贡献与快速成长之路

Interviewer: How in a year or a year and a half have you guys been, you know, made important contributions to your field?

Sholto Douglas: It goes without saying luck, obviously, and I feel like I've been very lucky in the timing of different progressions. It has been just really good in terms of advancing to the next level of growth. Um, I feel like for the interpretability team specifically, I joined when we were five people. We've now grown quite a lot. Um, but there were so many ideas floating around, and we just needed to like, really execute on them and have like quick feedback loops and like do careful experimentation, um, that led to like Signs of Life and have now allowed us to like really scale. Um, and I feel like that's kind of been my biggest value add to the team. Um, which, it's not all engineering, but quite a lot of it has been interesting.

Interviewer: So you're saying like you came at a point where like there had been a lot of science done and there was a lot of like good research letting around, but they needed someone to like just take that, like maniacally execute on it.

Sholto Douglas: Yeah, yeah. And and and there's, this is why it's not all engineering, because it's like running different experiments and like having a hunch for why it might not be working, and then like opening up the model or opening up the weights and like, what is it learning? Okay, well, let me try and do this instead. And that sort of thing. But, um, a lot of it has just been being able to do like very careful, thorough, but quick, um, investigation of different ideas. I just don't get blocked very often. Like if I'm trying to write some code and like something isn't working, even if it's like in another part of the codebase, I'll often just go in and fix that thing or at least hack it together to be able to get results. And I've seen other people where they're just like, 'Help, I can't!' and it's like, no, that's not a good enough excuse. Like, go all the way down. I've definitely heard like people in management type positions talk about the lack of such people. Where they'll check in on somebody a month after they give them a test, a week after they give them a test. I'm like, 'How's it going?' and they say, 'Well, you know, we need to do this thing which requires lawyers because it requires talking about this regulation.' It's like, 'How's that going?' I was like, 'Well, we need lawyers.' And like, why didn't you get lawyers? I think that's arguably the most important quality in like almost anything. It's just pursuing it to like the end of the Earth and like whatever you need to do to make it happen, you'll make it happen. If you do everything, you win. If you do everything, you win. Exactly.

Agency and High-Leverage Problems

Sholto Douglas: I think from my side, uh, definitely that quality has been important, like agency in the work. There are thousands, I would even like, probably tens of thousands of Engineers at Google who are like, you know, basically like we're all like equivalent, like software engineering ability, let's say. Like, you know, if you gave us like a very well-defined task, um, then we'd probably do it like equival-. Well, a bunch of them would do it a lot better than me, you know, in all likelihood. Um, but what I've been, like, one of the reasons that I've been impactful so far is I've been very good at picking extremely high leverage problems. So problems that haven't been like particularly well solved so far, um, perhaps as a result of like frustrating structural factors, like the ones that you pointed out in like that scenario before, where they're like, 'Oh, we can't do X because this team won't do Y.' Or like, and then going, 'Okay, well, I'm just going to like vertically solve the entire thing.'

Transition to Scaling and Early Career

Interviewer: We should talk about, uh, how you guys got hired because I think that's a really interesting story. So like the T- the of this is, I studied Robotics and undergrad, and in the meantime, on nights and weekends, basically every night from 10 p.m. till 2 a.m., I would do, uh, my own like research and every weekend for like at least six to eight hours each day, I would do my own research and coding projects and this kind of stuff. That sort of switched in part from like quite robotic specific work to after reading, uh, GW's scaling hypothesis post, I got completely scaling-pilled and was like, 'Okay, like clearly the way that you solve robotics is by like scaling large multimodal models.' I was trying to work out how to scale that effectively. And, um, James Bradbury, uh, who at the time was at Google and is now at Anthropic, um, saw some of my questions online where I was trying to work out how to do this properly. He was like, 'I thought I knew all the people in the world who were like asking these questions. Who on Earth are you?' Um, and, uh, he, you know, he looked at that and he looked at some of like the robotic stuff that I've been putting up on my blog and that kind of thing. And he reached out and said, 'Hey, do you want to have a chat? And you want to, um, like explore working with us here?' Um, and, uh, I was hired, I, as I understand it, later as an experiment in trying to take someone with extremely high enthusiasm and agency and pairing them with some of the best Engineers that he knew. Um, and so one another, one of the reasons I could say like, I've been impactful is I, I had this like dedicated mentorship from utterly wonderful people.

Interviewer: What you mentioned about being, um, being bootstrapped immediately by these people might have meant that since you're getting up to speed on everything at the same time, rather than spending grad school going deep on like one specific way of doing RL, you actually can take the global view and aren't like totally bought in on one thing. So not only can it is it something that's possible, but like has greater returns than just hiring somebody out of grad school. Potentially, you come at everything with fresh eyes, um, and come and locked to any particular field.

Sholto Douglas: Um, now, what, like, one caveat to that is that before, like, during my self-experimentation and stuff, I was reading everything I could. I was like obsessively reading papers every night. Um, and like, actually, funnily enough, I, I like read much less widely now that I, like, my day is occupied by working on things. Um, and in some respect, I had like this very broad perspective before where, not that many people, even, even like in a PhD program, you, like, focus on a particular area. Um, if you just like, read all the NLP work and all the computer vision work and like all the robotics work, you, like, see all these patterns just start to emerge across subfields, um, in a way that I guess, like, foreshadowed some of the work that I would later do.

