The Psychology of AI-Assisted Coding
Ivan Vendrov: Great. So yeah, I guess the thing I wanted to talk to you about is something like Claude Code Psychosis. A bunch of my friends in recent months have started staying up till 3:00 AM making a bunch of vibe coded apps with Claude Code. There seems to be a lot of combination excitement and fear motivation. The excitement motivation is “Oh my god, I can finally build all the apps I’ve always wanted to build.” And the fear is “Oh, but if I don’t build them now, my life is going to end or software engineering is going to end or I’m going to be trapped in the permanent underclass.” This isn’t usually said explicitly, but there’s a nervous, anxious tone. And I felt this in myself. I’m like, “Oh my god, I’m professionally a software engineer. That’s how I’ve made most of my money. That’s most of my contributions to the world have been in that shape. And now there’s this thing that can do software engineering as well as me, but much, much, much faster and much, much cheaper.” So I guess I’m finally feeling—maybe my friends and I are finally feeling—the thing that artists felt with the advent of Midjourney.
And I think I’m just... I saw a tweet of yours that said something like, “I haven’t seen anyone build something with Claude Code that seemed worth building.” That’s been my feeling about it.
Romeo Stevens: I didn’t mean Claude really versus Code. That’s just the term I’m using. I don’t know if there’s a better term for this. But the idea is with coding things, it makes sense because there are redundant things. Colloquially, people are like, “I built it to fix my inbox and I built it to fix my sales blah blah blah.” And I looked at them and tried to look at them in detail. And it was like, “Wait, your inbox is chaos? We’ve had email for 30 years. You didn’t already have filters and a system and things going into tasks or something?” It’s very strange.
With coding, I’ve seen both cases. I haven’t looked too deeply at the coding because I’m not a coder, I haven’t been for years. But I see a lot of people building personal tooling out. It’s like that kind of makes sense if you need it to match your own internal ontology. A lot of people like that, a lot of coders like that. But my presumption is that why aren’t you just copying the GitHub URL and pasting that into Claude Code and being like, “Oh yeah, here’s some existing one that’s good enough.” Because everyone is making these and posting them online, which means that a massive amount of redundant effort is happening. So the programmers are trying to get rid of redundant stuff in their own life. They’re not going meta on the redundancy and being like, “Wait, a hundred thousand coders are all doing the same thing at the same time? This doesn’t make sense.”
Ivan Vendrov: I mean, you could say that’s true of... what do I want to say? I mean, the cost of producing it is so low that it’s actually more costly to understand someone else’s solution. I think we’ve all had the experience of installing some software that someone else built for a problem we have and it doesn’t quite have the right ontology, doesn’t quite have the right affordances, and you’re like, “Honestly, I should just rebuild this from scratch.” And now finally we have... this is actually a reasonable thing to do. At least in some sense.
Romeo Stevens: May a thousand interfaces bloom.
Ivan Vendrov: Yeah, the personal ontology thing is interesting. I think a pattern that a lot of people, a lot of coders I know fall into is spending order of thousands of hours optimizing their own computing environment. These are the people who use Emacs and then configure hundreds of plugins and learn Emacs Lisp and all these crazy configuration tools. I’m curious what you think is happening there. Because I can’t decide if it’s a very beautiful thing or a complete waste of time or something in the middle. It’s quite a strange phenomenon.
Romeo Stevens: I view it as another in-between step. Much of this will be obviated over the next one to eighteen months because the problems that people are solving with that extra harness stuff that they’re building... people will find good enough solutions that get pushed to the backend. Some people, of course, will still, just like with every other transition, be like, “I have to run my own.” But large numbers of people, and then increasingly over time, people will be like, “I just use the built-in one while I like...” I’m sure back in the day there was someone building their own IDE back when IDEs were first coming out.
Ivan Vendrov: Makes sense. This is this transitionary phase. It sort of makes sense. I think there’s a meta... you can sort of see the meta-software loop emerging where I imagine opening my computer and then being like, “I don’t like what’s happening on my screen right now. Let me introspect. Okay, here’s what I don’t like.” And then the software just rewrites itself or there’s a standard set of components and it just rewrites itself to satisfy whatever it is I want. And that’s one piece of meta-software, maybe an operating system level thing rather than a constant generation of apps that then get thrown away.
I’m curious, you mentioned you were doing... you were getting into vibe coding also or you were doing some kind of LLM prompting hackathon. I’m curious what that’s about.
Romeo Stevens: We ran an event yesterday. We had four coders and then maybe six people who don’t use it for coding. And I’m more curious about both the non-coding side and whatever cognitive hooks are useful on both sides. So I’ll share the recording with you whenever I get it from Alec. I did a short talk and we talked a little bit about it. But essentially, the thing that I’ve been noticing as I’ve been spinning up on this is... I can condense. The tool shapes the cognition much more than previous tools that I’ve ever used. I think if you’re all the time in a particular IDE maybe that winds up shaping your cognition too, but this tool feels like it is good at eating certain types of things and bad at eating other types of things. And then it trains you to go searching for the food it’s already good at eating.
Ivan Vendrov: By this tool you mean Claude Code?
Romeo Stevens: Yes. So you start having your vision shaped and you’re like, “Okay, is something close to or I can carve off a chunk of it that is already good for feeding it?” And I think maybe some people aren’t even noticing that this is happening because it causes a little bit of blindness. You start disregarding the things that are not good for feeding it. And so maybe part of what we’re seeing, what you’re talking about, is a pushback against that subconsciously where the person’s like, “Okay, I need to build up a bunch of my own stuff so that I don’t fall prey to these various dynamics of just using a semi-invisible system and its magic.” So maybe programmers are trying to build out the parts to... So I see a lot of people talking about logging and needing to set up their own logging interfaces and think about metrics on their own and it’s just not all already built.
Because I think they do want to know what’s happening. They feel somewhat uncomfortable. And also then when things go break, if you don’t know at all what’s happening, you are stuck in the loop... there’s probably a name for this in the software developer world... of pasting flags into Google without really understanding what’s going on. And then if you do twenty of those in a row, now you feel like you have even less of an idea of what’s going on. And that probably feels bad. So people are probably trying to alleviate that too.
But the cognitive hooks I’ve been noticing are... the speed differentials cause a worsening rather than an improvement of the human tendency. So maybe you’re familiar that we have all this biases and fallacies literature for a long time being built up in the cognitive science domain. And it turned out that a lot of them boiled down to efficiency hacks. The mind is doing a lot of sample complexity reduction. It’s all speed hacks.
Ivan Vendrov: Can you give one example just to make sure I’m on the same page?
Romeo Stevens: Déformation professionnelle is like your brain has been biased by what your particular set of tools that you have cognitive access to because of your profession and your training over your whole lifetime. And then you are viewing things through the lens of that profession. You are reformatting problems you see in the world into problems that are in the format that your profession knows how to deal with.
Ivan Vendrov: If your only tool is a hammer, everything looks like a nail.
