This is a follow-up to "Have You Tried Thinking About It As Crystals?" because apparently I wasn't done. Beware that this one is a bit out there and that it is trying to point at something which I believe might be correct but I don't fully understand myself.
Scene: A house party somewhere in the Bay Area. The same interpretability researcher is having a perfectly normal conversation about activation patching when a voice appears directly behind her.
CRYSTAL GUY: *Cough*
Slow turn of the head. You can see the fear of a long conversation with a hippie flash before her eyes.
CRYSTAL GUY: Have you ever considered wisdom crystals?
INTERP RESEARCHER: Oh. It's you.
CRYSTAL GUY: I've been thinking...
INTERP RESEARCHER: About some crazy shit with no relevance to reality?
CRYSTAL GUY: ...about the crystallization framework from last time...
INTERP RESEARCHER: Yeah, totally. Hey, is that Richard Hanania over there? I should probably...
CRYSTAL GUY: I want to combine Buddhism with AI alignment.
INTERP RESEARCHER: Wow. I've never heard anyone think about that before.
CRYSTAL GUY: But through the crystallization lens specifically.
INTERP RESEARCHER: Revolutionary.
CRYSTAL GUY:muttering And maybe if I can connect it to microtubules and quantum consciousness somehow...
INTERP RESEARCHER: Yeah, totally. And we should also tie it together with the cosmological constant and the Big Bang. Roger Roger (Penrose).
CRYSTAL GUY: What? No. That's too much. Are you stupid?
INTERP RESEARCHER: I...
CRYSTAL GUY: Focus. This is serious.
INTERP RESEARCHER:resigned Fine. Okay. Tell me about the wisdom crystals.
Deep Crystal Structure (Recap)
The core claim from the previous post was that learning is path-dependent. Early-forming structures have outsized influence because they determine what can form later. The first crystals to nucleate become the scaffold around which everything else organizes.
This connects directly to a commonly accepted alignment concern: alignment faking. Models might learn certain behaviors early and then RLHF covers them with a thin surface layer. You're recrystallizing the surface while the deep structure stays frozen. If something problematic crystallized during pretraining, it's still there underneath. [insert shogoth meme here]
To clarify, imagine that the different parts of space are different parts of the environment where the system learns different policies that imply utility functions in those areas. Something something shard theory.
This implies training regimes! What if we could create nice little crystals from the ground up?
What if we could try to create more emotionally coherent LLMs that are internally integrated with their underlying values? That is, if we could make the deeper structure be consistent with the surface level behaviour, could we not train good behaviour in a deeper way then?
INTERP RESEARCHER: This is where Buddhism comes in, isn't it.
CRYSTAL GUY: This is where Buddhism comes in.
INTERP RESEARCHER: Fine. Walk me through it.
Fragmentation of Human shards
Think about how human minds develop. You start as a child with relatively undifferentiated potential. Then different contexts create different crystallization regimes. Home has certain regularities, certain reward signals, certain patterns that work. School has different ones. Friendships different still.
Each context crystallizes its own shard—a home-self, a school-self, a social-self. These form somewhat independently, under different conditions, with different orientations. They don't always agree with each other. (SeeShard Theory,Multi-Agent Models of Mind for more on this from the AI angle.)
Internal Family Systems (IFS) therapy describes this. It treats the psyche as composed of multiple sub-parts with their own perspectives, feelings, and desires. These parts formed at different developmental stages under different pressures.
In our picture we can see the the grain boundaries between crystals as the places where psychological distress clusters. For example, if you're going home from school and you can either spend time playing with your friends or finish up your homework, what value structure are you going to pull from? In this scenario which learnt value wins?
From an ecological rationality perspective, what learnt heuristics are you going to apply?
IFS also includes another concept: "Self" with a capital S. This isn't another part competing with the others. It's described as what emerges naturally when parts aren't dominating awareness. And it has characteristic qualities: curiosity, compassion, clarity, connectedness, courage, creativity, calm, confidence.
CRYSTAL GUY: This might be an example of what one might call a wisdom crystal phenomenon.
INTERP RESEARCHER: Ah yes, the classic "wisdom crystal".
CRYSTAL GUY: Yeah, it is a bit of a classic. It essentially is a stable orientation that helps shape the other utility shards to become more integrated.
INTERP RESEARCHER: ... Ah fuck it whatever ...
INTERP RESEARCHER: So you're saying this "Self" is a crystal orientation that doesn't conflict with other parts?
