Dwarkesh Patel: belief

Transcript context

…Exactly. I think from my side that quality has definitely been important: agency and work. There are thousands, probably tens of thousands of engineers, at Google who are basically equivalent in software engineering ability. Let's say if you gave us a very well-defined task, then we'd probably do it with equivalent value. Maybe a bunch of them would do it a lot better than me in all likelihood. But one of the reasons I've been impactful so far is I've been very good at picking extremely high-leverage problems. I mean problems that haven't been particularly well-solved so far, but perhaps as a result of frustrating structural factors like the ones that you pointed out in that scenario before, where they're like, “we can't do X because this team won’t do Y.” Well, I'm just going to vertically solve the entire thing. And that turns out to be remarkably effective. If I think there is something correct, something that needs to happen, I'm also very comfortable with making that argument and continuing to make that argument at escalating levels of criticality until that thing gets solved. I'm also quite pragmatic with what I do to solve things. You get a lot of people who come in with, as I said before, a particular background or a familiarity. One of the beautiful things about Google is that you can run around and get world experts in literally everything. You can sit down and talk to people who are optimization experts, TPU chip design experts, experts in different forms of pre-training algorithms or RL or whatever. You can learn from all of them and you can take those methods and apply them. I think this was maybe the start of why I was initially impactful, this vertical agency effectively. A follow-up piece from that is that I think it's often surprising how few people are fully-realized in all the things they want to do. They're blocked or limited in some way. This is very common in big organizations everywhere. People have all these blockers on what they're able to achieve. I think helping inspire people to work in particular directions and working with them on doing things massively scales your leverage. You get to work with all these wonderful people who teach you heaps of things. And generally helping them push past organizational blockers means that together you get an enormous amount done. None of the impact that I've had has been me individually going off and solving a whole lot of stuff. It's been me maybe starting off in a direction, and then convincing other people that this is the right direction, and bringing them along in this big tidal wave of effectiveness that goes and solves that problem. We should talk about how you guys got hired. Because I think that's a really interesting story. You were a McKinsey consultant, right? There's an interesting thing there. I think generally people just don't understand how decisions are made about either admissions or evaluating who to hire. Just talk about how you were noticed and hired. So the TLDR of this is I studied robotics in undergrad. I always thought that AI would be one of the highest-leverage ways to impact the future in a positive way. The reason I am doing this is because I think it is one of our best shots at making a wonderful future basically. I thought that working at McKinsey, I would get a really interesting insight into what people actually did for work. I actually wrote this as the first line in my cover letter to McKinsey. I was like, “I want to work here so that I can learn what people do, so that I can understand how to automate work.” In many respects, I did get that. I just got a whole lot of other things too. Many of the people there are wonderful friends. I think a lot of this agentic behavior comes in part from my time there. You go into organizations and you see how impactful just not taking no for an answer is. You would be surprised at the kind of stuff where, because no one quite cares enough, things just don't happen. No one's willing to take direct responsibility. Directly responsible individuals are ridiculously important and some people just don't care as much about timelines. So much of the value that an organization like McKinsey provides, is hiring people who you were otherwise unable to hire, for a short window of time where they can just push through problems. I think people underappreciate this. So at least some of this attitude of “hold up, I'm going to become the directly responsible individual for this because no one's taking appropriate responsibility. I'm going to care a hell of a lot about this. And I'm going to go to the end of the earth to make sure it gets done,” comes from that time. More to your actual question of how I got hired. I didn't get into the grad programs that I wanted to get into over here, which was specifically for focus on robotics, and RL research, and that kind of stuff. In the meantime, on nights and weekends, basically every night from 10pm to 2am, I would do my own research. And every weekend, for at least 6-8 hours each day, I would do my own research and coding projects and this kind of stuff. That sort of switched in part from quite robotic specific work. After reading Gwern’s scaling hypothesis post, I got completely scaling-pilled and was like, “okay, clearly the way that you solve robotics is by scaling large multimodal models.” Then in an effort to scale large multimodal models with a grant from the TPU access program, the Tensor Research Cloud, I was trying to work out how to scale that effectively. James Bradbury, who at the time was at Google and is now at Anthropic, saw some of my questions online where I was trying to work out how to do this properly and he was like, “I thought I knew all the people in the world who were asking these questions. Who on earth are you?” He looked at that and he looked at some of the robotic stuff that I'd been putting up on my blog.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.