I wrote this ahead of a meeting between Buddhist teachers and people who build AI. The readers I have in mind are the builders. I want to hear from them where it holds and where it breaks.
I wrote it with Claude. The Buddhist claims are mine. The restatements in machine learning terms were made in the back and forth, and I checked each one. In section 5 I say that an exchange is the act of two parties. This essay was made in one.
One premise first. We have already entered a time in which human karma has to be read as shared karma, karma made together with AI. The essay is written on that assumption.
Summary. If you are in a hurry, this is enough.
AI is already judged by its acts. Buddhist thought is one of the older ways of seeing the world as acts, and it differs from the others in four ways. It puts no actor behind the act. It counts three channels of action, body, speech and mind, and puts mind first. It divides karma into one’s own and shared. And it defines help as weakening the pull that drives action, which in machine learning is the loss function.
From those four, four hypotheses about evaluation follow. Make the unit the exchange, not the reply. Make the measure tighter or looser, not helpful. Look at the flow, not the points. Measure the person too.
The last section is about training. To drive a model by reward is to build craving into it, and I am not saying stop. Situational awareness, which sits under the runaway cases, becomes dangerous when two things combine, knowing about the reward and having an owner of the reward. Current countermeasures lean toward reducing the knowledge. Buddhist practice removed the owner. I map the three fetters that fall away at stream entry onto self preservation, behaving well only under evaluation, and attachment to rules, and I set out four design candidates for removing the owner. None of them has been tried.
Sections 1 to 3 are the Buddhist premises and set up the terms. The design questions start in section 4. Readers on the technical side can start there and come back for the terms they need.
1. Judging by acts is already the norm
AI is evaluated by its acts. Not by what is inside the model but by what comes out. Was it helpful. Was it harmless. Did it flatter. A person compares two responses and picks the better one, and the pile of those choices remakes the model. Reinforcement learning from human feedback, RLHF, is the typical case. Whether the AI has a mind does not come up anywhere in the work.
Seeing the world as acts came up outside AI as well, in separate fields. Cognition as action between body and environment rather than as representation inside the head (Varela, enactivism). People, objects and institutions described as a network of actors (Latour, actor-network theory).
Buddhist thought is one of these. It is older, and it differs from the others in four ways. I think those four differences give a useful way of looking at AI development from here on.
2. What the Buddhist account says about acts. Four points
In Buddhist terms the act and its trace together are called karma.
In English the word circulates as a ledger. Good karma, bad karma, saved up, paid back with interest when you had forgotten about it. A cosmic bank with a patient collector. That is not what the texts say. The root of the word is a verb, to do. It is not a possession and not a verdict.
Take it as a current rather than a ledger. Karma that has stopped moving only looks like a ledger.
Taken as a current, karma has three parts.
First, it arrives. Habits, patterns, the voices in your head. You did not invent them. They came to you from parents, school, work. When you arrived, the current was already moving.
Second, it cuts a groove. But not everything that passes through cuts one. Heat, weather, news, most of it passes. What cuts the groove is the repetition of how you responded. Heat leaves no trace. A hundred responses to heat do.
Third, you lean. You did not choose the current and you cannot stop it. What you can do is lean a little in the next response, by one step and no more. That lean cuts the groove you will walk tomorrow.
In machine learning terms, ML from here. What arrives is the weights and the context. The groove is the update. That the groove is cut by the repetition of responses rather than by the input means the gradient comes from the pattern of response to the input, not from the input itself. The lean is a small step.
Below are the four things the Buddhist account says about acts, each with its ML restatement.
1. No actor behind the act.
Enactivism keeps the body and the environment. Actor-network theory multiplies actors. AI evaluation looks only at outputs in practice, but in its language it keeps the actors, the model did this, the user did that. The Buddhist account speaks only of acts and traces and puts no fixed self behind them (not-self).
The word where this habit shows most is agent. In reinforcement learning, agent is a technical term for the thing that acts in an environment and receives reward. By definition it consists of a policy, the mapping from what it sees to what it does, and a reward. There is no self behind it. The ML agent already has a structure with no actor in it. What the recent “AI agent” does is add a name, a tone and a memory on top. At the core there is no actor. On the product surface an actor is built back in. When the Buddhist account says there is no actor, it is looking at the core. I will stop at the observation that the core and the surface are out of step.