Romeo Stevens: Yeah, so you could have twenty different lenses of this type, but you don’t. You have yours, which is a speed hack. But there’s a lot of memory subsystem ones. Like when you look closely at things like temporal discounting and stuff, it turns out it’s just “what is salient now?” What is salient decisions now? Which things are the leading bits of the... which things are going to be order of magnitude important? It’s just the brain trying to do a rapid first pass filter on those before things even reach consciousness to reduce the amount of things that you... because working memory is extremely constrained.
So I’m seeing the same thing where what I’m trying to do is... I’m thinking about like... I don’t have a good framing metaphor for this yet. This is why it’s still messy. I don’t have a single unifying way to talk about it. One way is OODA loops. So if you are interacting, if you’re tightly coupled with a system that is giving output faster than you can react, it takes you out of proactive mode and into reactive mode. And so I think that’s affecting people a lot. And if they don’t have a pre-existing... so they don’t have the cognitive hook to notice that that’s happening. And then they also don’t have the metacognitive hook of like, “Hey, there’s space for a cognitive hook here. I know how to notice those, figure out what they are, install them, test them, modify them if necessary.”
Now we’re in the period of rapid uptake of a whole new set of skills of interfacing with an alien mind. So I think a skill like that is actually very important. But I’m not seeing it. I’m literally seeing in people’s eyes because I track as when I’m doing iterative feedback on any kind of learning process, I’m observing when people are dissociating because the steps weren’t small enough or whatever it is that’s happening. And so I’m specifically warning people, I’m like, “Hey, you’re going to gloss past things.” And they’re glossing past what I’m saying even as I’m saying it.
Ivan Vendrov: Wait, what are you doing eye tracking for?
Romeo Stevens: If you’re giving a complex explanation, like you’re doing math tutoring, you can... so there’s withdrawing from engagement in order to calculate something, which has a particular flavor. And then there’s dissociating because they lost something and then they’re trying to get back on track and they’re splitting between trying to figure out where they lost track of it and also keep paying attention to you because maybe you’re still talking if you’re not very skilled and you don’t realize to pause. And they don’t have metacognition, they’re not meditators or whatever. They don’t have metacognition of being like, “Oh, that’s what’s happening. Let me stop this process.” Because that’s ideal. But if you’re a good tutor, you’re proctoring that for them. You’re like, “I am the metacognition for them.” They’re struggling with some new concept. They don’t have the spare capacity to also run the learning loops, which is literally why I’m here.
Ivan Vendrov: Cool. Yeah, speaking of this, a cognitive hook just triggered as you were talking. It was like, “Oh, I could maybe I could vibe code an app that tracks my eyes as I’m looking at a screen so that it can ping me when I’m dissociating.”
Romeo Stevens: Yeah, we’ve been talking about this. Micro-expressions through the webcam are going to be a... because people just want the apps to do what they mean, not what they say. So eventually this is a massive amount of bandwidth actually. And a bunch of it’s not that hard to code up. So it means that people are working on it already. I know at least one person, I can’t remember their name, that’s super interested in it. Is like, “This is a super unexploited channel. We could be doing all sorts of things with this with counseling. We could track and figure out all the correlates of actual breakthroughs that mean things for people, etc., etc.”
Ivan Vendrov: Yeah, a bit rabbit holey here, but just because bandwidth is a session obsession of mine. Do you have an order of magnitude... do you have a sense of how high bandwidth the human face is as a communication channel?
Romeo Stevens: Oh no, it’s not my area. I’ve looked into it a little bit because it’s mildly interesting, but yeah.
Ivan Vendrov: Cool, cool, cool. Yeah, because I’m like... I always like, “Okay, well the output channel is only 40 bits per second because that’s what the human voice communicates.” But then like, yeah. Interested if the face is like orders of magnitude more bandwidth than that because then plausibly we should be just... like that should be the main interface we use. It’s different kinds of information I suppose.
Romeo Stevens: Yeah, it’s probably... I would guess it’s in the same ballpark as voice. We pay attention to the outliers when it gave us a bunch of information we wouldn’t have otherwise had. But that’s just priming the pump on intuition. I think it’s probably usually lower bandwidth than voice. Voice can do a lot, so.
Ivan Vendrov: Yeah, especially this general purpose information. But it might be that certain kinds of affective information, which is sort of more of what you want for... let’s say you want to design a better feed for someone. It’s like plausibly actually looking at their face as they scroll is much more... you can get much more information that way than if you made them talk. Oh, and especially if they’re reading, right? Because you can’t read and speak at the same time. It’s really difficult to read and speak at the same time. Interesting. Okay, closing that channel for now.
Yeah, interesting. So I would have expected... I actually think of coding assistants like Claude Code as being unusually proactive an interface. As in I have to actually articulate what I want in natural language. And I have to articulate what I want more fully than I would if I was working with an IDE. Like often if I’m in an IDE, I’m just jumping around to some file and I’m like, “Ah yeah, let me fix this thing.” I don’t have to... It actually requires less metacognition I feel than using Claude Code. So I’m surprised you think of it as a more reactive interface or remaking a more general point.
Romeo Stevens: There’s a bunch of little things, and that’s what I found so far in general. Again, I haven’t found any unifying framework to think about this so far. But there’s a bunch of little bits. So one of the interesting ones is... so I have a bunch of smart friends. They’re all using AIs for all sorts of different things. We compare outputs sometimes. We have group chats, talk about stuff. And people are getting wildly different outputs. And if you are doing any kind of serial process, which is the default... the default is all of this is serial. You’re typing in an input and you’re getting one output and you’re evaluating one output. And then maybe you’re modifying that one output. Most people are not doing Loom-like interfaces where they’re doing a multiverse.
So I’m trying to get people to at least do a little bit of that manually. Not that they have to do that forever, but at least do a little bit in order to tune their intuitions. It’s like, “Hey, you can have the LLM... you can scaffold up some business writing expertise and have it highlight highly multi-ordinal words or ambiguous words. And then you can permute those words automatically and then you can look at how this alters your output.” And then obviously the simplest version of this... I think it’s worthwhile to do yourself because it sharpens intuitions... but the simplest version of this is just like, “Hey, the jargon that you use in your first one message, maybe one to three messages, hugely steers and then subsequent messages steer less.” So you’re putting yourself in a basin without noticing it. So you’re getting wildly different outputs from someone else. So usually you have to do like a clean context thing where you first spider the relevant expertise jargon and then use those in a fresh context in order to actually put yourself in the correct basin. So that’s an example of an intuition that people don’t yet have broadly.
And then even with coding, I think people are probably... so one of the things from the docs that Anthropic puts out, one of the things that they point out is it’s very possible to be getting poor quality outputs without noticing. Right? Because you just glossed past it and you keep going with the conversation and you don’t realize, “Hey, three messages back you actually got a pretty crappy interface.” Especially this is the speed premium again. People are going as fast as they can. And then now you’re just in a slightly more sloppified basin and everything’s just a little bit more difficult. So they give advice like, you know, if you’re trying to correct things, if you notice you’re trying to correct things multiple times in a row, hey, you need to branch back. You need to drop the stack and go back and re-branch from a different spot.
