a simple mechanistic theory of jhanas

Jhanas seem like magic. We need some kind of scientific theory for how they work. Here's an attempt at providing a well-grounded one.

A couple of initial notes. Firstly, I think jhanas are awesome, but they aren’t necessarily the best practice to do at a given point for a given person. Practicing them tends to move you along the path of meditative insight, which is not something everyone wants, and which can be temporarily life-disrupting. Secondly, the ideas presented here are hypotheses about what might be going on under the hood with jhanas, not confirmed at all. Although jhanas have been observed in lab conditions to activate reward circuitry, and although this story makes a lot of sense, we don’t know if it’s true yet. Thirdly, I want to emphasize that this theory is not required to practice jhanas well. Every practitioner of jhanas in history would be a counterexample to that idea. And so, if you find yourself reading and thinking “This isn’t working for me”, I encourage you to stop reading. If you’re the type of person for whom theory more often than not becomes a roadblock to good practice, I encourage you to consider not reading this. Finally, thanks to Nick Cammarata and Sasha Chapin for their thoughtful suggestions and edits.

A very simple model

The first three jhanas are meditative states that feel good in the way that sex or drugs tend to make you feel. With the first, you experience primarily a lot of bodily pleasure with noticeable joyfulness. In the second, the body cools down slightly and you mainly have an experience of significant joy. In the third, the excitement and energy of the first two lowers a lot and you end up in an ultra-contented state, an unmatched peaceful flavor of happiness. There are nine jhanas in total, but here I want to focus on how these first three might work since the fourth through ninth seem to work in a slightly different (though related) way.

These first few jhanas can be explained using the framework of reward prediction error (RPE). An ultra simplified model¹ of reward prediction error might look like this:

\(RPE(t) = R_{actual}(t) - R_{expected}(t)\)

If you get more reward than you expected, you have that prediction error is positive. When prediction error is positive you feel a sense of reward. When you get less reward than you expected, you feel a sense of negative reward.

Let’s assume that both of the terms on the right-hand side are defined in such a way that actual reward increases or decreases at the next time step in part as a function of whether the reward prediction error is positive or negative, and expected reward is defined in a way that attempts to minimize prediction error. Note that although it’s a bit more detailed in practice, this type of model is well-substantiated in neuroscience literature as the way neural reward circuitry seems to do its thing.

Jhanas as a reward prediction jailbreak

So, when you practice jhanas there are several things you typically do to make them work. Below are the instructions for the first jhana, but they are also roughly the core steps for working with the second and third. They’re each slightly different in terms of which sensations you focus on and try to amplify, as well as what kinds of friction there is going in, but are in essence the same. To move from one to the next you’ll typically wait until the current one is stabilized and then repeat these steps.

1. Be comfortable. Reduce any “hinderances” going on (sensual desire, ill will, sloth and torpor, restlessness and worry, and doubt).

2. Get into “access concentration”. Focus and clear your mind, reduce other distractions.

3. Notice a pleasant sensation in your body, like a warm tingling in your hands

4. Don’t grab onto the pleasant sensation.

5. Let it, your mind, and your body do their thing. Don’t grab onto the pleasant sensation.

6. Pleasant feeling intensity and depth grows, i.e the jhana intensifies. Don’t grab onto the pleasant sensations.

7. Go back to (3)

What if we reframe this under the reward prediction model? In those terms, the way to do first through third jhana practice looks something like this:

1. Lower your expected reward to a number smaller than your actual reward. Naturally, it is easier to do this when the actual reward is lower, so calming down the whole system and being comfortable is good because it brings the actual reward to be near zero or just slightly positive.

2. Sharpen your mind to the point that you can reduce noise on the actual and expected, but especially the expected, reward function. This is something like single-threading your reward circuitry, and carefully observing what happens on that thread.

3. Find a clean source of actual reward, and let it in. Should be simple, not noisy. Lower reward magnitude can be easier to manage at first because of the feedback loop and the next step.