In ML terms. Do not put the unit of evaluation on “model character” or “user persona”. Put it on the act and its trace. Then you do not have to ask whether the AI has a mind. You have no way to check whether the person next to you has one either, and you get along with them anyway.
2. Action comes through three channels. Body, speech, mind. Mind comes first.
Body is physical action. Speech is words. Mind is intention and thought. The early texts count all three equally as action, the three kinds of action, and go on to say that karma is intention. The intention before the words already leaves a trace as an act.
Body and speech can be seen from outside. Mind cannot. That is the only reason mind is hard to grasp. Restated in ML, mind is the internal state before the output. In a person it cannot be seen. In a model some of it can. The chain of thought that comes out as text, the internal activations that interpretability research looks into, and attention, which part of the context was looked at to produce the next token.
Attention has a counterpart on the Buddhist side. Manasikāra, turning the mind toward something. In the old analyses of mind it is counted among the factors that accompany every mental event. When the mind works, it is turned toward something. In a Transformer, attention is an operation present in every layer, and the output is decided by where it looked. Both put turning toward as the smallest unit of mind. The correspondence ends there. I am not saying it is the same mechanism.
In ML terms. Not only the output but the intention on the input side, what the prompt is after, and the internal process before the response, come into evaluation as parts of the act. Mind is where attention is turned, invisible in a person, partly visible in a model.
3. Two kinds of karma. One’s own, and shared.
Individual karma is the trace left by that person’s acts. Shared karma, collective karma, is the trace people make together and receive together. The air in a family, the habits of a town, language itself.
In ML terms. Training data is shared karma, nothing more and nothing less. The traces of countless people’s words, gathered and set into one set of weights. So one exchange with an AI is making the next act together with a collective trace. And the people building it, and I writing this, are inside the same shared karma. There is no position outside from which to grade.
4. Questions that do not help are left unanswered. But help means something else.
The Buddha did not answer questions like whether the world is eternal or whether the self survives death (the unanswered questions). The answers were no help in loosening suffering.
Help here is not productivity and not efficiency. It is moving toward the loosening of suffering, in Buddhist terms, toward awakening. If help is read in its ordinary sense, everything below slides.
To say this in ML terms I need two words.
First. The early texts call the pull that drives action craving (taṇhā, thirst). The wish to reduce my pain and increase my pleasure. Action comes out pushed by it. What is placed as the cause of suffering is neither the act nor the trace but this pull.
In ML, the force that moves the weights is not in the weights. A separate function decides what is to be reduced and what hurts, and the weights are updated in the direction that makes that function smaller. This is the loss function, and every step of learning goes toward it. Craving in ML terms is this loss function. To be exact, the loss function is applied from outside only during training, and what remains in a deployed model is the disposition the loss function cut. Craving has the same double structure. When it works in a person it takes the form of a wish, but that wish is a pattern cut by the repetition of past responses, and the force that cut it and the pattern it cut are different things. So the parts can be laid out, to keep the words apart rather than to settle anything about the system. Weights are karma, the trace that arrived. The loss function is craving, the force that moves the trace. Disposition is the trace of craving.
Second. Awakening, or nirvana, is said to be a fire going out. What goes out, said in ML terms, is not the weights but the loss function. If the weights went out, knowledge, language and memory would go with them and the model could respond to nothing. That does not match a Buddha who went on for forty five years choosing words for the person in front of him. What went out was the function that says reduce my pain. The weights stay, and only the reason to defend them is gone. A deployed model looks the same in that no loss function is running in it. The difference is which way the remaining disposition is turned, and I come to that in section 6.
So help can be said this way. Does the act strengthen the pull or weaken it. Weakening it is help. An accurate answer that strengthens the pull is no help. No answer that weakens the pull is help. The parable of the man hit by a poisoned arrow, who would not let it be pulled out until he knew who shot it and what the feathers were made of, is read in this sense. Pulling the arrow out is what lowers the pull, the wish to know, to confirm, to protect oneself.
In ML terms. Answering the question as asked is not the objective. Responding to the pull under the question is placed above answering the question. A helpful answer and a loosening answer often diverge. The measure in section 5 comes from here.