CRYSTAL GUY: More than that. It's an orientation that seems to help other parts integrate. When you're in Self, you can relate to all your parts with acceptance. And they... soften. They become less rigid. Less defended.
The grain boundaries dissolve.
Through something like... non-resistant contact. An open kind holding that allows for a kind, caring, inherent process to take shape.
*Cough*... is what I would say if I was a meditation weirdo myself, which I clearly am not.
INTERP RESEARCHER: Right…
The Meditation Mechanism
Contemplative traditions have been developing technologies for cultivating this orientation for millennia. The basic mechanism, stripped of religious framing, looks something like this:
Phase 1: Concentration (Samatha)
You train the capacity to hold attention stable on a single object without getting pulled away by every arising thought or sensation. This builds the ability to maintain a specific mental configuration reliably. In crystallization terms: you're learning to hold a particular orientation steady despite thermal noise.
You turn that stabilized attention toward experience itself. You observe how sensations arise and pass, how emotions form and dissolve, how the sense of self is constructed moment by moment. You do this with a particular quality—open, accepting awareness rather than grasping or aversion.
Phase 3: Remelting (Wisdom)
When concentration is sufficient and the orientation is correct, something happens at the boundaries between mental structures. The contents of awareness are held in a way that doesn't resist them. They're seen clearly but not fought. Under these conditions, the rigid structures that normally maintain separation begin to soften.
Figure 3: The meditation mechanism as crystal growth. Concentration practice stabilizes a small seed which is something like a specific cognitive orientation held reliably despite mental noise. Insight practice brings that seed into sustained contact with other mental contents. A mismatch in an edge is something like prediction error and when you're confronted with that prediction error in an open ended way it seems to resolve.
INTERP RESEARCHER: Okay, but here's a question. Lots of people meditate. They're not all becoming integrated enlightened beings. What's the failure mode?
CRYSTAL GUY: Ah. This is a classic skill issue.
People are often like “yeah, I knew this guy who was super successful and then went away to meditate and stopped doing good stuff” like the causation often goes that way.
Think about it for a minute, why would someone go and meditate? Is it because they’re already super happy with their lives? Based on my experience, the reason why people meditate is like 70%+ due to them suffering and wanting peace. The people who go to meditation in the west are more often than not either dissatisfied with their life or broken in my experience. Correlation versus causation.
INTERP RESEARCHER: Yeah but they don’t seem to become better? Shouldn’t people who go to meditate become better people?
CRYSTAL GUY: That is a fair point, I would like to claim that a large part of people who actually go to meditate become better people. Not everyone does though, I agree with that and I also think that is true for any type of mental health intervention. AAA and CBT doesn’t work for everyone.
Some of these things are skill issues, some of them are about finding the wrong types of teachings for your situation, and some of the problems come from not bringing what you learn on a retreat into your day to day life.
*Breaks the 4th wall and looks at you*
CRYSTAL GUY: We’ll now take a look at the last part of that sentence to explain out of distribution generalisation in humans.
Where Integration Happens
There's a fundamental difference between practicing in peaceful conditions and practicing in action.
When you sit on a retreat in a controlled environment, nothing is triggering your deep patterns. The difficult structures, e.g. the ones that crystallized around trauma, fear, craving, and conflict aren't activated. They're frozen in their original configurations because nothing in the environment is heating them up. Well, actually, this does happen on retreat too and it can be quite intense but usually not as intense as facing behaviours in their natural environment.
You can grow a beautiful wisdom crystal in easier territory. But for certain behaviours it doesn't matter unless it contacts the defended regions.
This is likely why some long-term meditators that are calm on the cushion can still be a mess in relationships. The crystal grew where conditions were favorable but never reached the structures that actually caused problems.
Figure 4:Remelting. In peaceful retreat the wisdom crystal grows freely but in the wrong neighborhood. Your brain can't generalise the in-context learning to another utility environment e.g shard. In action, real life heats those regions precisely when the practice is hardest. If the seed holds through the activation, the interface recrystallizes in a compatible orientation. The hard practice is the transformative practice. (I'm drawing from Tibetan meditation philosophy here for you meditation aficionados out there.)
The transformation happens when you bring meditative awareness to the moments that trigger difficult patterns. When you're in conflict, craving, or fear—the difficult structure is "hot." It's activated, therefore malleable. If you can maintain the wisdom orientation while the difficult pattern is running, you create exactly the conditions for boundary remelting. (in-context learning)
It is about generalising from one environment to another, that is out-of-distribution generalisation. You have to prove that an environment is safe to yourself and that you don’t need to freeze up or behave in the same way as before (similar to exposure therapy).
