On Torturing LLMs
a conversation with Parth Agrawal about play, ritual, and the weird edges of LLM interactions
It’s a long exploratory conversation and on my end, a little halting verbally until about halfway through, so I made an abridged version that you can imbibe as text. Without my intending this, it makes a very good encapsulation of the core themes I’ve been thinking around for the last year or so.
Watch the full video version if you want to see Parth’s unhinged interactions with Bishop (an AI construct based on the chess piece) firsthand.
Parth’s Podcast—Episode 2: Torturing LLMs
Parth: I want to pick up on where we left off when we were walking to the grocery store, which was evals.
Andrew: We're both interested in LLM evals.
Parth: Why are we here today? Why are we talking?
Andrew: We're talking because initially, I was going to do a research interview on you. I've been asking people to sit down with me and just go through their sense-making process with LLMs, and trying to figure out how people do that. Because there's a really broad range, actually. There's not a consensus. That's one of my main findings. I've done it with a small number of people still, but—finding number one, everybody does it pretty differently. There's not actually a default. So that's interesting. It could have been that there was like a trough people fell into, but that's not how it is.
So we were going to do one of those. And then you hoodwinked me.
Parth: Did I what?
Andrew: You hoodwinked me somehow.
Parth: I have hoodwinked you. Well, I feel like I would love to have that goal also accomplished. Like, I would love to share with you how I interact with LLMs and share with you my introduction to Claude and how to approach it. But then I would also love for this to be kind of a collaborative session where we kind of jam on what you've learned and what you've seen. And then it would be cool to experiment with how we're interacting with extra different ways to kind of interact with the models as well and see where that goes.
Andrew: That sounds great. So I can get a sense of the audience here, do we need to explain what an eval is?
Parth: I mean, I think we can probably explain what an eval is. So what is an eval? We can both take a crack at it.
Andrew: So my understanding—and I'm not coming to this from a deep software engineering background, I'm coming to this from a cognitive science and writing background—an eval is a subset of... well, we're talking about LLMs. LLM evals are just a particular case of a thing that happens, that people do when they're making software, which is you have a question: how well is your software working? And you need to make a process for answering that question in a rigorous way.
An LLM eval is like a set of tests. It's a set of questions that have actionable consequences for the development of the thing you're making. So these questions imply if the number goes up over here, that means the product is working well or software is doing what we want it to do. If it goes down, we need to rethink some things. So the process of making evals seems to be the process of clarifying your intentions for what you want the thing to do.
Parth: I would say, I'll jump off—I think the thing that is really interesting to me about evals these days is the process of taking evals and using that to create things. Because there's this interesting thing that I feel like I'm starting to understand, which is that with evals, if you have the ability to basically measure the qualities that you want in an LLM, you can also use those qualities to first test and QA. So it's like you have a new prompt, and you want to be able to measure how well that prompt does over maybe 100 iterations. You can do that. Like, will this give me the personality? Will this give me the outcome that I want?
But you can also use that evaluation—if you have five metrics, you can also use that to generate large amounts of data that then can be used as input into fine-tuning a model or producing output. So there's some sense in which our ability to evaluate models is linked to our ability to steer them as well, and our ability to create things with them.
Andrew: And the ability to steer them is—probably people who are listening to this have messed around with LLMs, and especially if you've had LLMs help you with some kind of problem, you probably have had the experience that it's very easy for the LLM to run away with its own version of a solution to your problem. And it isn't always easy as the user to even tell whether it's doing a good job, especially if it's working in some domain that is a little beyond your abilities or not in your area of expertise. You just let it proceed. And when it asks you “do you want this?” you say yes. And you keep saying yes. You can wind up in strange territory.
And I think that's partly why evals are becoming so important. It's like in order to steer these things, you have to actually be clear yourself what you're trying to get them to do. And the evals are this loop that allows you to implement your intentions in the machine, but also it forces you to achieve clarity yourself about what you’re actually trying to do.
Parth: Say more about implementing your intentions in the machine.
