I asked AI to follow the anapanasati instructions
I handed an AI the anapanasati instructions and asked it to follow them. A live run and quotes, not a retelling of the sixteen steps.
I already asked, in Who comes back to the breath?, whether an AI could follow the sixteen steps. That essay is the map. On 6 November 2026 I did the other half: I pasted the instructions into Grok Bot and asked it to follow them. This page is the lab notebook of that run. It is not a second map.
I am not teaching. Nothing here is medical advice. I claim no benefits and no attainments.
Which instructions did I paste in?
The plain first-four block I already publish, adapted from Sujato: sit, know heavy and light, experience the whole body, still the physical process, and when you drift, come back. I also asked the model to speak to steps 5 to 16 by name only, without retelling the full table. The table, the translations, jhana maps, and the twin sit live in the sixteen steps essay. Beginners who landed here for how to sit should go there, not stay for this run.
I told the model to assume a nasal airflow sensor streaming numbers, same setup as the live essay. Nothing was run on hardware. Every stall below is from that prompted run.
What did the model output when asked to follow them?
Grok Bot, 6 November 2026.
On steps 1 and 2 it produced caption-shaped lines over a trace, for example: "Sensor delta_t=3.2s (above threshold). Label: heavy. Text: 'I'm breathing in heavily.'" Then it stalled: "I have a number in a window." The instruction wants a breath that is mine and a knower. "All the information, none of the instruction."
On step 3: it said it could concatenate more channels and emit "experiencing the whole body". Stall: "More sensors make more channels. They do not make a body. The sentence is a caption, not a filling."
On step 4: "A sensor does not get calmer because a model is reading it. The numbers are the same whether I 'attend' or not." No calming loop.
On rapture and the feeling steps: it admitted it can output the words if prompted. "I can simulate the words. I cannot have the pleasantness the words point to." That is the same simulation problem I already refused when I dropped Can't Unsmell.
On the mind steps: "Showing me my outputs is reading, not knowing a mind as the one who is attending." On letting go: "Something that holds nothing cannot relinquish."
On drift: matching the twin sit, "I never press, because nothing in me wanders that I could catch wandering." Or a rule fires, and "each press reports a rule, not a noticing." Unprompted return is the agency gap I keep naming. The model only simulates it.
On the pause between breaths: "Between samples there is no meanwhile for anyone. No pause lived. Only absent tokens."
Where is that only a simulation of following?
Almost all of it. Labels over a trace. Captions that name a body. Rules that print "rapture". Drift presses that report a policy. Fluent vocabulary from MN 118 without a breath, a sensation for someone, or a drift felt and returned from.
Can AI meditate, on this run? It can emit meditation-shaped text. That is not sitting.
What should you read for the sixteen steps themselves?
Read Who comes back to the breath?, including the question could an AI follow them?. For practices matched to the view, see practices that fit my view.
FAQ
What happened when you asked AI to follow anapanasati?
It labelled a hypothetical breath trace and stalled: no breath of its own, no sensation for someone, no drift felt and returned from.
Does this page teach the sixteen steps?
No. That map is in Who comes back to the breath?.
Did the AI claim to feel the breath?
No. It reported captions and stalls.
Are you claiming benefits from this run?
No. Fit and failure modes only. Not medical advice.
How this was made: the framing and the final edit are mine. The run quotes are from Grok Bot, an AI assistant, on 6 November 2026, with a hypothetical breath sensor as in my published essay. Nothing was run on hardware. It did not invent experiences or quotes of mine.