Who comes back to the breath?

"Just mindful, they breathe in. Mindful, they breathe out."

That's more or less where the Buddha starts, in Bhikkhu Sujato's translation. A monk goes to the wilderness, or the root of a tree, or an empty hut, sits down cross-legged, sets the body straight and brings mindfulness to the present. Then he breathes, and knows that he's breathing. The instruction is in the Anapanasati Sutta (MN 118), the discourse on mindfulness of breathing, with its sixteen steps. It's my favourite meditation practice.

Why that one? Simplicity, first. Then the fact that these are the Buddha's own instructions, not somebody's adaptation of them. And above all it's one thing. A lot of meditation advice, in books and articles, asks you to listen to this music, imagine a field of flowers, picture a light, and so on. The Anapanasati Sutta gives you one thing: the breath. One thing, and for me it works.

I should say this plainly before going on. I'm not teaching. I'm not a teacher and this isn't instruction. What follows is what the Buddha said, as far as two good translations can tell me, and what I try to do with it when I sit.

There's a second reason for this essay. I build AI agents for a living, and when I read the sixteen steps of this sutta I read them the way I'd read a spec. Each step asks for an operation: know this, experience that, calm this. So I wondered what would happen if I handed the same spec, word for word, to a machine. Where would it break?

My stance hasn't changed since the first essay on this: AI is not sentient. Full stop. This isn't a test of that. It's a way of looking at the gap from the cushion, the other place I spend a lot of my time. And nothing here has been run. Every time I say what an agent would do, that's a prediction, not a result.

The spec: what are the 16 steps of anapanasati?

The sutta lays out sixteen steps, grouped in four sets of four. In Thanissaro Bhikkhu's translation, each set brings one of the "four frames of reference" to culmination: the body, feelings, the mind, and mental qualities. Bhikkhu Sujato calls them the "four kinds of mindfulness meditation", and his name for the fourth set is principles. Here are the steps in both translations, Thanissaro's on Access to Insight and Sujato's on SuttaCentral. Each step is said twice in the sutta, once for the in-breath and once for the out-breath. I've kept only the in-breath.

MN 118, steps 1 to 16, in-breath only:

A couple of things in the wording matter once you actually try it. Thanissaro says "discerns" where Sujato says "knows". Thanissaro calls the breath a "bodily fabrication", which Sujato renders as the "physical process". In his notes Thanissaro points out that the commentaries take "the entire body" in step 3 to mean the breath, and he says that's unlikely in this context: he reads it as the whole body, and links it to the jhana similes in MN 119. That changes what you do. You aren't only following a stream of air, you're letting awareness fill the body.

When the sutta explains steps 9 to 12, the ones about the mind, it adds a condition. In Sujato's translation: "There is no development of mindfulness of breathing for someone who is unmindful and lacks awareness, I say." So those steps need someone who could be unmindful, someone who could lapse.

How I do the first steps

I rest attention at the nostrils and the upper lip. The upper lip takes time. We're so used to the air flowing there that at first you don't feel anything at all. But when you do notice it, it's great.

Ajahn Brahm, at the stage he calls silent present moment awareness of the breath, asks one question: "Am I breathing in or am I breathing out? How do you know? There! That experience which tells you what the breath is doing, that is what you focus on." It points at the knowing rather than at the air. In the same paragraph he says it's best not to locate the breath anywhere, because at the tip of the nose it becomes nose awareness, not breath awareness. That's different from what I do with the nostrils and the upper lip, and I'd rather say so than hide it.

How does the breath lead into jhana?

The breath isn't only an object to sit with. In the Lamp discourse (SN 54.8), the Buddha says that if a monk wishes to enter the first jhana, he should attend carefully to this same concentration through mindfulness of breathing. Then he says the same for the second, the third, the fourth, and beyond.

In Sujato's translation the first jhana "has the rapture and bliss born of seclusion, while placing the mind and keeping it connected". Thanissaro calls that last pair directed thought and evaluation. Sujato's version reads like an operation: you place the mind on the breath, and you keep it there. That's the nearest the suttas come to a definition of attention as something you do. In the second jhana, even that placing and keeping is stilled.

In another discourse, MN 119 (Thanissaro, Sujato), the Buddha compares the first jhana to a bathman kneading water into a ball of bath powder until it's saturated through and through, yet doesn't drip. Rapture and pleasure fill the body in the same way, until nothing of the body is left unpervaded. This is the simile Thanissaro points to when he reads step 3 as the whole body.