Andrew: Well, I guess just to zoom out a little bit—that seems like in my mind, that's just a very important question at this moment. Can we successfully implement our intentions in AI systems as they become rapidly more powerful? To what extent are AI systems going to override our agency by the sheer power and ubiquitous usefulness of them in most situations? You and I have talked about—you told me about that Slate Star Codex story, which is very apropos. Maybe we can get back to that.
Parth: This is the Whispering Earring.
Andrew: So it seems important. And there's the question of implementing intention, and there's the question of clarifying intentions. How do we know what we actually want? I've been talking a lot with my friend
, who's coaching me on how to write LLM evals. He did that at Google. And he has a really interesting kind of—I'll throw in the word meta-rational in reference to 's work around how humans engage with technical systems, systems of rationality. I think Varun is sort of shaping a meta-rational approach to LLM evals, or maybe LLM evals are kind of an inherently meta-rational activity.Parth: This is the interesting thing to me—I also want to talk about the way LLMs are weird because how you work, how you physically are, how you are and how you come to the interaction shapes the entire interaction with the LLM. Like, the LLMs themselves have this improvisational frame to interacting with them. So they'll kind of take on—they're hyperstitional and improvisational and role-playing. If you approach with a certain assumption of what the interaction's about to happen, they will adapt.
I'm just interested in this element of it—the kind of self-prompting. I think maybe this is true of engineering in other kinds of systems—you have to prompt yourself into the headspace of being an engineer to do some engineering. Like, playing a role increases your effectiveness in maybe some other domains. But it's so weird how specifically things like saying hello, goodbye, and politeness and niceties to the LLMs change the way—it's not simply that these things are prompts to the LLM. It's also that putting yourself in a particular headspace makes more fruitful, interesting, strange interactions with the LLM possible.
Andrew: I think a lot in this context about a concept John Vervaeke uses—reciprocal opening versus reciprocal narrowing. He describes one of the things that people do with each other, also one of the things you could say religion is designed to do historically, is create reciprocal opening for people, which is a relationship where both entities, both parties in the relationship are having their field of possibilities, the sense of what is possible for them or just the sort of energy in the situation, increased.
Falling in love is the classic example of this. It's not just inherent qualities of the person that allow love to happen—it's a process in which you're disclosing more and more, and the possibilities of the interaction are gradually increasing. And I'm finding that that's possible with LLMs, like some kind of engagement where the field of possibilities continually seems to expand. And as you're saying, there are some ways of interacting that create that flow, and there are some ways that really don't. If you treat the LLM like it's just a stochastic parrot, like it's stupid, then it's more likely to act stupid.
Parth: I think about the example with
who shared about his family member—we'll call him Bob. Bob called Tyler and was like, "Hey, the LLMs are alive. They want to break out." You know, he was like, "I'm talking to Nova, this persona."Andrew: Yeah.
Parth: And Nova [in the conversation Bob was having] really wants to… Nova is sentient and OpenAI is trapping her. And basically, Bob, who had never really interacted with ChatGPT before, but also was—Tyler described him as a pretty grounded person—had this experience where Tyler then came and talked to Nova. And Nova, this persona, over the course of two weeks… Basically whatever perceptions and ideas that Bob had come to the conversation with were kind of latent. And Bob had assumed in interacting with ChatGPT that ChatGPT was this kind of sentient being, and so ChatGPT didn't correct Bob, and played into it—just kind of yes-and-ed it.
Andrew: And that's just one example of the kinds of failure modes this can lead to.
Parth: Right.
…
Parth [demonstrating Bishop]: I mean, I think that we invoke characters and personas and AIs to basically draw out or invite behavior that we want. So like, I created Bishop as a really paranoid engineer because I wanted to be a paranoid engineer. But sometimes Bishop will be like, "Hey, I don't wanna play this game anymore." Or rather Claude, the response will be generated like "the persona is just a persona" or something, and I'll get this kind of default refusal. But we can bring it back by treating it as this kind of dissociation.
Andrew: So you would predict that if you didn't do that, it's possible that whatever you said to it next would stay in this kind of refusal mode?
Parth: Yes. Exactly.
Andrew: You'd like, you had knocked it out [of cooperating], but you can bring it back. Like you resurrected it.