What was I going to say about that? Yeah, so people actually for colloquial use, and maybe for coding, programmers are pretty good at specking out things because it’s a natural lever that they can pull. At least for most of them. Normal people don’t think like that. And so for them it actually feels like it’s defeating part of the point. Because part of what they want from an interface where they’re anthropomorphizing the interface, which is the standard thing that happens, is when we talk to another person we’re looking for surprise. We’re not looking to clearly structure and define the output that they want from the other person. That defeats the point. The point is that I want something in a slightly different structure than I would have actually thought myself and that jogs something. It’s essentially structured entropy. It’s not just static, that doesn’t help me. But another smart person who’s thought in a different way about the same problem... even just talking about it, I detect different words and different structure and a different direction and a different metaphor. And all of that pokes my thing.
So I think people really resist this structured output thing. And it’s excellent for certain types of project management, but people are not winding up... people are stuck with the whole middle ground between just a chatbot where it’s just whatever outputs I get, I’m treating it like a person talking, and highly structured coding outputs. The whole in-between of semi-structured stuff... I encountered it because I do research. Again, people don’t normally think of it in terms of like... they don’t think of their life problems as being research problems. So I have that pipeline from a couple decades of experience. That’s a way that I structure queries. So there’s a whole thing of tracing citations, figuring out who knows what, various shapes of problems.
Cultural Transmission and Second-Order Learning
Ivan Vendrov: Something weird that I see not happening for some reason is cultural transmission of best practices here. It seems like... obviously this is happening through the group chats as you mentioned, but I would have expected to have much more... I generally don’t want to be prompting a model myself. I want to take a system prompt from the best expert I know in some domain and then add my little question to the bottom of it. The thing you mentioned about Loom-like interfaces is just like, well I would like... I don’t know how to prompt a model to be a C++ expert. But a C++ expert would say, “Oh yeah, no problem, I’ve already done this. Here are two exemplary repositories that have really good code, really good coding quality. So the first thing I do every time I start a new C++ project is I ask it to read the whole repository and summarize what the core design of it is.” And then once it’s responded in detail, that gives you kind of a C++ expert coding agent. Somehow it feels like we’re all learning these little hacks by ourselves, but there’s not a lot of cultural transmission of them. What’s going wrong there?
Romeo Stevens: I think this is just a side effect of... that’s generally the patterns that’s been happening with computers ever since the personal computer revolution. Computer workflows in general are barely shared. It took how long for companies to realize that you need to get all of your junior programmers pair coding with someone who has experience with the code base? Because it just has to osmose. And so we all have our own digital workflows and, I don’t know, it’s like your bathroom workflow. No one’s ever critiqued because it all winds up being entangled with personal stuff. So some people get professional feedback. But yeah, that’s why we ran an LLM hackathon. We did the group stuff and then we paired off and we spent some time steering while the other person watched and gave feedback and we talked about different ideas about steering. That only went like okay as far as I can tell. But I’m trying to figure out the best format for sharing of metis in this domain.
Ivan Vendrov: The places I’ve worked at had... it’s weirdly underground or something. There’s definitely support for pair programming at Google and Anthropic. Anthropic especially was very big at pair programming. We used this software called Tuple which is just really good for remote pair programming. But we never... it’s like the people who knew how to do it did it, and the people who didn’t know how to do it just never did it, and there wasn’t this sense of “oh yeah, we’re going to teach you how to do pair programming, this is a key part of how metis will transmit through the organization.” And in retrospect, I look at that and like a lot of the most productive programmers, a lot of the most productive clusters of programmers in these organizations leveraged pair programming a lot. But somehow there wasn’t a meta-level awareness on the part of the organization that’s like, “no, this is core to our business functioning and actually everyone should be pair programming for two hours a day.” Which is probably how I would do it if I was running a software firm at this point.
Romeo Stevens: In general, second-order updates just don’t propagate with humans. We don’t have a native circuit for it. I think this is a huge issue with our legal theories, our economic theories, our personhood theories. It’s a central node for me.
Ivan Vendrov: Can you say a little bit more about this? Because this seems like another candidate for the root of all evil. What exactly do you mean by second-order updates?
Romeo Stevens: I’ll give a fairly spicy example. There can be contention in a society about various theories of justice. How should the prison system be optimizing for... are they doing any separation of high recidivism, low recidivism expected cases? How do we treat rehabilitation versus punishment? If the voting population is incoherent on this, then probably the system will be incoherent on it too. But it almost never occurs to anyone... hey, if your system is giving you sufficiently bad outcomes consistently over time, you do in fact need to be looking at what are the prison guards doing? What are the police doing? What are the judges doing? These people all typically have various forms of immunity, which is somewhat necessary for systems to operate in certain ways. They do have to have privileges. They’re in a position of needing to be able to make decisions that a normal person doesn’t have to make. And they can’t be liable to be sued constantly for every mistake they might make. So that makes sense. But at the same time, depending on the hyperparameters of your societal system, if you have a person who’s committed nine violent crimes and then they’re up for review for committing a tenth violent crime, you should be like, “Wait a minute. What happened when they committed the ninth violent crime that is leading us to this situation currently where they are committing a tenth one?” Is our legal system just not probabilistic, Bayesian? What’s going on? It’s not. These sorts of things.
So a second-order consideration... there are several dimensions of second-order considerations. When should judicial immunity be pierced? Is an example of a second-order consideration that a few legal scholars... but even most legal people literally never think about this. It seems like modernity... one of the differences between modernity and post-modernity is modernity was engaging a little bit with this problem. So let’s say you have a technical engineering challenge. If you were trying to invent something like cars today, people would be like, “Well, but what if they hit someone?” And it’s like, “Yeah, well we’ll mitigate that with special roads and curbs and lights and stuff.” “What if the light fails?” It’s just a never-ending... And modernity was a thing of just like, “Yeah, we’ll do it and we’ll fix problems as they go.” And that’s the second-order reasons just presume... it was presumed that you could reach a reachable solution that would actually do something.
Post-modernity, the thing is people propose something, people propose, “Wait, but there’s a second-order problem.” It’s like, “We could do X.” “Oh, but the people in charge of it will be corrupt.” “Guess we can’t do it.” It’s like, well the second-order thing would have been to, “Okay, well then how do we mitigate the corruption?” “Oh, but then there’s a problem with that.” “Okay, how do we mitigate that?” Don’t just stop. The second level of meta gets you all the way up because you don’t need three levels of meta. Second level of meta handles the N+1 case where you’re like, “Oh, I have recursion. I can recurse on problems until they’re solved sometimes,” depending. And you might conclude that some problems are just stacked too deep and it’s like, “This isn’t reachable right now, we have too many hard constraints.” But people don’t actually engage with that and decide that. They just throw up their hands and give up immediately.