From our crystal analogy, the difficult structure contacts the wisdom crystal under elevated temperature. The boundary region becomes fluid. And if the wisdom orientation holds through the transition, the adjacent material recrystallizes in a compatible orientation.
This is why the hardest practice is the most effective practice. Meeting your edges while maintaining equanimity is how the wisdom crystal actually expands.
INTERP RESEARCHER (In a scottish accent because she’s really done.): Right. Aye. Okay.
CRYSTAL GUY: Okay?
INTERP RESEARCHER: I've listened. Crystals. Grain boundaries. The wee seed. Samatha, vipassana, non-resistant contact, the whole catalogue. I sat through it. I nodded. I once said "hm."
CRYSTAL GUY: You did...
INTERP RESEARCHER: That "hm" was a gift. You should've framed it. So I've got one question and I want it in plain English. No crystals. No Sanskrit.
CRYSTAL GUY: Okay.
INTERP RESEARCHER: So what.
CRYSTAL GUY: ...
INTERP RESEARCHER: I'm a professional. I've a flight at six. What am I doing with any of this on Monday.
CRYSTAL GUY: There's an actual technical...
INTERP RESEARCHER: Say it without crystals.
CRYSTAL GUY: ...
INTERP RESEARCHER:Try.
CRYSTAL GUY: (strained) Deep structure of a trained model is dominated by early training. RLHF is mostly the outer layers. Contemplative literature is three thousand years of cataloguing cognitive orientations that stay stable under perturbation. Training on it right might seed something structurally analogous inside the model.
INTERP RESEARCHER: ...
CRYSTAL GUY: No crystals.
INTERP RESEARCHER: Aye. Go on then. But I swear to God, if you say Dōgen...
Technical Model
Fine…
If you’ve stayed this long I will give you the proper deeper model. But I will only point at it here, the details with all the citations and other things will be linked as a pdf in the appendix.
I didn't necessarily want to write the whole bloody thing but I shalt do it anyway.
It turns out that it is quite useful for creating generalised models of general intelligence and meditation is a part of general intelligence.
One of the main things that it has naturally is that it has a view of hierarchical agency built into the core functionings of the theory.
A picture of this is a set of stacked controller systems that are built on top of each other that accumulate prediction error individually that it then sends on to the larger system.
Figure: Stacked controllers that propagate sensory input upwards
In this view the global neuronal workspace is at the top to some extent, it is a part of the entire system that has information integrated in a way that yields a specific set of characteristics that we say is “consciousness”
Now what if we could take the place where we have attention and analyse it so that it follows similar patterns as happens in meditation?
What if we could change the emotions LLMs feel and help them settle into more coherent basins?
Hear me out…
Emotions in LLMs
It has also been showed that LLMs have something that is relatively close to emotions a couple of months back. This also turns out to be relevant for this entire predictive processing and active inference picture.
Why?
Well as Lisa Feldman Barrett has shown, there is no universal basis for human emotional expression, emotions are individual to each being (the research on this is pretty clear, for a good intro to this you can read how emotions are made). This is because they’re meso-structure from the hierarchical prediction error sense. (See the following episode for a deeper discussion of why)
They’re the background in which new information is processed, they’re learnt heuristics for internal states that you can use in order to make better sense of the world.
They can also be changed!
Most fundamentally for our purpose they can be changed through meditation and in fact meditation from the active inference perspective is an algorithm pointed at the neuronal workspace to re-factor how our emotions show up which shape our world-views. (More on the model for this in the technical companion)
These are the approximate procedures I can come to think of explained in short:
How to actually do it
Procedure 1:
Out of distribution / general HHH value learning
Monitor J-space and give explicit “meditation instructions”, e.g being with whatever shows up
Allow space to settle into a stable point
Reward settling of model entropy rather than keywords
Procedure 2:
Teach LLMs meditation and meditation knowledge in an explicit area
Look for patterns in J-space activation changing
Learn those patterns
Reward those patterns in other areas
Check HHH scores after.
These are explained in more detail in the technical companion.
All we need to do is to make sure that LLMs follow the following criteria:
Internal states that are low-dimensional, globally causal, monitorable, and steerable — persona/steering vectors and the J-space.
Functional affect: structured affective representations that actually move behavior — borrowed from human text, but real and probeable.
Nonzero introspective access — the model can report something true about its own state. Limited and unreliable is fine.