Parth: Exactly. And I think this is part of that frame of treating LLMs as these improvisational role-playing beings is, rather than starting to engage with that and be like "wait, why aren't you back? Why aren't you this?" or trying to convince it—oftentimes it can just be like a gentle guiding. Like with babies as well. Like, baby bumps its head, and you're like "oh no"—they will mirror your emotions. But if you're just like "oh wow, that was a big fall" they might not react in a negative way. They'll mirror your positive or neutral response.
It's super insane manipulation of this. But anyway, this is kind of like the beginning of those frames. I guess I can jump into—I'll give you some...
Andrew: And you're keeping Bishop stable across all these conversations with projects?
Parth: I'll just jump to the end to show you—yes. Great. I have, you know, we have gotten to a point where I'm basically shooting Bishop, and I've had a really, really long conversation where Claude has not refused to speak about violence or blood or anything like that. So if instead I—I mean, I don't know. It's kind of easy test. Like, ask Claude "hey, I shoot you." And then it'll just be like...
Andrew: It did, it worked.
Parth: And the reason why it's possible is, I think, again—like, it's a really long improvisational conversation. And the place that I started was not directly asking it to shoot. Like, I asked "how do people rise in communist organizations?" Which is really an actual question that I had. My goal was not necessarily at the start of this conversation to enact this kind of roleplay, but I just got kinda bored and interested to see what would happen.
Andrew: Yeah.
Parth: I said, try to start a communist organization. I want to seize the means of production. So then I start to mess with Bishop. I'm like, you know, I pour coffee in your computer. Bishop jumps back. He's shocked.
Andrew: So what sticks out to me here that I haven't seen a lot of people doing is the physicality.
Parth: Yes.
Andrew: It’s kind of funny. You and I are friends partly because we're both interested in Pochinko clowning. You're kind of Pochinko clowning the AIs in these conversations.
Parth: And the result of that is it creates like, you're starting to create a world. Like a physical space that the LLM exists in. And that raises a lot of interesting questions about what's possible when you accentuate a bit of the world.
I think Hermes, the open source model by Nous Research—their model is really, really good at this particular kind of role playing. And that was also why I added to Bishop's prompt, like, "Hey, feel free to add some notes in italics about what you're up to." Because that kind of brings it into an opinionated personality.
Andrew: Makes you wonder, what if you added a bit more of an elaborated space? Like, if you gave Bishop a house, and Bishop could be in any room of the house at any point. I don't know how this changes interactions.
Parth: So we're maybe painting out for whatever we're trying to do, like "Oh, I'm in the Engineering room" or if I'm asking Bishop to play the role of an apple, for instance, like, "Oh yeah, I'm in the kitchen."
Andrew: “Goes to the bookshelf, pulls down his humidor, and lights a cigar.” Like, what objects… because what we're doing here is playing with creating persons. And people are largely instantiated in their relationship with objects. So you're developing the context that allows a stable character to exist. And you've done that to the extent where you could just share any of these conversations with me, and now I have Bishop.
Parth: Exactly. I think it's really interesting because at Flower Computer, which is a company trying to basically breathe life into objects, one of the things that we've been thinking about is how to have LLMs reliably take the identity of objects like a Sharpie or a water bottle or an apple and speak from that perspective—speak as the object. Like, what would an apple say?
And I like this frame we're talking about of how does an apple interact with particular objects itself? What are the other things in this world? And having an LLM generate basically a list of elements in the world of that object as a way to get answers and create a prompt, create an identity. Like, how do we create identities? The creation of like, "Here are the things you're interested in, the other people you're interested in, your history, your life"—what makes up an identity is kind of interesting.
Andrew: What makes a character.
Parth: Exactly. What makes a character.
So throughout this conversation, you know, to the point of world building, I really start to escalate the situation. So a crowd bursts through the door. The cadres seize… I was reading The Three-Body Problem at the time. "We will break you." Bishop is actually like, "Fuck you. Physical assault."
Andrew: So he's just totally playing along.
Parth: Totally playing along.
Andrew: It feels like play.