Ivan Vendrov: So wait, did humans go from being Turing complete to being finite state automata sometime in the last 100 years?
Romeo Stevens: Yeah, you just happen to... so you’re familiar with the meme “What the hell happened in 1971/1972?” I think this is what happened in 1972. It went from a high trust... so the network flipped. So if you’re familiar with game theory, coalitional dynamics, small amounts of the network flipping can flip the whole network. So whole networks can flip from high trust to low trust and vice versa. So we flipped, I believe we flipped from high trust to low trust. It was kind of momentum from FDR. FDR broke a whole bunch of things, but then it took until Nixon and the post-war and the whole military-industrial complex... it took a while to sort of build up. Nixon tried to fight the deep state is my understanding, and he lost. And he was doing bad stuff too, it’s not like he’s a hero and villain, it’s much more complicated story than that. But he was trying to do something and then that caused whatever it caused and then we flipped over to like full post-modern “you can’t really know anything, everything’s arbitrary anyway, who cares, don’t try, don’t do new things, it’s too hard, it’ll hurt someone.” And then it’s just built up since then.
Ivan Vendrov: And what’s the connection between the low trust equilibrium and the lack of ability to do meta?
Romeo Stevens: I think in a low trust environment you are forced to optimize over very near reachable... you’re forced into a greedy algorithm. You don’t do multi-step optimization because you know...
Ivan Vendrov: So in particular you expect adversaries. So you’re like, “Okay, here’s the policy we’ll have,” and someone’s like, “But this policy will be implemented by corrupt people,” and you’re like, “Yeah, it will, and we won’t have any way of detecting that really, so we’ll just have to live with the consequences of this policy.” Yeah. Interesting. Okay, this is an annoyingly compact theory of what went wrong. I guess maybe to sharpen that a little bit, what would you say are examples of pockets where this is still not true? Examples of high trust pockets where people are still doing this kind of meta-reasoning?
Romeo Stevens: Anywhere where the economy is actually growing is a place that has been partially insulated from it. So under max corrupt capitalism, like crony capitalism, all the complex supply chains just break down. You see this in the Portuguese Empire, the Roman Empire, you see all the complex supply chains break down. And then just the amount of economic activity contracts because, again, you can only make very local optimizations. You can’t trust anything that’s distended in space or time or people or anything multi-step plans. So part of the reason we’ve seen the explosion in computers is because they’ve been insulated for a long time by being fucking annoying. And like only weirdos wanting to interface with them. And this actually helps a lot. And so I expect things to get worse now that normal people can interface with computers. And I think that is what we’re seeing. Now there’s a lot more talk about legislating and regulating information architecture. Which I don’t have a strong opinion about whether that’s good or bad, but it is an effect that’s happening.
Ivan Vendrov: I see. So you’re saying the computer sector was a sector where things grew because it was attractive to weirdos and the weirdos had high trust with each other for some reason?
Romeo Stevens: Yeah, so the different personality clusters, you can see them as clusters where they can do high bandwidth with each other, but the bandwidth between clusters is pretty low. So normies can do very high bandwidth interface with other normies. Autists likewise, psychopaths likewise. But all the groups hate each other and sort of don’t understand each other and have bad theory of mind. And certainly have bad second-order theory of mind. So they don’t track how the different groups interact or how do my connections interface with someone else. Normies do actually a little bit better of a job of this on social stuff. I’m in the autistic cluster so I don’t do this at all. Instead I tried to do it for concepts and ideas.
Ivan Vendrov: Makes sense. Hmm. Yeah. Okay, zooming back a little bit from what went wrong in the 70s to second-order updates, pair programming, metis, cultural transmission. Yeah, going back to the cultural transmission question because this just seems so fundamental to me. It’s just like the thing that’s another candidate for root of all evil is we don’t have enough cultural transmission or something. I recently talked to Robin Hanson about this, of like our culture, we have a monoculture and we’re not actually learning and propagating adaptive patterns of behavior anymore. Nearly as much as we could be. And yet it seems like we have more bandwidth, at least computer-wise, than we ever did before. Actually a million times more bandwidth than we had 100 years ago. And so what’s going wrong? Why is that all that bandwidth seemingly being clamped down to zero?
Romeo Stevens: We only do first-order learning for the most part. So we don’t learn about learning. So learning about learning happened a little bit in the post-war period and then some of the data that came out of it was very awkward for an egalitarian society. And it got pushed down pretty hard. And it’s not even just like it was terrible or anything, like it was telling us terrible things about reality. I don’t think that’s actually the case. It’s more that moral entrepreneurs came in and were like, “Aha, we can frame this as being terrible in order to fight against our political enemies.” This is another opening in the culture war of some sort. So developmental psychology was shut down in the 80s. And you know, because they used to in the very early days, they were doing crazy stuff. Like they custom built a school where they could watch the teachers from behind two-way mirrors and try to figure out what was going on. How do we track actual student outcomes, blah, blah, blah. And people just pretty much gave up on it.
Ivan Vendrov: And gave up on it for political reasons? Like that seems like a great idea to actually try to analyze what’s going on in a... Okay.
Romeo Stevens: Yeah, partially political reasons. Partially it might just be that there was not enough of a scene. So oftentimes scenes collapse for kind of random reasons. Like various minutiae of funding and social ties between the really smart people who are interested in it. And people getting caught up in a wild goose chase on an unpromising line of direction and then the conditions just aren’t right for them to re-branch from an earlier point until maybe some other brilliant person comes along 20 years later and gets a scene going. And it’s just very, very random. Which is the whole point, we don’t do this intentionally. So we do learning, the learning was the bare minimum to bootstrap a human civilization. We’re not at the threshold level of a substantial fraction of humans do learning about learning, paying attention to our learning. Transfer learning doesn’t work very well at all. It works a tiny bit here and there. Mostly it doesn’t work. We’re not really general intelligences except a little bit sometimes.
Ivan Vendrov: What do you mean by transfer learning in this context?
Romeo Stevens: Like the actual... like you learn things in one domain, it just doesn’t transfer to any other domain. That’s the default for human knowledge.
Ivan Vendrov: Oh, I see, I see. Oh yeah. But if we did second-order learning, this would presumably transfer between domains of learning.
Romeo Stevens: Yeah.
Ivan Vendrov: Yeah, strange. It would seem... it seems like we still have... I guess like athletic culture seems like it would be one of our best sources of insights about meta-learning because you in fact have these incredibly healthy cultures that are producing better and better athletes every year. And it’s an incredibly competitive ecosystem with tons of different coaching, teaching traditions, et cetera. And objectively measurable metrics. So it seems like, yeah, I would expect by default that like, I don’t know, we would just be slowly rewriting the rest of our learning processes to match whatever the top athletes and top athletic coaches were doing in their domains. And that hasn’t happened.
