Being Safety Pilled and Fighting the Depill
I consider myself to be someone who is AGI-pilled: I believe that the capabilities of artificial intelligence (AI) will exceed human capabilities in most, if not all, productive tasks. I believe this has significant implications for the economy and broader society.
Not only that, I also consider myself safety-pilled: I believe that AI models will become capable enough to pose significant threat to our current way of life, and subsequently, I strongly believe that significant progress on AI safety is critical to ensure humanity’s survival and thriving.
In this piece, I do not seek to explain or justify this position (I’ll leave that to a later post). Instead, I want to write about something I’ve been experiencing recently: the struggle of trying to cling to such strong beliefs in an adversarial environment.
I recently was volunteering at Recursive, a conference gathering various researchers and thinkers to convene on the topic of AI R&D automation. Various presentations and discussions brought up the scaling laws and failure worlds and all the reasons that our default path in AI development is probably not going to go well. And then I reflected on how I’d been spending my time, and how little of my recent time spent actually sounds coherent under this worldview.
It’s not like that was the first time I’d heard these arguments. I’ve read many a blog post, agree with many of the claims, and semantically could produce strong arguments for AI Safety if asked. But I’ve not absorbed the implications of these arguments on a visceral enough level for it to change my base worldview. Only when I engage in direct discussion with others who “get it” do I actually experience a limbic response to “AGI development.”
The issue with this is that, on Stanford’s campus, I’m not frequently around people who even believe that AI with human-level capabilities might be developed, let alone that its deployment might go poorly and thus merit a worldview update. When I’m on campus, I spend most of my time doing work unrelated to AI Safety, like schoolwork and social dancing and spending quality time with beloved (not safety-pilled) friends. In the absence of safety-pilled discussion, I begin to de-pill.
It’s not that I lose my context on AI Safety and its importance, but rather that I lose the sense of urgency. If you ask me about my future plans and what I care about in the world, I will immediately bring up AI Safety. But the arguments have not changed my worldview enough to persist in my day-to-day decisions when I’m at Stanford, which for a number of other reasons is actually quite an adversarial environment for staying safety-pilled.
Fighting this de-pill is hard. My life at Stanford is full of so many deeply fulfilling relationships and endeavors, but the vast majority of them are not at all related to working on AI safety, and doing AI safety stuff means taking away from this lifestyle as it taunts me in the face for leaving.
The obvious solution is to safety pill all of my friends. Unfortunately, that would corrupt the very nature of my friendships. Another solution is to make a ton of new friends that become Stanford Effective Altruism (Stanford EA) and Stanford AI Alignment (SAIA) members. This is the current plan. Without a community of people, it seems very hard to hold onto strong fringe beliefs, and it is very easy to get de-pilled when the environment adversarially pushes toward mainstream beliefs, like grinding to work on frontier AI capabilities. Thus, my conviction moving into next year is to build Stanford EA and SAIA into havens for having open critical discussions of fringe opinions and preventing the de-pill for those who wish to remain pilled. I’m really excited for these organizations to fulfill their potential and begin absorbing the attention and interest that I’ve been seeing in nascent EAs and safety-pilled people, and I’m eager to see more people taking AI Safety seriously who’d never even encountered its arguments.
As a last note, this is why I’m super excited to be a Generator Resident this summer! I’ve met about half the residents so far, and besides being obviously very safety-pilled, they’re also generally very personable, thoughtful, hard-working, agentic (I wish there were a better word that didn’t have all the annoying connotations), and fun to be around! I’ve already had a lot of fun and deep conversations, and I’m really excited to spend this summer working on something that I feel is very important while being surrounded by fun people who also care deeply about the problems I care about.