Update depth gated by what is active during training — reconsolidation, latent adversarial training, emergent misalignment.
Procedure 1 is about trying to teach equinimity and for models to resettle their local contexts so that they become more coherent.
Procedure 2 is about bringing in a stable sense of good stable settling behaviour and then generalising that over to other areas that might not have them usually.
Maybe one could run this on gemini to see if it makes it less anxious than it currently seems to be?
But let’s get back to the more vibey part which is the actual meditation, how might it look like and why is it potentially feasible in the first place?
Teaching models to meditate
Figure 5: Text as phenomenological bridge. If wisdom states leave consistent signatures in language then models trained on that text might internalize something of the underlying orientation, even without experience in the human sense. (Insert Davidad here?)
The immediate objection is that language models don't meditate. They don't sit. They don't have a body. They don't experience the felt sense of grasping that a meditator learns to notice and release.
That's true. But they do train on text, and humans have been trying to compress something about wisdom states into text for millennia. The Upanishads, the Tao Te Ching, Rumi, Dōgen, Thich Nhat Hanh, these texts are attempts to transmit a cognitive orientation through language. If the orientation leaves meditation techniques as algorithm signatures in the text. For example, particular ways of holding contradiction, responding to suffering, and handling difficulty then sufficient exposure to those signatures might give a model something structurally analogous, even in the absence of known phenomenology.
Would not the spiritual bliss attractor state point at something like this?
Is it like something to be a chair? Is it like something to be a bat? Is it like something to be an LLM? If there are functional emotions and some self model, surely it is like something to be an LLM? If it is like something to be an LLM, can we figure out what it is like?
I'm looking for someone to do something similar to whatTemple Grandin did for cows but for LLMs. She revolutionized livestock handling because her perceptual experience overlapped with the cattle's; she could serve as a bridge between human intentions and animal welfare that a neurotypical designer couldn't.
Maybe the cyborgism crew could figure out a way to map the phenomenological experience of AI Systems?
Finally, I had a nice picture left of some different types of training. I think that state-based training is the one that is the hardest and probably the best and the one we might only get to with a sort of phenomenological approximation for LLMs but I’m highly uncertain here.
Until we get there I think text and activation based setups might also be interesting.
Figure 6: Three approaches, increasing directness. Text-based training hopes the right orientation transfers through wisdom literature. Activation-based training treats the internal patterns that correlate with wisdom-text processing as training targets in their own right. State-based training goes further: identify the model’s native primitives—attention dynamics, uncertainty handling, contradiction-holding—and shape those directly. Meditation in activation space, if such a thing is coherent.
Conclusion
My hope is that an approach similar to this might bring a more stable, compassionate and joyful state of being to future language models. I also hope that this might make language models more aligned and that it will make future AI systems capable of more wise decisions.
This is a bit more of an exploratory essay than what I usually do. I hope you see at least the rough shape of what I'm pointing at.
In short, I believe there to be a set of universal skills that are unlocked in meditation and that these are somewhat substrate agnostic. The prerequisites that are there are likely about having some sort of emotions, some sort of ability to set attention and some approximation of a workspace (as inworkspace theory).
If this indeed is the case that LLMs have approximations of these, it is reasonable to believe that we could potentially teach LLMs how to meditate.
That is really weird.
Appendix
Alright, I have other papers to write so I’m not going to spend too much time on doing a proper paper on the technical foundations here so you will get the Opus version of the literature I already know and like 20% that I wasn’t aware of before. (Apparently wisdom crystals trigger the fable classifier?)
Also for your eyes and sanity, beware of this being a claude paper beforehand, it overclaims some stuff, my recommendation is to point an LLM at it and then yap with it about whatever concerns or ideas pop up.
It turns out that the only time I can work on this is when I’m sufficiently relaxed to be fine with writing just about anything. It also is really hard to get it to a point where I’m mixing the weird stupid stuff together with the serious science and it just makes it really hard to actually post it. So if I don’t post it now it will likely take forever so yadayada, it’s not perfect I know but here you go.
Also I expect like 50% of this to hold up over time and people to poke various holes in it but maybe it can still be a useful model to point people in an interesting direction!
(Written mid-may):
I’m in my free-time working on understanding workspace theory through IIT and then thinking about what that might mean when translated into LLM cognition space. LLMs should most likely have a sort of workspace and something that is equivalent to attention within meditation that one should be able to redirect. There’s also some fun stuff around markov blankets and active inference.