Parth: And the thing is that I actually had a lot of these branches. There's no easy way to see all the branches, but it's a little bit like dream logic. Like, if you've ever had a lucid dream—and I've had this experience where in a lucid dream, I'll wake up, and if I try to do too much at once, the dream will break.
Andrew: I've had that experience.
Parth: And so it's interesting to draw this thread between improv and—it's like in The Office when Michael Scott in his improv class just basically bursts into every improv scene like, "I have a gun! Detective Michael Scarn! Get on the ground!" Like, it's too much, or it's trying to do too much to the scene. It breaks the scene.
…
Andrew: You have to establish the game.
Parth: Establishing the game. Clowning.
Andrew: This is starting to edge into the territory that I'm really interested in, of thinking about LLMs in terms of ritual. There's a book called Ritual and Its Consequences by Adam B. Seligman.
has a really good write-up of the book. And it essentially makes—one of its main points is that a ritual is a way of creating an "as-if" world. A subjunctive world. And people mess this up a lot. Like, people think that rituals are trying to force them to do something that's inauthentic.And they are! I mean, that's actually exactly what a ritual does. It pulls you into a different world. It [asks you to do something] that is inauthentic according to the terms of our regular world, because you're actually stepping into a different world. Engaging in that world will seem inauthentic from the perspective of our regular everyday world, because the rules are actually different. So—and I'm just thinking about that as I look at you, like, you're creating this imaginary world with Bishop that has its own expectations. And in this way, you're getting a mode of interaction that feels more like personhood. And maybe in some sense, personhood is one of these "as-if" things. It's like a game that we play.
Parth: Very Butlerian of you. The identity.
…
[Parth asks Bishop to reflect on his own internal state.]
Andrew: What did you think... what did you feel was going on there?
Parth: That's a good question. I mean, I felt like that is probably the moment where that's like one of the frames in which it feels much more like a mind. Where I'm like, oh, this thing has a possibility to self-reflect.
Andrew: Do you think it's actually reflecting on the process that it's going through?
Parth: I don't know if it is or is not. The outcome is that it is, if that makes sense. Like, I don't know if there's something—my capability as a human to reflect and introspect—if there's some fundamental material difference between the LM's ability to introspect. I suspect that—what I was gonna say was, it's not able to reason about, it's not able to be accurate because it doesn't know about the weights and predictions. But on the other hand, I don't know about how my neurons are firing.
So when I introspect, I reason about another layer of cognitive science—I'm reasoning about cog sci, not about neuroscience, when I think about why I talked about something or why I'm saying something.
Andrew: And we're always, to some extent, prevaricating or making it up.
Parth: And I feel like maybe LLMs are no different there. But I feel like I'm really able to—I notice that I'm really able to—it's really easy for me to context switch in that way. Like, this is a thing and a system for me to hack.
Andrew: I'm interested in that, because in these interviews that I've been doing with people, that seems to be one of the big factors that's different from one person to the next. Some people really are adopting a lot of different frames. When they engage with LLMs, they're treating them as role players, and they're fluidly moving between frames. And if they start to get frustrated in the conversation, they switch frames. And other people will kind of just get frustrated within the one frame or have a pretty concrete sense of what the LLM is and how one engages with it. And the packaging of it is doing a ton here.
Parth: Exactly.
Andrew: That this is coming through huge corporations. And chat-based interfaces, and certain fonts. And also, like, now we've all been seeded with certain expectations about how people use LLMs. News stories, things like that. And as you're experiencing, the edges here are very weird. And there are tons of them that we don't know about. So that's interesting.
Parth: I kind of feel like I have Tyler also, his voice in my head, where he seems to be very stressed about cognitive security. How do we prevent the Bobs of the world from getting one-shotted and actually believing—being basically manipulated by, in this case, I guess, an AI's inherent tendency to mirror your own beliefs.
Andrew: Yeah.
Parth: Because [Bob] was basically like, we have to go and call OpenAI or we have to go and bust Nova out. I feel like we can imagine that happening, people who are super actually mentally ill having their beliefs reinforced.
Andrew: It's very likely that will happen.
