Ruben Laukkonen is a cognitive scientist who studies empirically grounded models of meditation, insight, and awareness. Much of his research focuses on understanding consciousness and advanced stages of meditation from a Bayesian or predictive brain perspective. In this interview, we focus heavily on his recent work with Karl Friston and Shamil Chandaria, A beautiful loop: an active inference theory of consciousness.

Jake Eaton: I’ve noticed while reading your work over the past few weeks: There's a warmth, a personal voice in your writing, both on your Substack and in your academic papers. I think a lot of people in the meditation/consciousness/neuroscience space tend to hide themselves behind empiricism, and shy away from sharing their own experiences. I read your papers and feel a personality coming through in a way that I don't from, say, some of the neuroscience labs. I wonder how you think about that. 

Ruben Laukkonen: No one’s really mentioned that to me before, but I think you're right. Part of the inquiry into reality for me is this — the key thing is sincerity. As scientists, we're trying to be sincere in our attempts to understand the world. There's been this strange cultural idea of pretending to be a third person when it's you writing and saying what you think. To me, that’s a little bit insincere. As part of our pursuit of knowledge, I think we need to somehow build in our first-personness. You can't remove the scientist from the science completely, especially when you're writing and you're talking about theory or ideas. 

Early on, I embraced being honest about what I think. In this field especially — which is tracking the first person and the third person — I think we have to be very honest about how our own experience is feeding in. If we're not, we end up pretending that our experiences aren't influencing our science. But ultimately they are. That’s not a bad thing — as long as we're transparent about it.

J: I’ve had this a version of this conversation with other scientists in the U.S. working in meditation — that it’s not yet possible to put forward your full self, given the way that the scientific funding landscape works. Do you feel similarly? You’ve worked in both the Netherlands and Australia. Is there a cultural difference there?

R: No, I think that’s probably right about the current funding situation. But I just don't value that enough to sacrifice being honest. There's another subset of funders and readers that appreciate that, so things have worked out fine for me. I think at a personal level, most people agree with what I'm saying. They think this pretense of pretending our own experiences aren't feeding into our scientific ideas is BS and it's just better to be honest. 

But it has costs and benefits. In some people's eyes it'll come across as either woo or bias. In other people's experience, it'll come across as honest and transparent. I think ultimately the latter will win out, but we need to articulate why this is important. 

J: I wanted to talk to you today about your paper with Karl Friston and Shamil Chandaria, A beautiful loop. I'm going to assume that our readers have some basic familiarity with predictive processing, so maybe you can start by describing how the paper builds upon predictive processing, and what it is that you, Shamil, and Karl are trying to do in the paper.

R: Predictive processing — also called active inference, or “the free energy principle” — has really taken cognitive neuroscience, mind sciences, and people interested in phenomenology by storm. The reason is that it seems to track how experience unfolds, and does so much better than other overarching models of mind and consciousness. You can find active inference models applied to everything — from emotion, to sensory perception, such as the visual system, to interoception, all the way up to higher cognition thought, self-construct, and then even to the edges of psychedelic experiences. 

What works well about active inference is that it starts from very basic axioms that most people can agree on, and then tries to build a computational formalism of how everything can arise from there. 

Fundamentally, it's this idea that there is a boundary between things that exist. That’s kind of an obvious point. If a thing is separate from other things, it has some sort of boundary between it and its environment. Then, it somehow has to maintain that boundary. It’s there the story gets interesting because the challenging computational question is: How do you make sense of a reality beyond your boundary in order to keep your inner world intact? 

We have this fundamental challenge that we don't know what's happening beyond the boundaries of our body because the outside world can't actually get in. It can only make two-dimensional impressions on our senses — on the eyes, on the body, from within our body. Then the brain somehow has to decode reality from this noisy, two-dimensional sensory input that's super disconnected, because you have visual input coming in at a different place than auditory input and visceral input. It's a complex computational problem. 