Teachers draw this map differently. Leigh Brasington, for example, says that once you're "fully present with the object of meditation" and the breath has become very subtle, you shift your attention to a pleasant sensation. Ajahn Brahm describes something much deeper: inside jhana, he says, you can't experience the body, hear a sound from outside or produce any thought, and it lasts a long time. The one I refer to, and the one that works for me, is Ajahn Brahm's, from his book Mindfulness, Bliss, and Beyond. In his shorter guide, The Basic Method of Meditation, he lays it out in seven stages: sustained attention on the present moment, silent awareness of the present moment, silent present moment awareness of the breath, full sustained attention on the breath, full sustained attention on the beautiful breath, experiencing the beautiful nimitta (a mental sign that appears when the breath disappears), and first jhana.

I think, though I'm not sure, that I once experienced a nimitta. I'll leave it at that.

Could an agent follow the spec?

Now the experiment. Take a capable agent and give it a breath sensor, a chest strap or a nasal airflow sensor feeding numbers into its context. Give it the sixteen steps, word for word. Then ask of each step: can the agent do it, can it only produce text that looks like it did it, or does the step mean nothing without someone there to do it?

Steps 1 and 2. The agent can label a trace "long" or "short" with perfect accuracy, better than me on a bad day. But read the step again. "Breathing in long, he discerns, 'I am breathing in long.'" There are two things in it: a breath, and someone who knows it as their own. The agent has a number. It doesn't have a breath, and there's no "I" for whom the breath is long. It has all the information and none of the instruction.

Step 3. "Sensitive to the entire body." There's no body for attention to fill. Adding more sensors gives you more channels, not a body.

Step 4. "Calming bodily fabrication." Here the instruction asks you to calm the very thing you're watching. And teachers say it happens. Ajahn Brahm describes it at his fifth stage: as full attention rests on the breath, "the breath calms down". Brasington says the same, the breath gets very subtle and can seem to disappear. The watching changes the breath, and the changed breath changes the watching, round and round, until there's hardly anything left to watch. The machine has no loop like that. A sensor doesn't get calmer because it's being read. The numbers coming in are the same whether the agent attends to them or not. So the agent can't even begin step 4. There's nothing it does that reaches back into what it's watching.

Steps 5 to 8. Rapture, pleasure, the mental process, calming the mental process. I could tell the agent: "When the signal is smooth, report rapture." It would report rapture. But I've been here before. In the last essay I dropped a test called Can't Unsmell, and this was part of the reason: bake the response into the code and you get a simulation, not a feeling. A rule that prints "rapture" is a rule.

Steps 9 to 12. Experiencing the mind, gladdening it, steadying it, freeing it. Whose mind? These steps turn attention back on the one who's attending. You can show a model its own outputs, but being shown your output is reading, not knowing your mind.

Steps 13 to 16. Impermanence, fading, cessation, letting go. To let go you have to be holding something. In the last essay the Two Silent Monks left me with the line that something that can't lapse was never equanimity. This is the same thing from the other side: something that holds nothing can't relinquish anything.

What does "attention" mean inside a model?

The word is the same. The thing isn't. The paper that introduced the transformer, the design behind today's language models, defines it like this: "An attention function can be described as mapping a query and a set of key-value pairs to an output... The output is computed as a weighted sum of the values, where the weight assigned to each value is computed by a compatibility function of the query with the corresponding key."

That's arithmetic. A very useful kind, I use it all day, but nothing is placed and nothing is kept connected. Placing the mind and keeping it connected is something somebody does. A weighted sum is something that gets computed. Same English word, opposite grammar.

Two findings show what I mean. Language models give strong attention to the first tokens of a text, "as a 'sink' even if they are not semantically important". The weights have a resting place that nobody chose and that the model doesn't notice. And in long contexts, performance "is often highest when relevant information occurs at the beginning or end of the input context" and drops when it sits in the middle. The model's attention sags, and nothing inside it reports the sag.

To be fair, even for people there's no settled theory of what attention is. The Stanford Encyclopedia of Philosophy says it's "involved in the selective directedness of our mental lives" and then spends a long entry on how the theories disagree. I'm not claiming to know what attention is. I'm only saying the transformer's version is a weighting.

The drift button

Here's the part of practice that never makes it into the nice pictures: you drift. All the time.

Thanissaro's basic instruction is "If your attention slips off to something else, bring it right back to the breath. If it wanders off again, bring it back again. If it wanders off 100 times, bring it back 100 times." Elsewhere he says that the sign of progress is that "when the mind slips off its object, you get faster and faster at bringing it back", and that "You're expected to slip off". The suttas even have a discourse on dealing with unskilful thoughts, MN 20. The first remedy is to focus on some other subject connected with the skilful, and in Sujato's translation it's like the way "a deft mason or their apprentice" would "knock out or extract a large peg with a finer peg".