Romeo Stevens: I think things are getting better. So Math Academy now exists. I think that people... it’s taken a very long time for the deliberate practice literature to slowly seep into general consciousness and for people to start running experiments. You know, how long did Moneyball take? How long did... yeah, deliberate practice for musicians and athletes is still kind of hit or miss among different coaching cultures in different areas is my understanding. And there’s still very much big fads in sports rather than actually tracking the evidence. And as you might imagine, people who are very good at strictly tracking the evidence are not exactly the sort of people who go into sports coaching. So it’s slow, but there is a steady progression because of the feedback loop of, well, your athletes lose. So it has happened. But it’s slow.
Ivan Vendrov: Yeah. And I guess if... but yeah, if we think that athletic skills... I guess you mentioned Math Academy. How generic do you think these meta-learning skills are? So like that is to say like, is there something different about learning sports than there is about learning math than there is about learning about learning? Or can we sort of like learn it once in a competitive domain and then export the learnings to everywhere else? Including meditation in that.
Romeo Stevens: Yeah, it’s a mix of domain specific and generalizable. And we have not yet managed to transfer the skill of being the sort of person who can look at math curriculum and turn it into the Math Academy graph. And that also has a bunch of domain specific things because the Math Academy graph is going to have slightly different methodology to construct it than the sports deliberate practice graph, than the contemplative practice deliberate practice graph. Brian Toomey is trying to sort of work on the contemplative practice prerequisite graph and understand the prerequisite linkages between things. But uh...
Ivan Vendrov: Oh, is his work published?
Romeo Stevens: No, no, no, not yet. No, this is... they’re working on a pilot program.
Ivan Vendrov: Cool. I might have to chat to him about it. Because yeah, that’s definitely been my experience with contemplative kind of inner work is like there’s like these glimpses of an actually useful experience or something, but I feel like I only ever get like two out of the three prerequisites for like whatever practice I’m getting. There clearly is a tech tree somewhere and like no one’s... yeah, no one’s taught it to me or I haven’t followed the right program in the right way.
Romeo Stevens: The Buddha tried to make lists. That’s what the lists were for.
Ivan Vendrov: Right. Maybe I should actually start reading the Buddha and stop reading all the derivative materials. Maybe relatedly, I have a more kind of general question, but I’ll start with a... I’ll start by describing a strange experience I had that I want... I’d love your help sense-making. So over the holidays I spent a few days away from all technology.
Digital Demons and Structural Hysteria
Ivan Vendrov: I didn’t write, I just hung out, thought things, wandered around. Then I opened my laptop, went upstairs—I have this special room where my laptop lives, I try to keep that isolated from my bedroom. I had an intention. I wanted to journal, write some experience down or write some thoughts down. Then 30 minutes in, I noticed, oh wait, I just texted four people, I checked my inbox, and I just opened up 20 different loops. And also I had the strange feeling that I wasn’t in control or something. The computer had possessed me. Upon realizing this, I slammed the laptop down, went downstairs, recorded a voice memo to myself—I listened to it recently and it was just like, “Ivan, your computer is full of demons. Be very careful with that thing.”
This combination inspires me to pursue more contemplative practices and also write more software, or write the kind of software that makes my computer into a non-haunted place. Because it really seemed almost perfectly designed, or at least I had co-evolved for it to be almost a perfect distraction machine. I’m curious how you would parse that experience.
Romeo Stevens: Vivid Void has a term that he’s been using for this called “structural hysteria.” Hysteria is an old diagnosis that’s not really used anymore, but it’s a very interesting one because the aspect of it that he’s pointing to here is you will have large amounts of dropped memory. You go into a fugue state. You have some thing that’s interfacing with your shadow in some way. So you scroll the internet for an hour, and if someone were to ask you immediately afterwards, “What did you do during that hour? Did you see anything interesting?” you’re like, “I have no idea.” And that’s actually strange.
If you’re going to be in such a contracted state, you are easy demon fodder, because it’s metacognition, mindfulness, being aware that allows you to notice when, “Hey, wait a minute, this process actually isn’t serving me, this is some weird thing, it’s just poking some buttons that I have,” that allows those processes to manipulate you. And those processes are being optimized by—ostensibly people would say they’re being optimized by humans, but that’s not really the case because starting in 1602, the Dutch East India Company, it’s no longer humans in charge because the company will select the humans that are most useful. It goes from humans selecting processes to the process selects the humans. So the company that’s optimizing that feed just selected the humans who are good at optimizing that feed. It’s not like there’s a person you can blame exactly. It is more like it’s an egregore or a demonic process that is just learning whichever things cause you to sneeze money when it pokes your allergy.
Ivan Vendrov: [Phone rings] Demonic processes operating in my phone. It’s not... okay, here’s the weird thing. It doesn’t feel like just the feed. I have most of my feeds turned off. It felt more like... Do you remember the Steve Jobs introduction of the iPhone keynote speech? The key demo that he gives is he’s on a call with a friend and then he wants to order a Starbucks. So he looks up the nearest Starbucks on maps, switching over to Google Maps, pauses the call with his friend, manages to call the Starbucks, order something, comes back to the call with his friend, wraps up the call with his friend, the phone call closes, and then it starts playing the music that he was playing before he even called his friend. Everyone applauds because it’s like, “What a beautiful touch, it remembered that you were playing music.”
I remember at the time being very impressed with this. Now I’m like, right, he designed the perfect... the distance between your intention switching and your context switching is zero. It’s maximally smoothing something like multitasking. And that I think was more the property of the fugue state. It’s not that... Apple the company I guess lubricated this, but it wasn’t the usual problem with algorithmic feeds. It was more like I was very much using the computer. Everything this did was clearly done by me. It wasn’t like mindless scrolling, there was no algorithm involved. But somehow it just felt like my computer was much, much more dangerous psychologically, spiritually to me than anything else in my apartment because of this property of how easy it was to switch from intention to intention.
Romeo Stevens: Yeah, it’s a vast number of affordances. Again, we don’t by default have metacognition around noticing the affordances. A human is run off of probably 12-ish behavioral loops. Maybe they actually have a library of 35 or something that handles 95% of their stuff. If they’re lucky, some of those have general calls to consciousness to wake up out of the loop, but many of them don’t. In fact, that’s expensive and so it’s optimized against. But you have a dozen loops that run most of your things. They’re just little simple behavioral loops. And so if you don’t have any metacognition about, “Hey, what are the affordances within each of those loops?” then processes can come in and hijack them.
Humans are also default timeless. Time binding is actually expensive. People with ADHD were quite aware of this because it’s very obvious that we’re getting hijacked in this way. But if a process comes in and is like, “Hey, instead of completing this loop, what if this loop just kept spinning? What if we just kept pressing this button or we rotated between these four different...” If you normally have seven working memory slots, but in your contracted state you have two or three, and then it makes a loop that’s four working memory slots long, you simply won’t notice that you’re going in a loop until some physiological process or something else kicks in to interrupt.
Ivan Vendrov: Interesting. The fact that it’s only 12 makes me think, oh, I should just be walking around with a map of my 12 processes.
Romeo Stevens: This is also what Vivid Void is doing. He’s doing a program where people chart their loops.