Research Journey and Academic Connections

Interviewer: And Trenton, does this map onto any of your experience? I think Sh's story is more, more exciting.

Trenton Bricken: Um, mine was just very serendipitous in that I, I got into computational Neuroscience. Didn't have much business being there. Um, my first paper was mapping the cerebellum to the attention operation and Transformers. My next ones were looking at, like, you wrote that, uh, it was my first year of grad school, okay? Um, so 22. Oh yeah, but, uh, yeah, my, my next work was on, uh, sparsity in networks, like inspired by sparsity in the brain. Uh, which was when I met Tristan Hume, uh, and Anthropic was doing the solution, the softmax linear output unit work, which was very related in quite a few ways. Of like, let's make the, uh, activation of neurons across a layer really sparse. And if we do that, then we can get some interpretability of what neuron's doing. That started the conversation. I shared drafts of that paper with Tristan. He was excited about it. And and then, and and that was basically what led me to become Tristan's resident and then convert to full-time. Um, but during that period, I also moved as a visiting researcher to Berkeley, uh, and started working with Bruno Olous. And Bruno Olen basically invented sparse coding back in 1997. And so it was like, the, the, the, my research agenda and the interpretability team seemed to just be running in parallel, um, in in with just research taste and and so it, yeah, it made a lot of sense for for me to work with the team.

Sholto Douglas: Um, well, and it's been a dream. Since one thing I've noticed when people tell stories about their careers or their successes, they ascribe it way more to contingency, but when they hear about other people's stories, they're like, 'Of course, it wasn't contingent.' You know what I mean? It's like, if that didn't happen, something else would have happened. Yeah. But I mean, like, I literally met Tristan at a conference and like wasn't, didn't have a scheduled meeting or anything, just like joined a little group of people chatting, and he happened to be standing there, and I happened to mention what I was working on, and that led to more conversations. And I think I probably would have applied to Anthropic at some point anyways, but I would have waited at least another year. I, I, I, yeah. It's still crazy to me that I can like actually contribute to interpretability in a meaningful way.

Sholto Douglas: I think there's an important aspect of like, shots on goal, there so to speak, right? Where like you, even just going to choosing to go to conferences itself is like putting yourself in a position where you're, where luck is more likely to happen. My own was my own way of like, trying to manufacture luck, so to speak, um, and and like, try and do something meaningful enough that it got noticed. For the people who are like, 'just assuming that the other end of the job board is like just like super legible and mechanical, this is not how it works.' And in fact, like people are looking for the sort of different way, different kind of person who's agentic and putting stuff out there. And I think specifically what people are looking for there is two things: one is agency, and like putting yourself out there. Uh, and the second is the ability to do world-class something. Yeah, Andy Jones from Anthropic did an amazing paper, um, on scaling laws as applied to board games. It didn't require much resources, it demonstrated incredible engineering skill, it demonstrated incredible understanding of like the most topical problem of the time. Um, and he didn't come from a like typical academic background or whatever. As I understand it, basically, like as soon as he came out with that paper, both ends, R and Open AI were like, 'We would desperately like to hire you.'

The System and the Importance of Caring

Sholto Douglas: There's this line: 'The system is not your friend,' right? Uh, and it's not necessarily to say it's it's actively against you, it's your sworn enemy. Um, it's just not looking out for you, right? And so I think that's where a lot of the proactiveness comes in. Of like, there are no adults in the room, or like, and and like you have to come to some decision for what you want your life to look like and execute on it. And and yeah, hopefully you can then update later, um, if you're too headstrong in the wrong way. But but I think you almost have to just kind of charge at certain things to get much of anything done, not be swept up in the tide of whatever the expectations are.

Sholto Douglas: There's like one final thing I want to add, which is like, we talked a lot about agency and this kind of stuff, but I think actually, like surprisingly enough, one of the most important things is just caring an unbelievable amount. Um, and when you care an unbelievable amount, you, like, you check all the details and you have like this understanding of like what could have gone wrong. And you, like, you, uh, it just, it matters more than you think. Because people end up not caring, not caring enough. Uh, this is like LeBron quote where he talks about how when he sort of before he started in the league, he was like worried that everyone would be like incredibly good. And and then he gets there and he like realizes that actually once people hit financial stability, then they, um, like they relax a bit. And he's like, 'Oh, this is going to be easy.' Um, and I don't think that's quite true because I think in like AI research, because most people actually care quite deeply. Um, but there's caring about your problem and there's also just caring about the entire stack and everything goes up and down. Like going explicitly going and fixing things that aren't your responsibility to fix because overall it makes the stack better. I, something that a friend said to me a while back, but I think has stuck, is like, 'It's amazing how quickly you can become world class at something just because most people aren't trying that hard and like are only working like, I don't know, the actual like 20 hours that they're actually spending on this thing or something.' And so yeah, if you just go ham, then like you can, you can get really far, pretty fast.