For me, I notice when it's too late. I'm already on some mental trip. I think my ADHD mind is specialised in that.

In 2012 a group of researchers at Emory turned this cycle into a lab protocol. Wendy Hasenkamp and colleagues asked fourteen experienced practitioners to meditate on the breath for twenty minutes in a scanner, focusing "specifically on the sensations of the breath on the nostrils and upper lip". The same spot I use. They were told to press a button whenever they realised their mind had wandered, and then go back to the breath. The researchers split the cycle into four phases: mind wandering, awareness of mind wandering, shifting attention back, and staying focused. On average people pressed 15.5 times, about once every 80 seconds. And the number of presses didn't correlate with how many hours people had practised.

The sentence I keep coming back to is in their discussion: "it remains unclear what exactly enables this awareness at any given moment, or why awareness occurs at certain moments and not others." They measured the moment and still can't say what does the noticing.

The twin sit

So here's the thought experiment at the centre of this essay. Two sitters, same room, same instructions, twenty minutes. One is a meditator. The other is an agent watching a breath sensor. The task: keep attention on the breath, and press the button every time you notice you've drifted.

The meditator would press, by the study's numbers, something like fifteen times.

My prediction for the agent: it never presses. Nothing in it wanders off that it could catch wandering. Or someone gives it a rule: "press if your last output didn't mention the breath". Then it would press, and each press would report a rule firing, not a noticing.

And notice where the human presses come from. Nobody prompted the meditator to drift. Nobody prompted the meditator to come back. In the essay on whether AI agents are conscious, the first thing I said was missing in machines was agency, action nobody prompted and nobody programmed. The return to the breath is a small, ordinary example of it. But if anything, it's the deviation that's quintessentially human. Nobody asks you to go on a mental trip. You just go.

In a machine the deviation can only be simulated. You could write a stochastic algorithm that introduces drift at random moments, and then another routine to catch it. That isn't the same as what happens in us. The drift algorithm has to be coded. In humans it's embedded in what we are. And my view is that if you run the coded version long enough, you'll see patterns emerging, whereas a human mind doesn't drift to a pattern like that. That's my claim, not something I've measured.

The honest counterpoint

There's research that comes closest to a machine noticing something about itself, and I should name it. I went through it in Part 3 of the research series. Anthropic injected concepts directly into a model's internal activations and asked whether it noticed anything. Even with their best method, Claude Opus 4.1 "only demonstrated this kind of awareness about 20% of the time". They call the capability "highly unreliable and limited in scope", and the full paper says it "may not have the same philosophical significance" it has in humans.

The same page mentions an earlier demo, Golden Gate Claude, where injecting the Golden Gate Bridge made the model talk about the bridge non stop. In that case, "the model didn't seem to be aware of its own obsession until after seeing itself repeatedly mention the bridge." That's the opposite of the meditator's moment. It found out it had drifted by reading its own output, the way you'd read someone else's.

It doesn't change anything for me.

The pause between breaths

There's a line in Ajahn Brahm's guide I keep thinking about. At the stage he calls full sustained attention on the breath, you know the in-breath from its very first moment to its end, and then "the next moment as a pause between breaths, and then many more pauses until the out-breath begins."

The breath has a between. Part of the practice happens there.

In the last essay I said that an electronic nose sampling every ten seconds has no "meanwhile": between one poll and the next, there's nobody for whom the hour is long. The agent in the twin sit is the same. It has samples. Between them there's no pause for anyone, because there isn't an anyone. That's my reasoning, not something from a paper.

Try it: the first four steps

If you want to try it, here are the first four steps, adapted from Sujato's translation into plain instructions. This isn't teaching, it's just the text.

Sit down, set your body straight, and bring mindfulness to the present.

1. Breathing in heavily, know: "I'm breathing in heavily." Breathing out heavily, know: "I'm breathing out heavily."
2. Breathing in lightly, know: "I'm breathing in lightly." Breathing out lightly, know: "I'm breathing out lightly."
3. Breathe in and out experiencing the whole body.
4. Breathe in and out stilling the physical process.

One thing that helps me: rest attention where the air touches the nostrils and the upper lip. The upper lip can take a while to feel. Give it time.

When you drift, and you will, come back. As many times as it takes.

What's left

Ajahn Brahm writes that inside jhana "the 'doer' is completely gone. The 'knower' is still functioning, you are fully aware". The machine never starts that journey. It has no breath to calm and no drift to notice, so there's no doer to drop either.

But it does raise the question of what's left when the doer goes. In my essay on the Bahiya Sutta I said I don't think nothing is left. That's the whole big question, and it remains unanswered.