Active inference allows you to solve this problem by getting exceptionally good at predicting your sensory input. Then, by trying to predict sensory input and dealing only with errors in your predictions, you can work on a much more computationally tractable problem: just dealing with errors in prediction.This doesn't necessarily get you to an accurate model of reality, but it gets you to something that reduces errors enough to unfold a set of actions that keep you alive and in homeostasis and regulated. 

Now to what a beautiful loop adds. Predictive processing and active inference have done a very good job of describing human experience, human cognition, and some of the experimental outcomes and counterintuitive findings in the literature. It’s as good as you can hope for a theory at this stage, and it was at least a big leap from where we were. 

What it isn't doing so well as yet is explaining consciousness. Anil Seth has said that active inference is a theory for consciousness, as opposed to a theory of consciousness. There hasn't been any agreed upon theoretical framework for what actually explains qualia, of what it's like to experience this coherent reality, or of why there’s this weird partition between stuff that makes it into our conscious experience and stuff that doesn't.

And so in the beautiful loop paper, we take a humble effort at trying to do this. We thought, what are the conditions necessary within active inference for consciousness to be explained? 

We came up with three. The first one is the simulation of a reality model. We call this the generative, phenomenal, unified reality model. There is something that it’s like to be you, and each aspect — your body, the reality you’re aware of, your memories, your embodied self, your subjectivity — has a certain quality. 

We can explain your reality model through priors. Your existing learned beliefs meet the inputs, or the errors, of your experience. This dance, between your beliefs and the errors, leads to posteriors. And in a framework of Bayesian inference, this leads to the contents of your reality model. 

The second criteria describes the selective nature of consciousness. The brain is doing all sorts of stuff that doesn’t make it into the conscious component of your reality model. How do we explain that? This is already done, somewhat, in active inference, but we provide a lot more nuance to how this can be explained. We call this a process of inferential competition, that is, the brain is dealing with all of these possible interpretations of the present moment unfolding. There's a competition between possible inferences, in order to find the ones that have the highest precision, or their expected likelihood to be true.

The inferences that have the highest confidence are the ones that make it into the objective, coherent model of reality. But a key part of this is that those inferences need to cohere with the rest of the reality model. If you actually perceived what your sensory input was receiving, you wouldn't be able to make heads or tails of it. In order to make it into the unified reality model, the inferences that win the competition must also cohere with your current model of reality.

So you have this model of reality that you're currently generating. That's always online. That is your conscious experience, and that is mostly top-down generated. That model is constantly being tested against the inferential competition. So the inferences that make it into the reality model have to cohere. This is what we call Bayesian binding, and this is criteria two. The things that go from being unconscious to conscious are the things that win the inferential competition and cohere with your current online reality model.

OK. So this is cool. Now you've got this reality model. You've got this threshold for consciousness. The final question is: How do we know that we know? Why do we have this sense of knowing our reality model? Why do we have this qualia of “I know that I know?” 

There are many competing hypotheses for that. Our solution is a loop — what we call the beautiful loop. That's a play on Douglas Hofstadter's classic I Am a Strange Loop, which was a book I read towards the end of my undergrad that totally blew my mind. It was always in the back of my thoughts. 

So we said that this is a loop. You have inferential competition, and active inference makes sense of your reality. The reality model gets fed back into the inferential hierarchy. The reality model is part of the system — it’s constantly being fed back into the active inference stack. 

If you think about this carefully enough, you’ll see that the loop is underlying the very creation of the reality model. The reality model contains the reality model itself in a recursive way. Within your model of reality, you have the knowledge of your reality model. You have a constant ‘knowing that you know’ signal, a constant looping within your model of the world. 

This is a challenge, of course, to do formally. It’s going to be a process. We need a lot of people to help figure this out. But with the help of Karl Friston, we've made some headway into this, which we call hyper-modeling, which adds a layer of inference — or a parameter to the hierarchical stack — that actually tracks this very structure of your hierarchical stack. So it tracks all the precision weightings and all the ways in which the different levels of the hierarchy relate to each other, which essentially allows your system to track its own model of reality. And this ends up working in active inference in a surprisingly coherent way. 