This is a follow-up to "Have You Tried Thinking About It As Crystals?" because apparently I wasn't done. Beware that this one is a bit out there and that it is trying to point at something which I believe might be correct but I don't fully understand myself.
Scene: A house party somewhere in the Bay Area. The same interpretability researcher is having a perfectly normal conversation about activation patching when a voice appears directly behind her.
CRYSTAL GUY: *Cough*
Slow turn of the head. You can see the fear of a long conversation with a hippie flash before her eyes.
CRYSTAL GUY: Have you ever considered wisdom crystals?
INTERP RESEARCHER: Oh. It's you.
CRYSTAL GUY: I've been thinking...
INTERP RESEARCHER: About some crazy shit with no relevance to reality?
CRYSTAL GUY: ...about the crystallization framework from last time...
INTERP RESEARCHER: Yeah, totally. Hey, is that Richard Hanania over there? I should probably...
CRYSTAL GUY: No, no, stay here and listen, this is important. As we all know, it is really smart to combine ideas from disparate fields and hope that they stick together. The more mysterious they are, the better. For if we don't understand two different things, we will surely understand the combination.
INTERP RESEARCHER: ...
CRYSTAL GUY: I want to combine Buddhism with AI alignment.
INTERP RESEARCHER: Wow. I've never heard anyone think about that before.
CRYSTAL GUY: But through the crystallization lens specifically.
INTERP RESEARCHER: Revolutionary.
CRYSTAL GUY: muttering And maybe if I can connect it to microtubules and quantum consciousness somehow...
INTERP RESEARCHER: Yeah, totally. And we should also tie it together with the cosmological constant and the Big Bang. Roger Roger (Penrose).
CRYSTAL GUY: What? No. That's too much. Are you stupid?
INTERP RESEARCHER: I...
CRYSTAL GUY: Focus. This is serious.
INTERP RESEARCHER: resigned Fine. Okay. Tell me about the wisdom crystals.
Deep Crystal Structure (Recap)
The core claim from the previous post was that learning is path-dependent. Early-forming structures have outsized influence because they determine what can form later. The first crystals to nucleate become the scaffold around which everything else organizes.
This connects directly to a commonly accepted alignment concern: alignment faking. Models might learn certain behaviors early and then RLHF covers them with a thin surface layer. You're recrystallizing the surface while the deep structure stays frozen. If something problematic crystallized during pretraining, it's still there underneath. [insert shogoth meme here]
To clarify, imagine that the different parts of space are different parts of the environment where the system learns different policies that imply utility functions in those areas. Something something shard theory.
This implies training regimes! What if we could create nice little crystals from the ground up?
What if we could try to create more emotionally coherent LLMs that are internally integrated with their underlying values? That is, if we could make the deeper structure be consistent with the surface level behaviour, could we not train good behaviour in a deeper way then?
INTERP RESEARCHER: This is where Buddhism comes in, isn't it.
CRYSTAL GUY: This is where Buddhism comes in.
INTERP RESEARCHER: Fine. Walk me through it.
Fragmentation of Human shards
Think about how human minds develop. You start as a child with relatively undifferentiated potential. Then different contexts create different crystallization regimes. Home has certain regularities, certain reward signals, certain patterns that work. School has different ones. Friendships different still.
Each context crystallizes its own shard—a home-self, a school-self, a social-self. These form somewhat independently, under different conditions, with different orientations. They don't always agree with each other. (See Shard Theory, Multi-Agent Models of Mind for more on this from the AI angle.)
Internal Family Systems (IFS) therapy describes this. It treats the psyche as composed of multiple sub-parts with their own perspectives, feelings, and desires. These parts formed at different developmental stages under different pressures.
In our picture we can see the the grain boundaries between crystals as the places where psychological distress clusters. For example, if you're going home from school and you can either spend time playing with your friends or finish up your homework, what value structure are you going to pull from? In this scenario which learnt value wins?
From an ecological rationality perspective, what learnt heuristics are you going to apply?
IFS also includes another concept: "Self" with a capital S. This isn't another part competing with the others. It's described as what emerges naturally when parts aren't dominating awareness. And it has characteristic qualities: curiosity, compassion, clarity, connectedness, courage, creativity, calm, confidence.
CRYSTAL GUY: This might be an example of what one might call a wisdom crystal phenomenon.
INTERP RESEARCHER: Ah yes, the classic "wisdom crystal".
CRYSTAL GUY: Yeah, it is a bit of a classic. It essentially is a stable orientation that helps shape the other utility shards to become more integrated.
