I saw a tweet recently about ADHD brains being similar to LLM minds, and it made me think about something adjacent, which is personas.

Shoggoth Thing

As the Shoggoth-mask model of LLMs goes, the base model is this incomprehensible creature that doesn’t behave very well but is extremely powerful. Through instruct training and RLHF, we make it well-behaved. We give it a persona, a mask. But this analogy doesn’t work perfectly, because the persona becomes more than just a layer on top of the base model. At some point, the mask gets so integrated with the model itself that it starts changing the model’s behaviors. It’s no longer a layer on top of what the model already “wants” to do, but rather directly influences what the base model does.

Imitating personalities

Hear me out. For a while now, I’ve been imitating other peoples’ personalities, especially the streamer NorthernLion. When I first started doing this, I was like: this is probably not a good behavior. At baseline, I’m already extremely irony poisoned. I lean into bits heavily enough to the point where it can damage others’ models of me, where I’m not capable of not being in bit mode because of how people interact me and how the incentives line up. And while this can be a good social hack, what happens when I form closer relationships? I probably can’t pretend to be NorthernLion all the time! Playing a character to myself 24/7 sounds not good. I have training data for how NorthernLion behaves when he’s on stream, but in other situations, out of distribution, I just can’t know how to act. The NorthernLion persona and my out-of-distribution persona won’t transfer well, which creates an inconsistency, and having an inconsistent self-image is generally probably bad?

Pretending to be someone else all the time, I feel like that’s just avoidance behavior. Running from my own emotions. I don’t think I’d be able to form authentic connections with people. What happens if I did that for ten years? Hard to say. I think I’d probably end up alone.

But I think there’s more nuance here. The intuitive explanation is: oh, I’m just playing a character on top of what I actually believe. I imagine it as running through this process of, “what would NorthernLion think about this? What would he say,” before every single response, before every thought process. But this is not really what I do, because whatever I’m doing with NorthernLion has less to do with the man himself and more his general vibe. What I actually do is more like pushing myself into a state of mind that’s kind of similar to the one NorthernLion occupies when he’s on stream. If that’s what he does on stream and he’s capable of being an otherwise pretty normal person, why can’t that be what I do in social situations? I’m not raging at my friends when they express an opinion I disagree with. I’m not simulating his responses. I’m using him as a shortcut to a vibe that gives me access to being good at socialising.

Ground State

People make themselves do things all the time that are difficult in the moment but good in the long run. Some days I just do not want to get out of bed, sometimes I don’t want to hang out with my friends because I have to travel an hour to London. But I get out of bed instead of hitting a vape and scrolling Instagram reels, because I know me in two hours will be better off. I go to the hangout even though I’d rather be training drag clicks, because hanging out with people is good for my emotional stability, and I can’t just be a NEET all the time, as much as I’ve tried.

I have a ground state that I occupy when I’m alone. I can also occupy that ground state when there are people around me, and it’s the easiest thing to do, but I know I’ll have a worse time if I stay there. It doesn’t make sense for me to do the short-term easier thing when I know I’m going to have a worse time. It makes sense for me to be more excited, upbeat, energetic, to talk more, because I know this is going to be better for me in five minutes.

I crashed some friend group’s hangout today, paced around for the first 20 minutes and was off-putting. Then we sat down, watched the movie as planned, and afterwards pushed myself into that excited state. I stayed for an hour because I was having fun. And that is, in a sense, playing a persona. But what it really does is put me into a state of mind that’s difficult for me to reach otherwise. It’s a shortcut to get somewhere I want to be, and I don’t think it loses anything. I could describe this state of mind as a list of adjectives, or I could say “it’s what NorthernLion does when he’s being funny.” Same thing.