I Started an AI Safety Research Org and Think These 7 Things Matter

Thanks to Adam Jones and Ben Smith for suggesting I make this.How and Why it Started When I finished the 2024 AI safety fundamentals course by Blue Dot Impact, there was a small tick box saying something like 'I want to start an org in this space'. I was just coming to the end of my artificial intelligence MSc, and wanted to start an AI safety research organisation. I had tried once and done various other entrepreneurial things, and was looking for precisely this opportunity. Turns out the tick box was put there by then alignment course lead, Adam Jones, who invited me to co-work with him on starting a new AI safety organisation on something related to AI-enabled oligarchies.We brainstormed for a bit, and I ended up focusing on lock-in, which seemed highly neglected as a problem, and I was excited to explore what an org could do in the space. BlueDot offered me a small grant to co-work with them in an incubator-style arrangement at LISA. I spent 8 weeks working with them on what the org should do, and making longer-term funding applications to start Formation Research.Fast-forward to now, and I am recruiting a founding team to work on an empirical research agenda for secret loyalties, aiming to answer foundational science of deep learning questions that will help us build defences in AI labs against secretly loyal AI systems. While the org has not yet scaled or built a founding team, I have learned a lot about what is important when starting a venture like this, which I'm writing up here for other people looking to do the same thingWhat I think is Important1. Do things as in-person as you can.I spent some time working on the org remotely. Compared with living in London, working out of LISA full-time, and attending conferences and events, I think I made much less progress being remote.Feedback loops were possible when working remotely, but I was more isolated from the community. Even if I read blog posts and chatted with Adam, this only happened every now and then and