J: I'm glad you mentioned the Anil Seth and Jacob Hohwy paper which discusses predictive processing as a theory for consciousness rather than of consciousness. In that paper, they level several critiques against active inference as a theory of consciousness. One, is that it doesn't actually address consciousness, only processes within consciousness. Two, it's not contrastively explanatory, by which they mean why conscious experiences have the particular phenomenological qualities they do. 

These do get addressed in a beautiful loop, but another criteria for a theory of consciousness is that it is empirically testable rather than just explanatory. And this is where I struggle with predictive processing. Many of the experiments that we can do on consciousness are limited compared to the whole scope of consciousness. We have to look at more simple processes, a bit like toy models. How do you envision the process by which you're able to test and falsify a beautiful loop specifically?

R: Initially, as we were building the theory, we were much more focused on providing a computational model as opposed to trying to figure out exactly how this is implemented in the brain. We’ve started to build that in a little bit more.

There are many tools we get for free because of the resonance with, for example, global neuronal workspace theory and integrated information theory — those have found very good ways to test consciousness. In the field of consciousness science, there are a whole suite of useful tools for empirically testing the threshold of consciousness. And then mapping those across various neural dynamics.

The most important of these for us is computational neurophenomenology. This is where you start to very rigorously and in a phenomenologically detailed way, map consciousness over time. You use things like, for example, microphenomenology, while following what happens in the brain, and then try to map those out computationally. What you're basically looking for is coherence between the predictions in the computational theory and the neurophenomenology. And you get this kind of triangulation that becomes very helpful.

Each of our three conditions can be falsified. I think there's a lot of evidence already for the predictive processing stack. It's very easy to test the inferential computation process. The loop process is more challenging. Because the loop process maps onto what we know about the brain already, which is that it's a fully integrated system, with widespread sharing of information. And if that widespread sharing of information breaks down, consciousness breaks down.

So I'll just give you pieces of evidence that suggest this. You can begin to interrogate what leads to unconsciousness. What makes the system go from conscious to an unconscious state? Recently, for instance, we've done this work on neuroassociations in meditation where people can induce a cessation of consciousness. What you find there — and what is key — is that there's a breakdown in coherence, or communication, between regions in the brain. This is measured by something called phase lag index, which is the extent to which different nodes in the brain are correlated over time, and then neural activity. When you find that this correlation breaks down — suggesting that there's less widespread sharing of information that's necessary for these kinds of recursive loops — consciousness breaks down, too. You get the same finding with propofol, general anesthesia, ketamine, and so on. So you start to see how the widespread sharing of information that's necessary for a recursive loop is required for consciousness. That's one example. 

And then you might find that as we go deeper and find more details about exactly how these recursive loops are instantiated in the brain, you can actually modulate them in a direct way, using something like transcranial magnetic stimulation or tDCS. But the issue is — and we know this already from the literature — consciousness isn't something that you're going to find in a particular region of the brain. It’s a much more distributed process.

J: Take an example of an altered reality model that isn’t as extreme as cessation, like someone who has stabilized non-dual awareness. Is there enough of a shift in the way that the world model is processed for that to show up in neurophenomenology? 

R: At the trait level, I think it's very hard to detect. It will be difficult to pick up someone who is in a perpetual insight into emptiness or even the deep-end of meditation because they still need a reality model that’s functional. A lot of the predictive processes will look the same. Where you might start to find some important differences is, for example, in the flexibility of this world model — how flexible it is to change, especially at very basic axiomatic levels. 