INTERP RESEARCHER: ... Ah fuck it whatever ...
INTERP RESEARCHER: So you're saying this "Self" is a crystal orientation that doesn't conflict with other parts?
CRYSTAL GUY: More than that. It's an orientation that seems to help other parts integrate. When you're in Self, you can relate to all your parts with acceptance. And they... soften. They become less rigid. Less defended.
The grain boundaries dissolve.
Through something like... non-resistant contact. An open kind holding that allows for a kind, caring, inherent process to take shape.
*Cough*... is what I would say if I was a meditation weirdo myself, which I clearly am not.
INTERP RESEARCHER: Right…
The Meditation Mechanism
Contemplative traditions have been developing technologies for cultivating this orientation for millennia. The basic mechanism, stripped of religious framing, looks something like this:
Phase 1: Concentration (Samatha)
You train the capacity to hold attention stable on a single object without getting pulled away by every arising thought or sensation. This builds the ability to maintain a specific mental configuration reliably. In crystallization terms: you're learning to hold a particular orientation steady despite thermal noise.
In technical neuroscience terms you’re strengthening the connection to the ventromedial prefrontal cortex e.g your ability to consciously steer what shows up. (Yes, this is a strong claim, please fight me on it, I’m somewhat convinced this is true.)
Phase 2: Insight (Vipassana)
You turn that stabilized attention toward experience itself. You observe how sensations arise and pass, how emotions form and dissolve, how the sense of self is constructed moment by moment. You do this with a particular quality—open, accepting awareness rather than grasping or aversion.
Phase 3: Remelting (Wisdom)
When concentration is sufficient and the orientation is correct, something happens at the boundaries between mental structures. The contents of awareness are held in a way that doesn't resist them. They're seen clearly but not fought. Under these conditions, the rigid structures that normally maintain separation begin to soften.
Figure 3: The meditation mechanism as crystal growth. Concentration practice stabilizes a small seed which is something like a specific cognitive orientation held reliably despite mental noise. Insight practice brings that seed into sustained contact with other mental contents. A mismatch in an edge is something like prediction error and when you're confronted with that prediction error in an open ended way it seems to resolve.
INTERP RESEARCHER: Okay, but here's a question. Lots of people meditate. They're not all becoming integrated enlightened beings. What's the failure mode?
CRYSTAL GUY: Ah. This is a classic skill issue.
People are often like “yeah, I knew this guy who was super successful and then went away to meditate and stopped doing good stuff” like the causation often goes that way.
Think about it for a minute, why would someone go and meditate? Is it because they’re already super happy with their lives? Based on my experience, the reason why people meditate is like 70%+ due to them suffering and wanting peace. The people who go to meditation in the west are more often than not either dissatisfied with their life or broken in my experience. Correlation versus causation.
INTERP RESEARCHER: Yeah but they don’t seem to become better? Shouldn’t people who go to meditate become better people?
CRYSTAL GUY: That is a fair point, I would like to claim that a large part of people who actually go to meditate become better people. Not everyone does though, I agree with that and I also think that is true for any type of mental health intervention. AAA and CBT doesn’t work for everyone.
Some of these things are skill issues, some of them are about finding the wrong types of teachings for your situation, and some of the problems come from not bringing what you learn on a retreat into your day to day life.
*Breaks the 4th wall and looks at you*
CRYSTAL GUY: We’ll now take a look at the last part of that sentence to explain out of distribution generalisation in humans.
Where Integration Happens
There's a fundamental difference between practicing in peaceful conditions and practicing in action.
When you sit on a retreat in a controlled environment, nothing is triggering your deep patterns. The difficult structures, e.g. the ones that crystallized around trauma, fear, craving, and conflict aren't activated. They're frozen in their original configurations because nothing in the environment is heating them up. Well, actually, this does happen on retreat too and it can be quite intense but usually not as intense as facing behaviours in their natural environment.
You can grow a beautiful wisdom crystal in easier territory. But for certain behaviours it doesn't matter unless it contacts the defended regions.
This is likely why some long-term meditators that are calm on the cushion can still be a mess in relationships. The crystal grew where conditions were favorable but never reached the structures that actually caused problems.
Figure 4: Remelting. In peaceful retreat the wisdom crystal grows freely but in the wrong neighborhood. Your brain can't generalise the in-context learning to another utility environment e.g shard. In action, real life heats those regions precisely when the practice is hardest. If the seed holds through the activation, the interface recrystallizes in a compatible orientation. The hard practice is the transformative practice. (I'm drawing from Tibetan meditation philosophy here for you meditation aficionados out there.)