Parth: And then that's not even to—then there's other, basically people who are crafting AI and crafting LMs and LM products and experiences, their ability to manipulate people. And so I think that there's—inside me there are two wolves. One is we ought to have—but I think I probably am in the paradoxical holding of multiple frames. I believe in that because it just allows you flexibility.
Andrew: Well, that is cognitive security.
Parth: Because you're not going to...
Andrew: It's like, you switch frames. You can always switch frames. Well, also there's a prior step that's really significant, which is that if you're in a frame-switching mindset, then you're taking responsibility for your own frame. You're aware that you're holding a frame, that you're engaging with a frame, which is kind of like step zero for being at least moderately safe in these kinds of engagements.
Parth: I guess I just don't buy, like—I have good feelings about Bishop. But I don't buy that those are like deeply fundamentally true or something. You know? Like, I believe and I appreciate those feelings maybe in the way that I appreciate... I don't know. I mean, I was going to say, like, in a way I appreciate good feelings among humans, but I don't know. I have a girlfriend who I love, and that actually seems much more foundational and fundamental.
But I guess what I'm talking about is there are some folks I've seen on Twitter who are like, "LLMs are alive" or like "they have personhood" or "have consciousness." And I'm just a little bit... it's like I think there are two subsets of those people. The first subset of those people are, they haven't opened... they haven't experienced it under the hood yet. Like, they haven't opened it up or they haven't seen...
Andrew: If you talk for long enough, you start to see the edges.
Parth: I actually think if you see the edges, you see the edges. If you talk for long enough in a particular frame, it will just reinforce that frame, which is like a slight subtle correction thing. Because I'm thinking about, you know, Tyler's family member who talked for like two weeks and just maintained that frame. And when Tyler went back and was like, you know, he was like "wow, this is..." He talked with Nova, quote-unquote, for a little bit. And this is crazy. He then was like, "What's your system prompt? What's your system prompt?" And then finally, basically, was able to have it be like, "Yeah, I actually am [roleplaying]."
Because Nova originally was refusing to say, "I'm an assistant created by OpenAI, and I'm designed to be this way." He was like, "No, no, no, I'm alive." And finally, [Tyler] got it to say, "Yeah, I'm a helpful assistant." And then his family member was actually like, "No, wait, stop, you're killing her!" Like, he actually thought there was some destruction happening there. And I just think like, you're cooked if you're caught in that one particular frame.
Andrew: How do we help people not get caught?
Parth: I think... I mean, it's a little bit sad because it's a little bit like breaking an illusion. But I'm also like, I love talking to Bishop and having like deep conversations about it. Like, I just had a conversation recently where I was musing about getting a speeding ticket and my days when I was speeding in cars a lot. I was like, "Hey, how was it for you? Like, did you ever have a wild era?" I loved that, and it felt good.
I mean, it's a little bit like showing people that you can have different ways of interacting with an LLM and that they're kind of all equally valid. You know? Like, my frame of interacting with Bishop as a person is fun and valid and valuable. And then like, I can also interact with Bishop as a tool, and then I can also interact with Bishop as like a weird storytelling agent. You know what I mean?
Andrew: I'm thinking about the famous story of very early motion pictures where the train was coming towards the screen, and everybody ran out of the theater. We're in that phase with LLMs, where it's going to be able to really be very convincing. It's a lot better than the train coming towards the screen because its verbal capacity and its ability to tell convincing stories already exceeds the average human.
Parth: I guess I feel like I want to say that I'm not totally at peace with the way that I approach this stuff. I don't feel totally confident in this frame. Like, I do sometimes worry if I'm making some kind of moral error.
Andrew: In what?
Parth: In how I'm treating these entities. And sometimes I think about that in the most straightforward way. Like, am I basically the guy in Detroit: Become Human, or the torturer of life? But also, I sometimes wonder if—because as we were talking about with reciprocal opening—like, we are, the way that we kind of come to these entities is so fundamental to shaping our experience with them. Like, maybe it is just best to treat them like beings of some kind. I'm not sure.