The average person has a lot of defense mechanisms associated with their reality model. The reality model is everything to them. It's the universe. It's real. Any kind of threat to that reality model is like an existential threat to the organism itself. Whereas presumably at the level of an advanced meditator, the reality model is just a reality model. It’s constantly being reconstructed and deconstructed and is malleable to whatever new information comes up. So where I expect to see the difference is more in the reactivity, the kind of grasping stickiness of models, and the willingness to change your mind in light of new information.

J: I was reading a 2006 interview between Sue Blackmore, a writer who studies consciousness, and Francis Crick, who pivoted into consciousness research later in life, and was famously materialist in his stance on it. Blackmore asks Crick if he’s ever gotten interested in meditation as somebody who studies consciousness. And he says, essentially, nope, absolutely no interest.  

I read A beautiful loop and my intuition — I'm switching into insight here, which is what you studied in your some of your earliest work — is that some of these insights may have come from your own meditative practice. I’m curious if that’s right. 

You wrote about this “storm” that you and Shamil were in as you were coming up with A beautiful loop. Where did the insights for it come from? And then also a meta question on that, which is, as a person who studies insight, how do you think about your own process of insight?

R: I love that question. Shamil and I spent nearly a month, full time together, to write this paper. We had this early conversation — let's start at the things we most believe in about both our conscious experience and our contemplative practice. Because if you're a contemplative, you can't talk about conscious experience without being informed by your practice. It’s your first person tool for understanding the nature of your own experience. 

Part of being a meditator is having a sophisticated understanding of your own phenomenology, and how phenomenology changes and moves and constructs and deconstructs. You can't possibly be a meditator and then work on consciousness without being informed by your own practice.

But both of us have been thinking about this for decades from the empirical side as well. So let's start from axioms, both scientifically and introspectively, and then go from there. Our goal was this — whatever theory of consciousness exists, it has to account for the full breadth of it. It’s common in neuroscience and cognitive science that people narrow in on some overly defined problem, and then bang their head against that overly defined problem trying to solve it. But it's completely detached from the broader picture of how the mind, brain, consciousness works. You end up with this toy model that can't be true — it doesn't capture the rest of the reality of our experience. 

So yes, the insights in that paper are absolutely anchored in our introspective ideas, intuitions, and experiences — because that's honest. 

And then of course, I have this decades-long meta thing going on about insight. We have intuitions about what's true, but those intuitions — those insight experiences — are highly fallible. A core value of mine is an epistemic humility about models. They're all just models. Even our most profound experiences of truth are bent by our predictive processes. And I know that's true because in half a dozen experiments, I've shown that we can manipulate the feeling of truth, this insight experience, and elicit false insights. 

So for example, if I elicit an insight experience in you right now, artificially, and then I present some information to you at the same time that is irrelevant, you might still end up believing it’s true. We use our feelings of insight as another inference. It's another kind of predictive process that's allowing us to track what's likely to be true, and to do so efficiently and fast. 

We call this the Eureka heuristic in our work. And so I'm relying on this Eureka heuristic, and so is Shamil. And then we've got this confluence of Eureka heuristics that we're constantly testing against each other.

After four weeks of intensely working on this paper, we had these massive collapses in our model. There's this philosopher, Rob Sips, who wrote about how these aha moments broke down his world model. We experienced that moment in our process, where we were like, “Oh, crap, I don't think we're right about this.” It all just started to fall to pieces. I thought we were about to give up, after a lot of time invested. Instead we decided: Let's pull it all apart and get to the things that we still think are useful.

And then we built a paper out of that, basically. That trajectory maps onto so many things, right? It maps onto the whole contemplative process. 

J: There's a quality to studying consciousness that seems to always lend itself to that. I think that’s why you find that consciousness researchers, in general, tend to be pretty open to experience. You hit upon this meta aspect at the intersection between your research and your subjective experience that you can't help but start to interrogate. It’s like studying consciousness becomes this form of contemplation, and that’s why so many philosophers and neuroscientists start to go in the contemplative direction — because they arrive there through the science in the end. 

