A recent post on LessWrong argues that the AI safety field is failing to utilize a surge of highly qualified interest, creating an artificial scarcity of researchers that threatens to stall critical progress.
\nIn a recent post, LessWrong contributor lessw-blog discusses a counter-intuitive dynamic currently shaping the AI safety landscape. For years, the prevailing narrative within the field has been that progress is constrained by a lack of talent-specifically, a shortage of researchers possessing the unique combination of technical capability and philosophical alignment required to tackle existential risks. However, this new analysis suggests the reality is quite different: the field is currently drowning in qualified applicants it cannot absorb.
The post highlights the severe scaling issues facing major AI safety fellowship and residency programs. Using the Anthropic residency as a prime example, the author notes that the program reportedly received over 2,000 applications for a very small number of positions, resulting in an acceptance rate of less than 1.3%. While high selectivity is often viewed as a marker of prestige, the author argues that in this context, it represents a systemic failure. The rejection of candidates with "absurdly high qualifications" indicates that the bottleneck is no longer the supply of talent, but the field's capacity to mentor, manage, and direct that talent effectively.
This distinction is critical for stakeholders in the AI ecosystem. If the constraint were truly a lack of talent, the solution would be outreach and education. However, if the constraint is organizational capacity-a lack of senior mentors or funding structures to support junior researchers-then current strategies are misaligned. The post suggests that the field is effectively turning away capable individuals who wish to contribute to safety, potentially pushing them toward capabilities research or out of the industry entirely. This creates a "scaling problem" for safety research itself; as AI models scale exponentially, the human systems designed to ensure their safety are failing to scale alongside them.
The analysis points to a need for new infrastructure within the AI safety community. Rather than simply seeking "more dakka" (more raw firepower or effort) in the form of applicants, the field requires better mechanisms to onboard and utilize the people who are already knocking on the door. Without solving this operational bottleneck, the gap between the pace of AI development and the pace of safety research will likely continue to widen.
For those involved in technical recruitment, research operations, or AI policy, this post offers a sobering look at the inefficiencies hampering risk mitigation efforts. It challenges the community to move beyond elite gatekeeping and toward scalable research management.
Read the full post on LessWrong
\n\n