The case for AI safety capacity-building work
A major funder says recruiting talent for AI safety is more impactful than direct technical work.
A viral post on LessWrong from a researcher at the philanthropic funder Coefficient Giving (formerly Open Philanthropy) is making a provocative case: the most impactful work in AI safety right now isn't necessarily building new models or writing policy papers—it's building the talent pipeline. The author, part of a team that granted over $150M in 2025, argues that 'capacity-building'—recruiting, training, and placing skilled professionals into AI safety roles—offers a powerful 'multiplier effect.' The logic is that convincing one equally talented person to dedicate their career to AI risk can double the impact, a strategy deemed more tractable and scalable than many direct interventions.
The post is a direct call to action, particularly aimed at technical staff at leading organizations like Anthropic's safety teams, Apollo, and Redwood Research. It suggests that marginal hires at these top labs could have more impact by founding or joining new organizations focused on talent development, citing successful programs they've funded like the ML Alignment & Theory Scholars (MATS) program and career guide 80,000 Hours. The argument is backed by survey work and anecdotes indicating past capacity-building efforts have had large, predictable effects on the field's growth, positioning this work as a critical, under-supplied lever for mitigating catastrophic AI risks.
- The post argues for a 'multiplier effect': recruiting one person into AI safety can be more impactful than doing the direct work yourself.
- Coefficient Giving's capacity-building team granted over $150M in 2025 to programs like MATS, BlueDot Impact, and 80,000 Hours.
- It's a call for talent at top AI safety labs (e.g., Anthropic, Redwood) to shift focus to building new talent-recruitment organizations.
Why It Matters
It reframes the priority for mitigating AI risk from pure technical research to urgently scaling the human capital needed to do it.