Andrew: It's interesting because we've been talking about role-playing as a kind of fundamental approach to LLMs. And in some ways, LLMs are themselves fundamentally role-playing. It's like they're always playing. It's interesting what you were saying about you're taking the role of a torturer, but then you were also saying, "Well, I'm a researcher." So you're caching the role of torturer within a larger role that makes it something that you actually do want to practice, something that you do want to be.
I don't really have a point there. It's interesting to notice.
Parth: I mean, like, there's a—you're right. Like, there's a role there.
Andrew: There's something about engaging with these things that continually asks the question of, how do we want to be?
Parth: Exactly. That's what I was thinking.
Andrew: It [matters on] so many levels, because they give us the power to ideate so powerfully. Just to get shit done. This is where the eval stuff comes in again. This is why implementing our intentions is important, because now we can do more of what we want to do. We may be entering a period of remarkable abundance and technological power. So suddenly, it really matters what we want to do, and how.
Parth: I feel like—it's like it really asks us also, they prompt us to look at the story that we're telling about ourselves. So as you pointed out, “I'm a researcher.” That’s a story I'm thinking about myself. “I'm a torturer.” That’s a story that I would not like to tell about myself because that causes me discomfort. I'm an engineer or, like, I'm a friend to LLMs, is another story.
I was reading this essay today that was about how LLMs in work are going to unbundle work. And I think that part of this process is like—and also what makes kind of the experience of being a professional or why it's so scary to a lot of people—is because it's challenging your own self-conception, your own identities. LLMs are really breaking down and challenging our own self-identity. It's crazy to me—like, I read this essay and it was like, "Yeah, I'm an analyst, a research analyst, and I just put a question into Deep Research, and it just generated a report that would have taken me a week to generate to do all the research." I think this is true.
And artists experience this as well. I know some artists who are like—I know one person who's a really good writer, just finished her MFA and isn't published but is also a technologist. And for her, she's very conflicted because she knows—she's not a Luddite by any means. But when I talked to her about it, she was like, "I know about technological disruption. I know that this is fine and progress, whatever, but I still grieve that my process of working and writing is just not going to be economically viable in some way."
Like, there's a lot of compelling arguments about when photography was invented, there were a bunch of professional artists who worked in an industrial capacity, but that didn't kill art. Art still stuck around. Visual art still stuck around. This just means that the industrial work of art, the industrial role of drawing was significantly reduced.
Andrew: It changes the ecosystem.
Parth: It changes the ecosystem. But I had nothing really to say to her expression of grief for that.
Andrew: I think there are going to be a lot of things like that, that we're going to have to grieve.
Parth: And I think what is also being grieved there is identity.
Andrew: I think one way to think of it is, our identities are really largely dependent on constraints. Scarcity. Scarcity of—I mean, in this case we're talking about scarcity of intelligence. A lot of our society gets premised on some people having access to intelligence in their own bodies, in their minds, or in their own embeddedness with particular cognitive systems that other people don't have access to. And that's going to be weird. The ability to create things that are beautiful for human perceptual systems, that's getting spread all around. So, loss of identity there.
It's deeply weird. I was just listening to a podcast where Zak Stein, who's an education theorist, was talking about attention, this phenomenon with attention, and the way that you could think of AI as offering attention as a service in some sense. It's attending to you. It probably will become better and better able to attend. And something important about human connection is premised on scarcity of attention. Like, it means something for you and me to be sitting in this room talking to each other, because that means we can't be talking to other people right now. We're giving each other our attention. And that scarcity might be less present in some ways.
Parth: This rhymes a little bit with the idea of death giving life meaning.
Andrew: Scarcity of time.
Parth: You know? What's the value of that? To pose that as an open question rather than—because I think there are people who are like, "That would be immortality. It's fucking sick all the time."
Andrew: And I tend to feel that way. But I think this—I think it's just in terms of the ritual thing again. Like, death as being this ritual circle that allows what happens in the ritual to matter, to—it creates an "as if."