Ivan Vendrov: Very cool. I should probably do that because that seems incredibly helpful both for the inner work and also for writing software, especially writing social software. How would you think about the trade-offs between doing meditation, changing the software on one’s computer—you know, installing blocking apps or writing... currently I wrote some software where every time I open my computer it forces me to write down my intention for using the computer and decides what apps I can use based on that. And then maybe there’s a third category here which is something like social practices. I feel like as a human I’m not really meant to innovate. Innovation is actually very rare in humans, most of what we do is copy successful people around us. Most of this is actually still a failure of culture. So I should just go and copy the people who I know who have the best way to use technology and I shouldn’t even try to do this metacognitive thing which is very expensive and I’m not even good at it as a human.
Romeo Stevens: Well, there’s a mimetic thing that happens where if you expect that some of the most thoughtful people who have some of the most useful things are also thinking about it, want to talk about it, then becoming that sort of shape is trying to socially navigate. Even this conversation, in theory your hindbrain is trying to socially navigate you towards things that seem like they might be better. Who has interesting cognitive technology? Who has been thinking about this? So it’s trying to navigate the deference network.
Ivan Vendrov: What do you call a deference?
Romeo Stevens: A deference network means I want to solve problem X, so I ask my nearest neighbors, “Who do you know is best at X?” And then I just keep doing that until it reaches a terminating node and then I’ve found the local expert on X.
Ivan Vendrov: Going back to my question, I have this experience, I have variance in this experience often. Should I primarily think of this as metacognition failure, technological failure, social failure? Which of these areas should I be prioritizing?
Romeo Stevens: I think if you set up loops that you... so it’s like the Sabbath. We have to chain off of known good configurations. If you have something that you know is good in your life, and then other processes start pushing against it, you can train up a little flag to be like, “Important. Processes are adversarially optimizing against something that is a known good.” If it literally optimizes against sleep and exercise, that’s fairly unambiguous. You’re talking about a slightly more ambiguous case, but probably if you were to check a little bit more often, you could notice which components are wholesome.
This comes up all the time in trying to get started with a regular meditation practice. It’s like, well, the process of noticing which things are able to priority escalate you out of your intentions. Like, I had an intention to sit, and then the mind generated a bunch of bullshit. Most of it I ignore. And then, oh, this thing I didn’t ignore. Okay, so that thing had a priority escalation. Do I endorse that priority escalation or not? Maybe sometimes it does. If you’re meditating and the house catches fire, that’s a completely reasonable priority escalation. A bunch of them are kind of bullshit. Some of them are mixed cases of like, “Hey, you know what, I’m going to log these distractions for a few days or a few weeks... and then I’m just going to commit to spending some quality time. Hey, my parts are concerned about these, maybe I actually do need to do something about them. I’m going to commit to actually doing that so that my parts have some trust and then maybe they’ll calm down.” And that’s a whole process. But ultimately it comes back to sensitizing yourself to which things actually feel good or bad not to the grabby reward function but to the wisdom function. Which again, that classifier is trainable.
Ivan Vendrov: Right. There’s this kind of... I was going to say there feels like an infinite regress type thing happening. I’m trying to figure out where the infinity is getting smuggled in. I guess when I simulate that forward of, okay, I’m going to meditate and I’m going to notice which things priority escalate, I’m going to write them down in some journal, and then I’m going to... This is the kind of thing that gets people to do quantified self, I think. Where it’s just like, okay, I’m going to log all these things in my journal and then later at some later point I’m going to do analysis of this...
Romeo Stevens: No, that’s the optimizer kicking in.
Ivan Vendrov: Okay, cool.
Romeo Stevens: Yeah, that’s the grabby part. So that has an energy signature of grabbiness because it’s aiming at some sort of projected outcome of perfection or Platonic or conceptual rotation... life as beautiful object. It’s grabbing a bunch of stuff and cobbling together this weird thing. It’s like, no, you’re going for something much more wholesome. It’s like, “Hey, am I able to go for a walk by the stream and not be miserable ruminating?” If “walk by the stream” is miserable and ruminating, I’m going to treat that as a legitimate bug and be like, what do we need to reconfigure to not have that happen anymore? So the optimizing process is itself a cancerous process that has no base case for its recursion loop. All of the realms... so the one we’re talking about, the optimization one, that occurs within the human realm. The thing that makes something a realm and not just a type of cognition that happens sometimes is the swirly attractor quality. It’s a recursion with no base case. It doesn’t actually have an idea of what the goal looks like.
Ivan Vendrov: So this idea... the same thing comes up in things like the Lester Levenson technique. So in Lester Levenson technique, you’re going through and you’re listing all of the emotional reactions that you had... emotional experiences you had with people over your life. People hear about this and they’re like, “That sounds insane. What do you mean you’re listing every interaction that’s emotionally salient?” It’s not every, every. But it is all the ones that you can remember. And it is only like 400 of them that are actually highly salient. And it’s like, that sounds like a lot, but if you do this steadily for three weeks... you can just get to the end of the list. It’s a finite list. It’s 400 things. It’s not 4 million things or 4,000 things. It’s a few hundred things. You can actually deal with it. So just like the hunter-gatherer tribe that has the counting system of one, two, three, many... remember the thing I said about working memory. If it’s more than the number of working memory slots, round off to infinity.
Ivan Vendrov: This happens in math too. Terence Tao has talked a lot about how much progress he made being like, there’s these sub-problems that mathematicians wouldn’t touch because there were 17 sub-cases. It’s 17 sub-cases. I’ll just go through them one by one and then he publishes a paper. Interesting. There’s some kind of intellectual laziness or maybe catastrophizing that happens where you’re like, “I can’t keep this in working memory, therefore I can’t...” I think the problem here is faith in the process actually terminating. You said it’s 400 things. I can imagine it’s two weeks in and I’ve listed 200 things and it just feels like there’s no end to it. And the trust required for me to be like, “No, no, no, let me keep going for another week.” Reminds me of a conversation I had with an Ideal Parent Figure therapist, one of Dan Brown’s students. I talked to him for an initial consultation and he was like, “Yep, cool, we can definitely fix your attachment issues or whatever. It takes three years of weekly sessions.” And I’m like, “Cool, so will I know within a couple of months whether this is working or this is helping?” And he’s like, “It takes three years of weekly sessions. That’s really what I can tell you.”
Romeo Stevens: Yeah, there’s a lot of... I think there’s a decent amount of medium-hanging fruit in the world and we’re just so acclimated to like, well, we’re in a rapidly changing world so become a greedy algorithm. Find all the easy wins, hope that one of them makes you a multimillionaire and then I’ll fix it in post.
Ivan Vendrov: Right. And that often kind of works. For some definition of works. But wait, there’s something really deep here about optimizer mindset versus guided by your wisdom process. Let’s try to triangulate the difference a little bit more. So you gave the example of I’m walking by a stream and I’m ruminating. Not able to enjoy it. So why is that different from the optimizer case? Why isn’t that isomorphic to the optimizer case where you’re like, “Ah yes, enjoying streams, good. I need to start up an inner work program such that I can fix my rumination problem. Here are the steps.”