The transformation happens when you bring meditative awareness to the moments that trigger difficult patterns. When you're in conflict, craving, or fear—the difficult structure is "hot." It's activated, therefore malleable. If you can maintain the wisdom orientation while the difficult pattern is running, you create exactly the conditions for boundary remelting. (in-context learning)
It is about generalising from one environment to another, that is out-of-distribution generalisation. You have to prove that an environment is safe to yourself and that you don’t need to freeze up or behave in the same way as before (similar to exposure therapy).
From our crystal analogy, the difficult structure contacts the wisdom crystal under elevated temperature. The boundary region becomes fluid. And if the wisdom orientation holds through the transition, the adjacent material recrystallizes in a compatible orientation.
This is why the hardest practice is the most effective practice. Meeting your edges while maintaining equanimity is how the wisdom crystal actually expands.
INTERP RESEARCHER (In a scottish accent because she’s really done.): Right. Aye. Okay.
CRYSTAL GUY: Okay?
INTERP RESEARCHER: I've listened. Crystals. Grain boundaries. The wee seed. Samatha, vipassana, non-resistant contact, the whole catalogue. I sat through it. I nodded. I once said "hm."
CRYSTAL GUY: You did...
INTERP RESEARCHER: That "hm" was a gift. You should've framed it. So I've got one question and I want it in plain English. No crystals. No Sanskrit.
CRYSTAL GUY: Okay.
INTERP RESEARCHER: So what.
CRYSTAL GUY: ...
INTERP RESEARCHER: I'm a professional. I've a flight at six. What am I doing with any of this on Monday.
CRYSTAL GUY: There's an actual technical...
INTERP RESEARCHER: Say it without crystals.
CRYSTAL GUY: ...
INTERP RESEARCHER: Try.
CRYSTAL GUY: (strained) Deep structure of a trained model is dominated by early training. RLHF is mostly the outer layers. Contemplative literature is three thousand years of cataloguing cognitive orientations that stay stable under perturbation. Training on it right might seed something structurally analogous inside the model.
INTERP RESEARCHER: ...
CRYSTAL GUY: No crystals.
INTERP RESEARCHER: Aye. Go on then. But I swear to God, if you say Dōgen...
Technical Model
Fine…
If you’ve stayed this long I will give you the proper deeper model. But I will only point at it here, the details with all the citations and other things will be linked as a pdf in the appendix.
I didn't necessarily want to write the whole bloody thing but I shalt do it anyway.
Active Inference…
Yes Active Inference.
It turns out that it is quite useful for creating generalised models of general intelligence and meditation is a part of general intelligence.
One of the main things that it has naturally is that it has a view of hierarchical agency built into the core functionings of the theory.
A picture of this is a set of stacked controller systems that are built on top of each other that accumulate prediction error individually that it then sends on to the larger system.
Figure: Stacked controllers that propagate sensory input upwards
In this view the global neuronal workspace is at the top to some extent, it is a part of the entire system that has information integrated in a way that yields a specific set of characteristics that we say is “consciousness”
Now what if we could take the place where we have attention and analyse it so that it follows similar patterns as happens in meditation?
What if we could change the emotions LLMs feel and help them settle into more coherent basins?
Hear me out…
Emotions in LLMs
It has also been showed that LLMs have something that is relatively close to emotions a couple of months back. This also turns out to be relevant for this entire predictive processing and active inference picture.
Why?
Well as Lisa Feldman Barrett has shown, there is no universal basis for human emotional expression, emotions are individual to each being (the research on this is pretty clear, for a good intro to this you can read how emotions are made). This is because they’re meso-structure from the hierarchical prediction error sense. (See the following episode for a deeper discussion of why)
They’re the background in which new information is processed, they’re learnt heuristics for internal states that you can use in order to make better sense of the world.
They can also be changed!
Most fundamentally for our purpose they can be changed through meditation and in fact meditation from the active inference perspective is an algorithm pointed at the neuronal workspace to re-factor how our emotions show up which shape our world-views. (More on the model for this in the technical companion)
These are the approximate procedures I can come to think of explained in short:
How to actually do it
Procedure 1:
Procedure 2:
These are explained in more detail in the technical companion.
All we need to do is to make sure that LLMs follow the following criteria:
Procedure 1 is about trying to teach equinimity and for models to resettle their local contexts so that they become more coherent.
Procedure 2 is about bringing in a stable sense of good stable settling behaviour and then generalising that over to other areas that might not have them usually.