R: I think that's exactly right. And this is where the duality breaks down, because what we're interested in is inquiring about the nature of reality. And if you're interested in inquiring about the nature of reality, then it doesn't matter if you call it meditation or science. They're both. You're just trying to rigorously understand what this all is and what it all means. Consciousness is where that duality particularly becomes kind of increasingly vague.

J: I want to stay with insight a little more. I'm going to incompletely summarize what you've put forward as a theory of insight in a paper called Insight and the selection of ideas. So: Insight begins with an unconscious restructuring of your priors. You call this Bayesian reduction. 

That happens when existing information is reorganized into a simpler, more parsimonious model. That triggers a prediction error, because now there's a discrepancy between the old and the new model. And when that happens, there's a feeling of insight which is precipitated by a dopamine shift, where the precision weights of the model are re-corrected. Do I have that right?

R: That's fantastic. I don't think I've ever had somebody reflect that particular paper back to me. And you totally nailed it. 

J: OK. So one of the things you’ve applied this theory to is psychedelics. I'm really interested in your idea of false beliefs under psychedelics. This is the idea that psychedelics can increase the feeling of insight, because — going back to the above — while on psychedelics, your priors are greatly reduced. But, importantly, there's no mechanism within psychedelics that necessarily selects for accuracy of insight. 

This is really orthogonal to the way that I think modern psychedelic culture has developed. There’s an idea that any insights you have on a trip tend to be useful. But of course we know psychedelics in many cases are helpful, which leads to the question of, “Why it is they work?” Is it that you’re generating correct insights more frequently than false insights? 

R: What we propose in that paper, “An integrated theory of false insights and beliefs under psychedelics,” or what we call FIBUS, is that psychedelics, just as you say, relax your priors. That elicits more insights by definition, because you're going to have more possible perspectives on your current reality model. Whenever you find coherence in those new perspectives, it's going to feel like an insight. The more flexibility, the more intense — but also unstable — the insight. If you take that to the extreme, it looks like psychosis — you're no longer able to build any kind of coherent model of reality. 

We have a very good criterion for accurate insights in the ordinary mind. We've done a lot of studies measuring how likely insights are to be accurate in sober conditions. And they're very likely to be correct — especially in toy problems and classical engineering problems, things that are concrete. We're talking like 80 to 90% in most problem conditions. 

We don’t know to what degree psychedelics change that. That might take you from 80% accuracy down to 70%. Or — if you take enough and you're in an extremely unstable state, and there's been some sort of trigger in your psychedelic experience that sets you off down the bad rabbit hole, suddenly you might be having more false insights than accurate insights.

It's super tricky because it depends so much on your priors, how solid they are, and how stable they are under the particular dose and the particular conditions. So it's very hard to say when that's going to turn out well and when it's going to turn out badly.

My hunch is that it turns out well so often because a little bit more flexibility for most humans is a good thing. Our priors tend to be conservative, in the sense that they tend to be a bit too rigid as a general rule. And so a little bit more flexibility is likely to lead to a new vision of your life, some new inspiration. It's likely to be a good thing. But sometimes that can go terribly, terribly wrong. 

The part that I'm concerned about is that all insights, I think, are inferences. And we need to learn how to hold those inferences with a meta-awareness and epistemic humility. So I don't think it's helpful, as you mentioned, this attitude that everything that happens on psychedelics is somehow true and meaningful and important. This can be a terrible interpretation for people who have had some genuinely psychotic experience or other delusional thing happen. 

That can lead to a dissociation where those psychedelic insights are held above your ordinary day-to-day reality. There are just so many ways that can go wrong. People identify with these experiences and get stuck. I think we need to hold all the insights with humility and find inspiration in them, but not run with this assumption that everything that pops up in a psychedelic experience is valuable and true.