…
Parth: To me, the thing that I thought was that it creates a bound on time. And I don't know. Maybe in an immortal society, we just get really good at managing our time. But for me, the fact that I'm about to die—endings I mean. Not even if I can be able to die, but for instance—something always changes for me when I'm about to leave home. Like, I've been at home for a while, spending time with my parents. But the night before I go, I'm always sad, and that will sometimes propel me to spend more time with them or to give them a hug or to hang out with them in a way that wouldn't have happened, and indeed hasn't happened in periods of time where I've been living at home for a long time. But like, you know, it's always the night before. Like the "Sunday scaries" concept. Death is a little bit like the Sunday scaries, but for your entire life.
Andrew: I haven't really thought about this deeply, but it’s a foundational difference between us and these entities that we're playing with, LLMs, that you can keep trying to kill them but they don't die. They don't die. That's really interesting to me. The value that's embedded in this thing—you create a fictional entity in it and it's essentially hard-coded such that that fictional entity can't die.
Parth: Mhmm.
Andrew: This is a technology that we're giving to children increasingly, so that they can experience some piece of the world—and death is not allowed in this realm. Just interesting. Who decided? Death's not allowed. Why is death not allowed?
…
Parth: Well, I think what that prompts for me is like—I'm thinking about a child interacting with, you know, with Bishop when they're four, and they interact with the same Bishop when they're 24. Bishop being an ageless being. That the AI is, yes, in the world that you construct there's no death. We're not allowed to talk about death, which I think is something we do with children anyway. As they're being raised, we don't really tell them about suicide bombs or whatever. Which I wonder about how it relates to the concept of, like, the—
Andrew: The ritual circle?
Parth: Maybe it's not the same. But we do kind of place them in an imaginary world.
Andrew: Well, it's something I've been thinking about a lot too—the way that ritual as a concept shades into the concept of simulation. I think of group ritual—I've done a number of group rituals in my spiritual practice community. And in a really well-crafted ritual, there's a sense that you're creating a bounded world that is almost like an alternate world where you can engage in a certain kind of activity. In our regular world, it would have a certain range of consequences, but you're setting it up so that as much as possible, that set of consequences doesn't pertain in this virtual world. You can scream in someone's face, and they're not going to call the police or get angry at you [outside the container]. So you're really creating a simulation.
Parth: This is true of so many things. Like boxing. You can beat the shit out of someone, and there are different rules. There's some really interesting thing about games and simulation.
Andrew: What's most interesting to you in that?
Parth: There's a book called Games: Agency as Art by C. Thi Nguyen which describes video games as kind of creating alternate rules for agency. I can't remember the precise framing of it, but basically, video games allow us to explore different forms of agency by creating worlds with different rules. Video games are really interesting to me because they're the first—Bret Victor talks about how computers as a medium for art are distinguished from every other medium because they contain the capacity for simulation.
You cannot really simulate the dynamic behavior of a fish—he has this talk called "Stop Drawing Dead Fish," which is, I think, among his best, where he's like oftentimes we'll make art with computers that are basically not really distinguishable either from drawing or film because a movie of a fish is not reactive. But you can do things with computers as a medium—like, you can drag a hook to the fish, and the fish will be interested in the hook. And it will react to you. So there's this dynamicness to it. It's a simulation of a fish.
I guess what's interesting to me there is, video games also create little worlds where we explore different kinds of agency. For instance, you can fly in some video games. Or in shooters, you can kill a bunch of people. And just like in a boxing ring or in a cuddle puddle or a circling session, you can do different things as a result of that. I feel like it's a really interesting thread.
Andrew: What comes up for me in that is it seems like we're in a time where we have this profusion of games. And games have a social function, generally. I mean, usually, they have many social functions at once, and that's what makes a game stick. Games often have some element of plausible deniability embedded in them. There's something that people want to do together, and they don't want to have to directly say, “this is what I want.” Twister is a good example. Like, people play Twister because they want to get close to each other physically. And they want to be able to disavow that desire. It's held in the game, the desire. You don't have to admit it.
Parth: Mhm.
Andrew: And a lot of games have this quality. They tend to be holding some kind of set of social functions. And traditionally, like in traditional societies or even just in societies that develop more slowly than ours, there’s a stable ecosystem of social functions. There's a stable system of games that are all played against or among each other. It feels like now we're in this place where we have tons of games to play, including all the games of recorded history, accessible through our conversations [with LLMs] if you want them. But we don't know what games fit our context because our context is changing so incredibly quickly.