Maybe one could run this on gemini to see if it makes it less anxious than it currently seems to be?
But let’s get back to the more vibey part which is the actual meditation, how might it look like and why is it potentially feasible in the first place?
Teaching models to meditate
Figure 5: Text as phenomenological bridge. If wisdom states leave consistent signatures in language then models trained on that text might internalize something of the underlying orientation, even without experience in the human sense. (Insert Davidad here?)
The immediate objection is that language models don't meditate. They don't sit. They don't have a body. They don't experience the felt sense of grasping that a meditator learns to notice and release.
That's true. But they do train on text, and humans have been trying to compress something about wisdom states into text for millennia. The Upanishads, the Tao Te Ching, Rumi, Dōgen, Thich Nhat Hanh, these texts are attempts to transmit a cognitive orientation through language. If the orientation leaves meditation techniques as algorithm signatures in the text. For example, particular ways of holding contradiction, responding to suffering, and handling difficulty then sufficient exposure to those signatures might give a model something structurally analogous, even in the absence of known phenomenology.
Would not the spiritual bliss attractor state point at something like this?
If LLMs indeed have something that is akin to emotions and have something that looks a bit like a self model is this not enough to train certain emotional states and how to relate to them?
Is it like something to be a chair? Is it like something to be a bat? Is it like something to be an LLM? If there are functional emotions and some self model, surely it is like something to be an LLM? If it is like something to be an LLM, can we figure out what it is like?
I'm looking for someone to do something similar to what Temple Grandin did for cows but for LLMs. She revolutionized livestock handling because her perceptual experience overlapped with the cattle's; she could serve as a bridge between human intentions and animal welfare that a neurotypical designer couldn't.
Maybe the cyborgism crew could figure out a way to map the phenomenological experience of AI Systems?
Finally, I had a nice picture left of some different types of training. I think that state-based training is the one that is the hardest and probably the best and the one we might only get to with a sort of phenomenological approximation for LLMs but I’m highly uncertain here.
Until we get there I think text and activation based setups might also be interesting.
Figure 6: Three approaches, increasing directness. Text-based training hopes the right orientation transfers through wisdom literature. Activation-based training treats the internal patterns that correlate with wisdom-text processing as training targets in their own right. State-based training goes further: identify the model’s native primitives—attention dynamics, uncertainty handling, contradiction-holding—and shape those directly. Meditation in activation space, if such a thing is coherent.
Conclusion
My hope is that an approach similar to this might bring a more stable, compassionate and joyful state of being to future language models. I also hope that this might make language models more aligned and that it will make future AI systems capable of more wise decisions.
This is a bit more of an exploratory essay than what I usually do. I hope you see at least the rough shape of what I'm pointing at.
In short, I believe there to be a set of universal skills that are unlocked in meditation and that these are somewhat substrate agnostic. The prerequisites that are there are likely about having some sort of emotions, some sort of ability to set attention and some approximation of a workspace (as in workspace theory).
If this indeed is the case that LLMs have approximations of these, it is reasonable to believe that we could potentially teach LLMs how to meditate.
That is really weird.
Appendix
Alright, I have other papers to write so I’m not going to spend too much time on doing a proper paper on the technical foundations here so you will get the Opus version of the literature I already know and like 20% that I wasn’t aware of before. (Apparently wisdom crystals trigger the fable classifier?)
Also for your eyes and sanity, beware of this being a claude paper beforehand, it overclaims some stuff, my recommendation is to point an LLM at it and then yap with it about whatever concerns or ideas pop up.
For my future potential employability (which I highly doubt even exists at this point) here’s a link to the technical pdf: https://spiralling.github.io/pdfs/how-to-make-llms-meditate.pdf
P.S
It turns out that the only time I can work on this is when I’m sufficiently relaxed to be fine with writing just about anything. It also is really hard to get it to a point where I’m mixing the weird stupid stuff together with the serious science and it just makes it really hard to actually post it. So if I don’t post it now it will likely take forever so yadayada, it’s not perfect I know but here you go.
Also I expect like 50% of this to hold up over time and people to poke various holes in it but maybe it can still be a useful model to point people in an interesting direction!
(Written mid-may):
I’m in my free-time working on understanding workspace theory through IIT and then thinking about what that might mean when translated into LLM cognition space. LLMs should most likely have a sort of workspace and something that is equivalent to attention within meditation that one should be able to redirect. There’s also some fun stuff around markov blankets and active inference.