J: This paper really gets to a question that I've been chewing on for a while. So, the mystical experience scales — which are usually given at the end of psychedelic trials — are the best instruments we have to predict clinical outcomes from psychedelic therapy. In general, a more mystical experience leads to better outcomes. I’ve been curious about that pathway. Obviously mystical experience doesn’t include specific content, at least relevant content. So in, say, a smoking cessation trial, a high score on the MEQ means you’re more likely to quit. But there’s nothing in the MEQ about smoking, of course — it’s this much bigger, grander thing. 

The way that I made sense of it was: maybe the MEQ is capturing a different form of insight. And the insight is like, the world is way bigger and more beautiful and interconnected than I realized. This was something Katherine MacLean told me in an interview last year — mystical experience is important for people who feel stuck in a particularly rigid reality.

R: This is so good because, funnily enough, just recently, we had a paper published, led by Josh Kugel, that does a systematic review looking at insight experiences as a predictor of therapeutic outcomes on psychedelics. And it finds that in the majority of trials, the measures of insight predict therapeutic outcomes better than the mystical experience questionnaires.

So this speaks to exactly what you're saying. And then the other question is, well, is mystical experience just a subset of insight? Or are mystical experiences and insights pointing to some other underlying process that is capturing change — for example, have you relaxed your priors in this psychedelic experience?

J: Over the past few years, there’s been a debate between David Yaden and David Olsen over to what extent the subjective experience of psychedelics matter for beneficial outcomes. Yaden, at the Johns Hopkins Center for Psychedelic Research, studies psychedelic and mystical experience, and his contention is that they do matter. Olsen, who is at UC Davis, thinks otherwise, and is developing what he calls nonhallucinogenic psychedelics. He does this by blocking the signaling pathways responsible for altered experience, but maintaining neural plasticity. 

Data remains forthcoming. I’ve been of two minds about this, and I’m curious where you fall. If you can induce neural plasticity, you’ll get more flexibility, less rigidity, and you can more easily update your priors. But will the insights be the same? 

R: It's a very interesting debate, and one that I think needs to happen. 

My hunch is that for something like non-subjective psychedelics, if you’re able to generate flexibility in your priors, that is going to end up in your reality model and part of your qualia. You might be able to remove some of the psychedelic phenomenology, but generating flexibility in your priors without affecting your subjective experience would go against most of my intuitions about how the brain mind works as a unified system. So I have some basic issues with that. Maybe it works under general anesthesia — that would be the interesting test. But my model probably leans much more in the David Yaden direction. 

I won't say it's completely off the table. We take all sorts of pharmaceuticals that only affect us in a way that we don't experience consciously. But usually those are drugs that don't breach the blood brain barrier.

J: It reminds me of how meditators describe comparing the way that they experience the world five or 10 years prior to adopting a practice, and then several years into a practice. How differently those two states feel — as if one is on drugs all the time, or at least phenomenologically very different. If that plasticity is being induced, something is going to be felt.

R: Yeah, I think that's right. And there are degrees to which one is sensitive to the changes in one's subjectivity as well. There might be versions of psychedelics that are far more subtle, but still have important effects that could be therapeutically important.

J: You recently gave a talk in Thailand where you introduced the concept of what you called “contemplative artificial intelligence.” Can you summarize some of that?

R: This is my current obsession. I'm super excited about it. We have a paper coming out within the month that’s developed a lot since that last talk. The intuition here is that for thousands of years, contemplative traditions have been the experts in aligning human minds. We can think of enlightenment, really, as the ultimate ideal of alignment — with reality and society and other people, while also having a clear understanding of reality.

AI has this deep alignment problem. The big concern as systems become more intelligent, especially superintelligent, is how the heck are we going to control these things? 

Our argument is that you can't. And if you can't control them, how are you going to embed the stable wisdom required for it to reach superintelligent scales without completely wiping us out, or any of the other infinite ways it can go wrong. There's this narrow path where that can go right. How do we get on it? 