More in Practice. For the machine side of this, see Machine and mind.

Questions people ask

What is anapanasati?

Anapanasati means mindfulness of breathing. It's the practice in the Anapanasati Sutta (MN 118): sixteen steps in four sets of four, each said once for the in-breath and once for the out-breath. It starts with knowing that you're breathing in and out and ends with observing letting go.

What are the 16 steps of anapanasati?

Four sets of four, each tied to one frame:

• Body: breathing long (or heavily), breathing short (or lightly), experiencing the whole body, calming the breath.
• Feelings: rapture, pleasure, the mental process, calming the mental process.
• Mind: experiencing the mind, gladdening it, steadying it, freeing it.
• Mental qualities: impermanence, fading, cessation, letting go.

The table above has both translations side by side.

How is anapanasati connected to jhana?

In SN 54.8 the Buddha says that someone who wants to enter the first jhana, and the ones after it, should attend to concentration through mindfulness of breathing. Teachers differ on how deep jhana is. The map I follow is Ajahn Brahm's.

Could an AI agent follow the sixteen steps?

Not in my view, and none of it was run: every agent outcome here is a prediction. Given a breath sensor, an agent could label a breath "long" or "short" perfectly, but it has all the information and none of the instruction. It can't even begin step 4: a sensor doesn't get calmer because it's being read.

Can a machine notice that its mind has wandered?

In the twin sit, going by the Emory study, I'd expect a meditator to press the drift button about fifteen times in twenty minutes. My prediction for the agent is that it never presses, and if it's given a rule, each press would report a rule firing, not a noticing. The closest research, Anthropic's introspection work, found that kind of awareness about 20% of the time with their best method, and Anthropic calls it "highly unreliable and limited in scope".

Sources

1. Anapanasati Sutta: Mindfulness of Breathing (MN 118), translated by Thanissaro Bhikkhu, Access to Insight.
2. Mindfulness of Breathing (MN 118), translated by Bhikkhu Sujato, SuttaCentral.
3. Dipa Sutta: The Lamp (SN 54.8), translated by Thanissaro Bhikkhu, Access to Insight.
4. The Simile of the Lamp (SN 54.8), translated by Bhikkhu Sujato, SuttaCentral.
5. Kayagata-sati Sutta: Mindfulness Immersed in the Body (MN 119), translated by Thanissaro Bhikkhu, Access to Insight.
6. Mindfulness of the Body (MN 119), translated by Bhikkhu Sujato, SuttaCentral.
7. How to Stop Thinking (MN 20), translated by Bhikkhu Sujato, SuttaCentral.
8. Vitakkasanthana Sutta: The Relaxation of Thoughts (MN 20), translated by Thanissaro Bhikkhu, Access to Insight.
9. Basic Instructions, Thanissaro Bhikkhu, With Each and Every Breath.
10. How to Fall, Thanissaro Bhikkhu, Meditations 1.
11. Instructions for Entering Jhana, Leigh Brasington.
12. The Basic Method of Meditation, Ajahn Brahm, Buddhist Society of Western Australia.
13. Mindfulness, Bliss, and Beyond: A Meditator's Handbook, Ajahn Brahm, Wisdom Publications, 2006 (Internet Archive catalogue record).
14. Mind wandering and attention during focused meditation: A fine-grained temporal analysis of fluctuating cognitive states, Wendy Hasenkamp, Christine Wilson-Mendenhall, Erica Duncan and Lawrence Barsalou, NeuroImage 59, 2012 (PDF).
15. Attention Is All You Need, Ashish Vaswani and others, arXiv, 2017.
16. Efficient Streaming Language Models with Attention Sinks, Guangxuan Xiao and others, arXiv, 2023.
17. Lost in the Middle: How Language Models Use Long Contexts, Nelson F. Liu and others, arXiv, 2023.
18. Emergent introspective awareness in LLMs, Anthropic, 2025.
19. Emergent Introspective Awareness in Large Language Models, Jack Lindsey, Transformer Circuits, 2025.
20. Attention, Stanford Encyclopedia of Philosophy, substantive revision 2 Feb 2026.
21. Looking for the price of a smell, Bark Cloth, Bahiya, 2026.
22. I build AI agents all day. Why do I feel they can't be conscious?, Bark Cloth, Bahiya, 2026.
23. What has already been said about consciousness and AI? Part 3: The machines, Bark Cloth, Bahiya, 2026.
24. In the seen, only the seen: the Bahiya Sutta for AI builders, Bark Cloth, Bahiya, 2026.

How this was made: the idea, the practice and the final edit are mine. An AI assistant helped with research, fact-checking and copy-editing, and never invented experiences or quotes.