A good way to model the strategy of most AI safety groups is as career pipelines.

They raise awareness about catastrophic and existential risks from AI, and they help people do productive work to reduce these risks. A pipeline is a set of activities that increase the probability that someone does quality work in AI safety. Your activities should fit well together to create a path for a person to go from “interested in AI safety” to “working on AI safety”.

Note that “my group is a career pipeline” does not mean that, at every step of the way, you should ask, “How is this increasing the number of careers?” There are many valuable proxies to optimize for, such as quality discussions about AI safety and the number of highly capable people signing up for internships, and these are often more practical to think about. But ultimately, it’s important to think about your group’s impact and what success actually looks like.

Also, it’s important to account for the quality of careers. While the number of AI safety jobs is growing, it’s still quite limited, meaning that if you optimize for the number of graduates who apply for AI safety jobs (versus the number that are likely to get them), many of them might end up rejected and disillusioned.

If you don’t think someone is likely to get a career right away, it might be worth optimizing for them to stay in touch with the field and not drift away. However, it’s difficult to predict how the demand for AI safety talent will evolve. We foresee the number of jobs increasing rapidly in the next few years, especially in the AI governance space, where new bureaucratic institutions are popping up around the world and have the capacity to hire more people. We also predict technical research and operations roles will grow as well. However, if the only role a group member is willing to consider is a research role at Anthropic, this seems less tractable. To see a list of some core companies and organizations that have AI safety opportunities that we are keeping tabs on, check this section out.