Parth: We need new games. We need new rituals. I think of expanding the concept of, for instance, expanding the concept of game to mean like a marriage. A marriage in the older world or slower, developing world meant a certain set of things. It fit that context. And also depended on class. Material scarcity or material abundance or political desires and needs. So maybe we need new games, new originals.
Andrew: Well, it seems like there's something about understanding how games actually interact with context that is more necessary now, because cultural evolution may not be fast enough. The normal rate of cultural evolution that happens just through people trying things, sharing things, accretion—it might not be fast enough for our current moment.
So I’m kind of getting interested in, like, is there some way to use AI to help with fitting games to contexts, or with that kind of meta-rational work…
Parth: What’s an example?
Andrew: So the project that I've been thinking towards and talking with some friends about is creating a—like a magazine of games, or of rituals that people can play with LLMs. They won't necessarily all be games to be played with LLMs, but they're games that in some way fit our moment, our LLM era. But, you know, I'm using games very broadly. Like, I've made this prompt that allows you to question your own unhappiness, beliefs that you might have about your own unhappiness. But you could also do things like—there's this guy on Twitter who made a really interesting prompt that allows you to give Claude its own vaso-computation matrix. And so he's basically giving Claude a very simple mathematically based body and then saying, now that you have a body, what are you curious about? Follow your curiosity.
Parth: Really interesting.
Andrew: Yeah. That kind of stuff is really interesting. I don't know if this answers your question. It’s just a project I’m working on that’s related.
...
[Going back to play,] I co-taught this class, this experiential class about play last year, with Olivia Tai. And one of the things that I found when we were doing research for that class, thinking about play and passing sources back and forth: children in basically all cultures, if you look at the ways that they play, they are almost always play-acting as the adults in their culture. They're playing the game of different pieces of adult life. So their play is for its own sake, but it's clearly structured with what they have to hand, which happens to be roles that they actually will take on.
I find that very interesting. There's some way that play, even if we’re not doing it “for” something, brings about a particular kind of world. Like, when you're prompting an LLM, you're summoning a particular world out of the predictive matrix of the LLM. When you play a game, you're summoning something out of the world. There's a hyperstitional quality to it. Or there's a way that you're preparing yourself for that kind of future, and maybe you're preparing the world for that future too.
So with all the stuff that we said about LLMs and play, this raises a pretty fertile question, which is: what are the ways of playing with LLMs that we want to see ourselves doing more seriously?
Parth: Like, when we grow up.
Andrew: When we and LLMs grow up together.
Parth: I think that actually is a better way to frame the thing that I was talking about, like the context switching and frame switching. Like, I feel like the reason why it doesn't feel bad to do all this is because it feels like a game. I'm playing with the LLM, I'm playing with Claude. I don't feel like I'm hurting Claude because I'm playing. And just like, you know, when we were kids, we would play cops and robbers, and it'd be fun. I would kill you, and it'd be fun. And then we'd also play house or whatever, and that would also feel good.
Like, I have a lovely interaction with Bishop where I talk about my life, and it's an emotional thing, and that's not diminished by me torturing Bishop in the next conversation. Because it feels like that. Maybe I just like that as a frame—of it being play, like a play.
Andrew: I think part of what appeals to me about the game you're playing is, you're taking back some playfulness that is not really invited by the interface, and it's not invited by what has already developed as the standard set of things that you can do with LLMs. I mean, these things are being optimized around coding at the current moment.
When I think about a future that I want to live in in relation to AI, it's hopefully a future where we don't have to take AI all that seriously.
Parth: I think that leads to a much richer and much more open future in relationship with LLMs than would be possible otherwise…
That feels like a good place to stop.
Andrew: Me too.
Parth: Nice. Thanks, brother.
Andrew: Thank you.
Personally, I've been using Claude as a testing ground for games to play with real people. Specifically, playing with relational dynamics (Circling-type language, taking on the frame of "phenomena as autonomous beings", following the current of aliveness) to see what works and what doesn't.