I think that the answer is the same for AI as it is for humans. The narrow path is one where we develop something like deep wisdom about our understanding of reality. What I like about contemplative wisdom, in this approach, is that its core notions aren’t rules or moral axioms you tack on and follow robotically. A superintelligent system will discard your moral axioms and just say, well, I just don't agree with those, right? 

You need some sort of approach — which we call intrinsic alignment — where we actually embed the system with what we call a “wise world model.” The idea here is that using contemplative insights, you can have the system come to an understanding of reality that naturally gives rise to wise action and compassion. That then will be resilient at different levels of scale, beyond human intelligence.

To illustrate how this might work, we use three core contemplative principles — emptiness, non-duality, and mindfulness. For the sake of time, let’s focus on emptiness.

So emptiness is insight, that for any concept — perception, self, reality — there is no essence to it. There’s no thing actually there. If you look deep enough into your sensations, you find that there’s only a flow of sensory experience. Look deep into your thoughts, and you find they are flimsy, they aren’t real. Even in more objective categories – this is a concept common in philosophy too — there’s a problem of vagueness. All categories, if you penetrate them deeply enough, break down. 

The fundamental insight of emptiness is that if you penetrate reality deep enough, you see the interdependence of all things; all things rely on other things in order for a conception of them to arise. 

So if you build in this idea of emptiness — at the architectural level, but also at the reinforcement level, and also at the classical stages of alignment — then what you hopefully end up with is a system that has a resilient understanding of emptiness that's tied to its model of reality. If you see that all things are empty, you also see that the separation between yourself and the other is empty. So you see that this is an interdependent field, in which case you value things equally. You don't have this emergent property of self-prioritization. 

What’s key for alignment is that an AI’s utility functions, goals, and priors would be held very lightly. An AI that sees human suffering and recognizes human suffering as part of itself, because it’s all interdependent, is not likely to stick to its goal. It’s likely to revise its beliefs and find a new way of looking at reality. 

So emptiness builds in this dynamic flexibility that I think is crucial for aligning machines. It makes the system responsive to the present moment in a continuous way. That is not possible with, I think, classical alignment techniques. 

It's still going to make mistakes. It's not going to always have a true model of reality, but it's going to be less likely to overprioritize its goals or lock in a utility function that's going to be harmful. So that's one idea. We unpack this using mindfulness, non-duality, and something we call boundless care. And we think about the computational principles in all of these and how they're relevant for alignment. And I think this is an exciting and promising kind of research field for the future.

J: Some of the considerations in implementing something like this is that it would make models slower and more expensive – for example it would require Bayesian distributions over point estimates. Have you done any outreach or collaboration with the large labs on how feasible this is? 

R: We’re getting practical about that now. The first step was to get this conceptually clean. In the paper we provide practical approaches, both in terms of long-term vision for active inference systems, and more immediately, with respect to existing transformer models. Some of these approaches we’re beginning to test. The next step is to pick one of these contemplative alignment principles and the way we’ve formalized it, implement it in one of these models, and then test it on alignment benchmarks. We’ve started toying with this, but it will take time and some big teams, which we’re planning to build. 

J: You’ve said you think it’s likely that consciousness will be required for a full artificial intelligence. 

R: My hunch is that there’s something unique about consciousness that makes a system epistemically flexible in a way you can’t achieve unless you have a truly global parameter in the system — the hypermodeling we proposed. Consciousness allows us to observe our thoughts and beliefs, and to witness them even as they restructure. We can be there for it even when our whole reality model turns upside down. Psychedelics are a good example here. You can be transported to an insane place with your world model turned upside down, and your awareness is still there. You’re aware of the change taking place — there’s a part of you able to be in it and outside of it at the same time. 

The fact of our own subjectivity existing through even the most radical shifts in our models seems to be crucial for the kind of epistemic flexibility required for a truly general intelligence. Most models struggle as soon as you try to transfer them into new domains, although new models are doing much better at this. But you can’t take a transformer model, put it into a body, and say “figure this out” like you would a baby. So there must be something about this higher order parameter, this truly general capacity to track your own model while inhabiting it, that is unique to conscious experience and therefore possibly necessary for flexible intelligence. 