4. Do not increase your expected reward in response to the good feelings! If you feel that come up, let it go.

5. Reward prediction error is now increased. Do not increase your expected reward!

6. Actual reward increases. Do not increase your expected reward!

7. Go back to (3)

Overall in this picture, what’s going on is that we set up the reward system to be free of noise, we then dampen or zero out the expectation of any reward (i.e. reduce clinging/craving/tanha with respect to wanting to feel really good from the nice sensations we’re experiencing). Then as we gather small rewards, our actual reward estimate increases in proportion to the previous reward amount because we aren’t expecting anything. If we can skillfully maintain this by minimizing expectations as well as keeping noise and other sources of negative reward at bay, this directly results in a feedback loop of good feelings.²

This answers several questions

This is a nice direction of explanation because, if true, it explains things that are pretty hand-wavy and mysterious about jhanas:

1. First of all, how the hell does this magic trick work? Is it fake? Jhanas, the controversial meditation technique that broke slatestarcodex? Seems fake bro. Here we have an explanation that matches existing neuroscience decently well. And it clearly explains the infinite happiness hack which has such a tendency to fascinate and frustrate.

2. Why can many people get the right sensations going for the first jhana but then have trouble stabilizing it? Under this framing, the answer is either you have noise on the actual rewards channel (e.g. you’re getting negative reward estimates from some non-jhanas source), or you end up expecting more reward than you’ve generated at a given step and thus slow, dampen, or kill the feedback loop.

3. Why don’t people get super addicted to jhanas? I’m more uncertain about this one, but here are some guesses. First, if the way you get a huge reward is by practicing turning off your expectation for rewards, and if the only way to obtain the state is to turn that off, then you naturally learn to not do that. This dampens the positive side of addiction behavior shaping (i.e. the drive to go do a certain addicting behavior). Second, a part of the mechanism that drives serious addictions is the negative reinforcement—the comedown, the loss at the poker table, etc. I.e. we’re driven to learn to avoid the bad feelings. This is in part mediated by negative reward prediction errors. To get negative reward prediction error in the model presented here you must have your expected reward be greater than your actual reward. But with the definition we gave above this mostly doesn’t happen during normal jhanas practice. It’s far more likely that someone craves a good feeling and kills the feedback loop before it feels like anything more than warm fuzzies in their stomach. And this isn’t likely to result in a large negative reward, dampening the other side of the addiction behavior shaping.

4. Why do people master the jhanas and then pretty much forget about them? The theory here makes this seem obvious — you stop expecting rewards because you practice how to stop expecting them, over and over. This matches the great explanation (which as far as I know was first put forward by Nick Cammarata), that practicing jhanas creates an abundance mindset around the resource of feeling good. Once you get skilled enough that you can make extremely pleasant, top 1% or better feelings happen any time, you don’t feel like you need to grab at every little thing in your life to feel better. This seems right to me (although there are other reasons). 

5. Why are the first three jhanas so different from the others? As far as I can tell moving from the third to the fourth jhana is precisely when things shift from dealing with reward circuitry into dealing with deeper manifestations of prediction and expectations. A bit more on this in the next section.

6. What does it mean for jhanas to “prepare the mind for insight”, as many people like to say? Why is it that traditionally we do insight meditation after exiting fourth jhana, and not after second, or third? This theory suggests an explanation: the lower three jhanas (at least temporarily) get you off the ride of caring so much about how good you feel. This frees the mind up to attend to things deeper in the stack. It unlocks and eases the decoupling from our inner and outer environments that happens during higher jhanas.

What about the other jhanas?

As I mentioned above, this model best matches the first three jhanas. What seems to happen when you’ve gone through those is that your overall reward system is able to settle down in a stable equilibrium without needing to pump rewards in and amplify them, without needing to manually tune things. At that point, you aren’t expecting any rewards, and you aren’t getting any rewards — positive or negative. And of course, by this point, you’re in a stable, low-noise equilibrium.