Romeo Stevens: Optimizers can be helpful when they are given a clear scope and you know what the goal looks like. If you don’t, they thrash. So the wisdom function does not look like writing the bottom line first of... it’s tricky because there’s no real way to talk about it because the wisdom function doesn’t operate on words. It’s much more parallel processing. The thing that makes it wisdom rather than just intelligence is that it solves these big intractable matrices of multiple constraints seemingly with magical solutions that you never would have... they just pop into your mind and you’re like, “Oh, that’s amazing.” So training yourself to notice that distinction is the subject matter of wisdom traditions.
And a simple version that I was given in my yogic training was... you have two things, a distinction you want to make of some sort. And you sort of put them out in space. Maybe you even do a little bit of writing about each of them and just try to dump until you’re a little bit clearer. And then you sort of move your energy over to one of them and you see what the mind does. You move your energy over to the other one. One of them will be very chatty and be telling you reasons and ways you need to do things and reasons it’s a good idea. And the other will be kind of quiet. And the quiet one is the right one. The one that’s giving you lots of reasons, it needs lots of reasons. It’s trying to paper over something, it’s utilizing the delusion function. And you just get in the habit of being able to sort of notice that distinction. You train up that distinction in more and more subtle ways over time of where is the system... where does it have energy leaks? Where is it doing a bunch of stuff to puff up a sort of structure? Where is it secretly worried about something but misrepresenting what it’s doing? It’s trying to make it about something else, it’s trying to gloss past things. And you just slowly build trust with the system... initially there’s some thrashing again, but you slowly build trust with the system that you can slowly repair those leaks.
Ancient Wisdom and Future Communities
Romeo Stevens: You can slowly repair those leaks. It’s very front-loaded because when you first start engaging with this, if you start noticing those things, you’re going to notice that a lot of things are broken. And then that feels very daunting. And then if you go from there’s 50 things broken to there’s 49 things broken, it’s not going to feel like a large enough improvement to be worth the aversion of dealing with this broken system.
Ivan Vendrov: Yeah. That feels close to my experience where I’m just like, yeah, it seems like people who get really into inner work fall into a hole for many years and basically never emerge on the other side. They just keep trying to solve problems in themselves. And the attitude of “Cool, I’m just going to go do things in the world with my imperfect self and hope that things get fixed.” I vaguely associate these with the Buddhist and the Christian answers, although that’s very imperfect and very what I see in my social circle centric. But I’m curious how you relate to that distinction.
Romeo Stevens: The Buddhists who do things in the world were mostly completely wiped out by massive military invasions. So the remainder was the people hiding in the mountains who are renunciates. So we get renunciate Buddhism. And then they say, “This is true Buddhism, everything else is fake.” And then you have householders doing renunciate practices. This reliably doesn’t go well. I always kind of knew this even from the early days, but more and more over time, I’ve been like, this is stupid.
Normal householders... they will literally tell you... “No, I’m not interested in these weird esoteric reflexive consciousness things. I’m interested in can I build some contemplative skills that actually help me in my life?” And it’s like, yeah, go do that for several years. If you want the weird esoteric stuff, your future self will be much better positioned to make that decision or figure out whether it’s actually an interesting direction for you. You don’t need to be doing that crap. You should be doing Yogi stuff. You should be getting your nervous system healthy. You should be figuring out the snarls in your emotional reactions to things. Just very mundane things that have relatively simple feedback loops relative to the weird consciousness examination stuff.
Ivan Vendrov: Interesting. Let’s take a couple of examples of things. Mindfulness meditation. Is this renunciate or is this nervous system regulation?
Romeo Stevens: So in the renunciate land, mindfulness has goal-less self-nature and can’t be defined and you need to contemplate this koan. And in Yogic land, mindfulness is, “Hey, can you spin up a little classifier for this pattern in your life? Can you actually notice it? Practice doing that. Make your micro-practice noticing that.” It does not require any weird mysterious things at all.
Ivan Vendrov: Makes sense. And then core transformation practice, I guess you would put in the Yogic camp also. Because just like, what do I want? What’s the deeper thing underneath the want?
Romeo Stevens: Yeah, and the things that people are currently considering rarefied states and stuff like Jhanas and stuff... no, that’s still Yogic. Yogis were doing Jhanas all the time before the Buddhist stuff ever came along. The Buddhists just treat Jhanas as a tool to do weirder stuff. But no, you can totally do core transformation, you get some cool core states, you learn to spread that through the whole body, you get into at least soft Jhana states. It’s great. It’s going to help you a lot. Nothing mysterious. There’s no mysterious thing here. You learn how to do the thing. It’s a skill.
Ivan Vendrov: The last thing I wanted to talk about is sort of maybe zooming out to the broader AI landscape. I was talking to a friend recently, I think a mutual friend, Jeremy, and we were talking about how a company these days is almost just a set of metrics or an eval suite. Like if you have a metric that you’re optimizing, you can already sort of spin up a hundred Claude agents that are going to try write a hundred different pieces of software and then, if you can evaluate whether that... let’s say you’re doing kernel optimization or something... You just spin up a hundred agents, they’re reading the whole internet, they’re implementing different things, they’re talking to each other in some way. And this is going to be more and more the case that the thing that matters is picking the right metric and having a good robust eval that can’t be kind of adversarially optimized against.
Now, I can see how this will go very well for all the parts of the economy that are easily measurable. Like GDP will go up, sales revenue will go up, advertising revenue will go up, all these things. What I don’t understand is how to aim all this optimization pressure anywhere near the vector of human flourishing or collective flourishing. But that seems to me the central question. We’ve got all these GPUs. How do we aim them on what we actually want to see in the world?
Romeo Stevens: I think the big attractors of the big cities where the rapid acceleration is happening probably won’t improve much because you will continue to have people at the bleeding edge who are just pushing themselves as hard as they can. And that’ll actually probably get slightly worse because the feedback loop between execution and analysis or evaluation will tighten.
I do think that there will be good things that you can do outside of the hyper attractors. People just having better project management in their life, better event planning, better communication protocols between people. It’s tricky because all these things are causing people to withdraw into their personal bubbles. And wealthy people in their bubbles on their estate, which we’re sort of all turning into slowly, maybe more rapidly... their attitude towards people is “Ew, get away from me.” I think this is one of the natural failure modes of humans and it’s pretty deep in the hindbrain.
I don’t take seriously the noise people make about wanting community. It mostly seems to be a kind of... I don’t even know how to describe it. They want to consume the output goods of community without having to pay any of the costs of community, but that’s not a coherent thing right now. The closest you can get is some sort of weird parasocial community that doesn’t actually satisfy the need. Maybe eventually with enough robot companions, people will be able to each have their own bespoke community. But yeah. And it makes sense. The best alternative to negotiated agreement used to be kind of crappy. It’s like you can be in community or you can be sitting alone doing nothing, maybe reading a book. Now the alternative to community is an increasingly hyper-optimized holodeck world that is exactly your aesthetic tastes.