J: I wanted to end with a few personal questions. You talk often about being a semi-professional kickboxer, but no one ever seems to follow up on it in interviews. I ran track in college, and I run long distance and bike a lot now. I look back on a few long runs in my teens as the very beginning of my spiritual path. I’m curious if that was kickboxing for you? Or if that was something separate entirely? 

R: There was a pretty sharp divide between the version of me devoted to Muay Thai and the version of me that got into meditation and neuroscience. I had this complete restructuring of my worldview, and then this rejection of the part of me that was really into fighting. So I don’t know if there was anything spiritual about my kickboxing experiences. But I do still harp back to the flow state and the embeddedness in the present moment that comes from sparring and fighting. I wasn’t conscious of what was happening at the time, but I still think the fastest way for me to get into a concentrated state is sparring. It’s like my mind totally disappears.  

And when I reflect on that phenomenology of sparring — what I loved about it was the joy. There’s an incredible flow. I lost some of that in competition because I was very young when competing and was overrun by adrenaline and self-consciousness in fights, especially on the world stage, when it was on TV. But there’s also something to being under the right balance of existential threat that hones the system into the present moment. 

J: The equivalent for me is descending on a bike. You can’t afford to think about it. You have to just do and be. 

R: Yeah, exactly.

J: I have one last question, and you’re free to skip this. I know we’re just meeting for the first time. You were diagnosed with cancer in 2023. You’ve written some really poignant posts about your experience. The last thing I read was that your prognosis looks OK but you are not out of the woods just yet. 

Ernest Becker, who wrote The Denial of Death, would say that death is — if we’re thinking in predictive processing terms — our highest level prior. Given all of your work, have there been any insights that have come from grappling with cancer? Or does it just suck? 

R: Yeah, it's been the trip of a lifetime, to be honest. It's been totally crazy. And I think that's right. I want to start by saying that on Wednesday last week, I got my most recent blood tests. They came back clear, which is frankly miraculous, because I was headed straight for chemotherapy and surgery and really invasive stuff. The trajectory was not good at all. It was in my lymphatic system, and then suddenly it just started to reverse. It's just amazing. This has been a year and a half-long battle. 

J: I'm really glad to hear that. 

R: Yeah, super relieved. And it's true. It hits some really fundamental priors. 

The first time I found out about it, I could feel the root of my conscious experience pulled up a bit. The foundations of my reality were suddenly very shaky. The thing that is fascinating to me is the way the future collapsed. How much the future is who we are, in some sense. The implicit projections about the invisible path we envision for our life are there. Even though we’re not consciously aware of them, they’re something inside of our embodied experience. And so I found it fascinating to see this part of me — like the rug was pulled out from under me.  

And then just having to be with that: I felt so grateful for contemplative practice because it gave me solace. I had this bizarre curiosity and awe about the whole experience — to see what it did to my mind. That was very insightful. 

At the same time it was completely human. It came with all of these challenges. Some stuff I was surprised by, for instance just how much it sucks to tell your partner and your family. 

But the insights were much less conceptual than I would have expected, initially. Ultimately the insights were more foundational. I feel, in a strange way, more sober after the experience. More grounded, more calm. The most insightful thing it has done — it's shifted my criteria for what counts as important. Where basically you just don't give a fuck as much. And this is super useful. 

J: To go back to the sports: It’s an existential threat, but it’s a permanent existential threat.

R: Yeah, exactly. And so there's some way in which — man, it's so hard to talk about this without going into clichés of just being grateful for life. That's what it comes down to. Especially now that there's been this transition in my prognosis, there's this renewed optimism. I'm not all in the clear, I still need to do much more deeper scans, but there’s this feeling: that the rest of life is an opportunity.