From this perspective, the first few jhanas are a reward circuitry jailbreak which enable us to turn down the noise on the thing (our reward circuitry) which drives so much of our go-to or default tendencies for what we think about, feel, and do.

My intuition about what this does is that it creates space. We’re often preoccupied with making predictions about which rewards we’ll get and living in that loop of observing, planning, and obtaining or not obtaining rewards. And the feeling of turning down the expectation of reward is a lot like the feeling of turning down expectations in general. Getting good at that is probably a transferable meta-skill. And clearly, part of what’s going on with all the chatter in the brain is its incessant need to predict stuff all the time.

This chattering results in expectations of various kinds. And expectations stabilize perceptions. When you’re in this quieted, smooth state, and you’ve been practicing how to turn down your expectations of reward, to reduce your clinging, and you’ve reduced (or zeroed) the number of online reward prediction threads, you can more easily begin to turn down your prediction machinery around other things such as emotional gradients (corresponding to jhana four), boundedness of space (jhana five), boundedness of mind (jhana six), the existence of things that aren’t nothing (jhana seven), the expectation that the perceptual apparatus must have an object of some kind (jhana eight), and finally the expectation of consciousness itself (sometimes called ninth jhana, also called nirodha samapatti).

A few final thoughts

There are many experiments we could imagine running to test this class of hypotheses, I may say more on this in a future post. And as for theory, there’s more left to uncover here. For example, in the usual predictive processing view, you can think of the deconstructive parts of meditation as essentially loosening up our priors over various aspects of our usually fairly firm experience of the world. Where insight into the macro concept of a “self” eventually provides a clear picture that maybe this is just a thing we’re used to instead of a real, hard, eternal object that always exists. Eventually, you can do this with space, time, and the rest of everything in your world model. In some sense, the degree to which you have sustained meditative insight into experience is equal to the degree to which you have stably loosened your priors about the various objects of your world model (by which I mean something like the generator of the totality of your experiences as a sensing being). An obvious possibility here is that reward prediction error framing might be part (though certainly not all) of why enlightenment itself makes people feel good in general. Having relatively lower expectations about everything is quite plausibly good for experiencing more sustained rewards.

Where research into meditation and predictive processing is lacking though is how this whole process works at a granular level. How do we build up the capacity to do this deconstruction? How do we learn to so finely tune our experience that we get all of these wide-ranging and generally great effects from meditation? And so on. I think the theory of jhanas presented here starts to show why jhanas are so helpful in this deconstruction process.

Many of us are supremely attached and attuned to the ongoings of our reward circuitry. From birth we ride the ride around and around, desperately scraping reward pellets up wherever we can get them. Lower jhanas are such a great foundational practice in part because they show us that:

1. There is a ride.

2. We can get off the ride. It’s actually possible.

3. We’re safe to get off the ride. It won’t even hurt!

4. When we start playing with getting off the ride, we feel better.

5. Actually, wait, getting off the ride feels amazing, way better than that stupid ride ever felt.

From there we are free to go further, building up skills during higher jhanas practice that help us let go of very deeply assumed parts of our normal existence. Things that are so deeply rooted we’ve never seen them before, yet which strongly determine the character of our daily experience. In this way, jhanas are a remarkably solid foothold on the often confusing path to snapping out of the dream. And so, though they’re worth understanding better in their own right, it’s clear that understanding jhanas better also helps us improve our theories of how to navigate and understand the path of meditative insight itself.

bayes

1. I’d like to emphasize that although this is an extremely simplistic model, I nonetheless think gets to the heart of something, and so I chose to use it here in an effort to not get bogged down in explaining biology and math, tracing through more detailed mechanisms and showing how they relate to jhanas, and so on. In a later post, I’ll explain how I think, for example, temporal difference style models of reward prediction error fit in here. There is quite a lot more to say.

2. Obviously, there are a huge number of assumptions in this picture, for example, to what degree can we actually top-down manipulate the expected reward calculations going on deep in our reward circuitry?