generalists from legible generalists. From the outside, it’s hard to know which signals actually matter - public writing, volunteering, technical side projects, eval work, operations, community‑building. My sense is that many strong people aren’t lacking motivation or skill; they just don’t know how to turn their past experience into proof that AI safety orgs can actually evaluate. It would be great to have a clearer list of concrete proof‑of‑work examples for different generalist paths, instead of assuming public writing is the default.
Thanks for writing this, it really resonates. I’m also trying to move into AI safety from a non‑traditional background: media and editorial work, AI evals, and now small safety‑focused technical projects. I’ve noticed that 'write publicly' is helpful, but often too vague. For many generalists, clearer proof‑of‑work might be more useful - like a small eval setup, a case study, a reproducible repo, or a field‑building experiment with concrete results. I’d love to see more guidance that connects different generalist backgrounds to specific proof‑of‑work projects, instead of treating public writing as the default path.
I think the strongest part of this post is the emphasis on shutdown as a signal, not only as direct risk reduction.
That said, I wonder whether the relevant move has to be full organizational shutdown, rather than a public frontier-development freeze: no new frontier training runs, no capability-scaling race, continued service of existing models, and a major reallocation toward safety research, evaluations, external auditing, and coordination.
That might send a stronger signal than 'we are racing, but responsibly,' while avoiding some obvious objections: loss of internal safety capacity, investor/employee backlash, and the public interpreting shutdown as financial failure.
Maybe the key distinction is: do we need an exit from the company, or an exit from the race?
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.
One fieldbuilding gap I’d like to see discussed more is how people from non‑English or non‑US backgrounds can enter the AI safety space. A lot of AI safety material assumes not just the technical ideas, but also familiarity with EA or LW culture, US/UK career norms, and a specific style of public reasoning. For people with relevant skills from other countries or languages, the challenge isn’t only translation - it’s also making the field’s norms, expectations, and opportunities understandable. I wonder if localized ‘on‑ramps’ for skilled mid‑career people could be especially valuable.
I get the idea of a 'generalist gap,' but I think it helps to separate generalists from legible generalists. From the outside, it’s hard to know which signals actually matter - public writing, volunteering, technical side projects, eval work, operations, community‑building. My sense is that many strong people aren’t lacking motivation or skill; they just don’t know how to turn their past experience into proof that AI safety orgs can actually evaluate. It would be great to have a clearer list of concrete proof‑of‑work examples for different generalist paths, instead of assuming public writing is the default.
Thanks for writing this, it really resonates. I’m also trying to move into AI safety from a non‑traditional background: media and editorial work, AI evals, and now small safety‑focused technical projects. I’ve noticed that 'write publicly' is helpful, but often too vague. For many generalists, clearer proof‑of‑work might be more useful - like a small eval setup, a case study, a reproducible repo, or a field‑building experiment with concrete results. I’d love to see more guidance that connects different generalist backgrounds to specific proof‑of‑work projects, instead of treating public writing as the default path.