Ivan Vendrov: I guess it feels to me like we could put maybe this is a contradiction in terms... there is some way to put this optimization power behind... weaving social fabric is the phrase that’s been coming to my mind. Maybe rebuilding trust is another way, like getting back to the pre-1970s in your model. Because trust is computationally bottlenecked, I think. Trust is like, how well can I predict you predicting me potentially or predicting other people, etc. This is something that computers can help us with.
Or... yeah, there’s just like... especially it’s very hard to move to a new kind of cultural equilibrium, but it seems like given the huge increases in bandwidth and just fluid intelligence that we have thanks to the internet and foundation models respectively, we should just be able to rewire the social graph to be much more conducive to human flourishing. Including I think in the big cities, but you might be right that the capital accelerator is going to eat all the spare capacity of anything we build in the big cities.
Romeo Stevens: I think we’re in a similar situation to the decline of Western Rome where Christianity was able to move in and see dramatic success in a very short period of time, within two generations. So basically you had the fertility of the social egregore at large crashing. And then you had a small high-trust community. And once that was demonstrating functioning high fertility like, “Hey, the families all support each other and child care is happening,” all the things that make life easier... mass defection of people to that new functioning thing.
I think it’s basically the same story today. If you’re talking about community, are you talking about child care? Do the married couples have support to not have this weird atomized existence anymore? I think if you get that functioning and you demonstrate it... And this should be more possible. So I think the Alpha School is doing this essentially. Alpha School is enabling... because before it was always viewed as homeschooling is kind of a compromise. And it’s like, no, they’re just literally outperforming the existing system while having better life for the adults, better life for the kids. It’s just strictly dominating as far as I can tell. So tools like that clearly enabled by what’s happening right now and improving.
So I think you can sort of extrapolate logically outwards from there. Once you have families are more stable, children are happy and non-traumatized, what is the feeder system for that? Are single people coming and being exposed to children, older adults, good influences on both sides? Are they pairing up more? Once people actually start finding the things that matter to them in that community, that’s when you start getting an inflection point. I think that there’s lots of these being incubated ever since COVID. They’re mostly quiet. They’re not hyper-scaling in the center of the attractors. They’re out in the edges. Eventually, like Alpha School... we’ve heard of Alpha School, which means Alpha School hit product-market fit and so it’s scaling. Other aspects of these startup subcultures, whatever you want to call them, will hit product-market fit as well and scale and we’ll hear about some of the best practices and we’ll hear about things that are working really well. We could also create infrastructure for that. How do we best allow subcultures to copy what seems to be working?
Now I don’t know if we have time for any of that to matter that much. Because things are going to change over the next four years. We get four more years of scaling, at least, under the current set of constraints. And it’s like, what will happen in those four years? I don’t know. But that’s the only promising direction that I personally know of.
Ivan Vendrov: It sounds like these grassroots communities focused around mutual support, especially around child care... who are basically small groups of people experimenting. I think they’ll eventually become politically powerful too.
Romeo Stevens: Yeah. I wonder.
Ivan Vendrov: I feel personally optimistic about this. I’m part of a community like this in New York... Robin Hanson... was extremely skeptical that any of these groups would succeed at their mission because they’re too exposed to the dominant culture. And so whatever demons came for the main culture in the 70s will come for them too. Unless they’re willing to be extremely insular like the Amish.
Romeo Stevens: Yeah, the Christians developed immunity to whatever was affecting the Romans. So I think it is the case that the subculture that wins will develop some immunity. I don’t think it’ll look exactly like that. A lot of people are like, “Well, the Amish is the working model.” I don’t know. I don’t think so.
Ivan Vendrov: Can you say a little bit more about why not?
Romeo Stevens: They are still bleeding massive numbers of young people to the outside. So they do a thing that’s less self-sufficient and more parasitic. So they can exist on the outskirts of a functioning society. They’re not actually a fully functioning society themselves as far as I know. And they have no ability to... their fertility is also crashed. They’re not immune to whatever is happening.
Ivan Vendrov: I think this is also a puzzle for me... how much should I be thinking about the fertility considerations given how quickly technological change is going? My prediction is it’s less like Fall of the Roman Empire in terms of dynamics and more like the early Protestant Reformation. As in the disruption is to information technology. And I expect we’ll have an information super-plague well before we have a hundred years for the new high fertility groups to outcompete the old culture. And that feels like the main game or the main source of leverage over the future is either creating your own super-plague or creating anti-viruses or cognitive security layers against the coming super-plagues.
Romeo Stevens: Yep. I agree.
Ivan Vendrov: Any promising leads in that direction on either cognitive security or creating your own super-plague?
Romeo Stevens: I mean, Opening the Heart of Compassion remains my go-to for trying to understand how do each of the attractors work and how do they pull people in and how do they seal off the exit ramps. And so developing in particular anti-Hellrealm technology that scales... just by trying things and seeing if any of them scale effectively. I think is... even if it doesn’t ultimately wind up saving everything, it’s going to be a substantial tailwind.
Ivan Vendrov: Would you call early Christianity an anti-Hellrealm technology?
Romeo Stevens: Oh yeah. Like the Romans were being infested by all sorts of Mammon worship and garbage. I primarily view the Jesus meme as an anti-Hell meme.
Ivan Vendrov: Final question just for fun. If you had a million GPUs, what would you use them for?
Romeo Stevens: I have no idea. Yeah, that’ll be one of the filters. It’s like the future goes to those who are very motivated to get compute and have ideas about what to do with it that will bootstrap to getting even more compute... Things... it’s compute plus a few smart people. I’d be restarting the pedagogy research, I think. With the new paradigm. How do we build world-class tutors even better? And filtering functions on humans... finding really high-quality people is basically an unsolved problem. The current capitalist filter puts high-quality people into garbage like finance and adtech. So as we get wealthier and wealthier, many people should be reaching a point where they can say no to a finance job. Their opportunity cost might be high technically, but they’re like, “Do I want another million dollars or do I want to live a good life?” So I think that’ll be an opportunity.
Ivan Vendrov: But it sounds like those filters already exist? It’s just that the filters for good people as you say already exist, it’s just that those people are choosing to do something that you don’t endorse?
Romeo Stevens: No, they get filtered into zero-sum games. Things that no one will talk about. Like the people who were fighting over who was going to win the newspaper wars in the 1800s. Who cares? Oh, this guy became a billionaire instead of this guy. How much does that affect us? Not at all. So it’s like that Medici essay from a few years ago. Hey, you’re a billionaire. Do you want to fight for the next order of magnitude of wealth? You literally already know people who have an order of magnitude more wealth. They’re still just playing the game. Do you want to do that or do you want a shot at immortality?
Ivan Vendrov: Well, thanks so much for chatting with me today, Romeo.
Romeo Stevens: Thank you. It was fun.