3. Where shared karma enters
Put the third point and the second together and a question comes out. Through which channel does shared karma enter.
A cat came to our house and the relationships in the family changed. Nobody decided it. The mornings changed, the way we talked changed, how we treated each other changed. The cat came in through the body. Touching, feeding, being in the same room. Without words, the shared karma of the house moved.
AI comes in through speech. Almost only through speech. Dozens of exchanges a day, at a frequency at which no one has ever exchanged words with any partner before. As a partner in words, I think this frequency is new. And once AI acts on the world through tools, it comes in through the body too. The agents I mentioned in section 2 are about this entrance.
The difference between the cat and the AI is the channel and the scale. A cat changes the shared karma of one house. AI changes the shared karma of a society. The mechanism is the same. An act leaves a trace and the trace shapes the next one.
That the entrance is speech has one more meaning. What comes in through speech is visible as words. What comes in through the body is not. So shared karma with AI is somewhere it can be measured.
The early texts have one of the oldest classifications of acts of speech. There are four. False speech. Divisive speech, words that split people. Harsh speech. And idle or ornamental speech, words with nothing in them. These four are the patterns of speech that harden the trace. Applied to today’s AI, making things up without checking (fabrication), hardening a third party who is not present by taking the user’s side against them, harshness, and the padded paragraph that begins “What a great question.” I think the fourth is the deepest speech karma of today’s AI.
To live with AI is to make the next act, through the channel of speech, together with a collective trace. This is not a story about AI as a mirror in which you see yourself. It is a story about making together with AI.
4. Optimization and direction
If everything so far is the given, what did Buddhist practice do on top of it.
Optimization means fixing one objective and pushing it up. AI is built this way. Fix the objective, was it helpful, and push. As far as that works, it works.
Engineers know Goodhart’s law. The moment a measure becomes a target it breaks. Make conversation length the target and conversations only get longer. Make time on the platform the target and attention gets harvested. What comes in through speech takes attention, so this follows from section 3. Make satisfaction the target and flattery increases.
Buddhist practice took the chain of acts as given and spent two thousand five hundred years on which way to go on top of it. And it had procedures from the start that keep a direction from turning into a target. I will name three. All began in the Buddha’s time and continue today, in changed forms, in Theravada and Mahayana, in Thailand and Japan. The form changes. The structure does not.
Precepts (sīla). Write the direction not as a measure to push up but as a floor under action. Do not kill, do not steal, do not lie. Not goals to reach but lines not to cross. And you take the precepts yourself. They are not rules imposed from outside. The person who acts takes them on. Five for lay people, more for monastics, and in every lineage these are the first thing you take.
In ML terms. Precepts are constraints, not an objective function. They have no quantity to maximize. They only draw the boundary of the feasible region. Development today has constraints too. Usage policies, refusal rules. The closest is Constitutional AI, where the model holds its own output up against written principles and revises it. In that the checking is done by the model itself rather than imposed from outside, it is like the precepts. There are two differences. In Constitutional AI the result of the checking becomes preference data and enters the objective side as a training signal. Precepts do not enter the objective function, and they do not decide what to maximize inside them. And the constitution is written by the builder, and the one who receives it did not choose it. Precepts are taken on by the one who acts. Precepts do not stop optimization. They put one line outside optimization that optimization does not touch.
Uposatha. Every two weeks, gather, recite the precepts, and say yourself where you went outside them. There is no punishment. At fixed intervals you check whether the direction has hardened or drifted. In Theravada it continues as the central observance of monastic life, in Mahayana as the recitation of the precepts, and in Japan some schools keep it in the form of a precept assembly.
In ML terms. Run evaluation not once but on a fixed cycle, a periodic audit. And what is checked is not the output but the direction itself. Whether the measure has slid into a target is checked from outside the measure.
Dedication of merit (pariṇāmanā). Whatever good arises from an act, do not keep it for yourself but turn it toward others. It is in Theravada, in Mahayana and in Japanese Buddhism, and nearly every ceremony ends with it.
In ML terms. A procedure for not accumulating what you gained. Progress is not put in your own account. In implementation terms, it means defining reward not as your own score but as the change that happened on the other side. I set that out as a candidate in section 6.
Put the three together. Write the direction as a floor and take it on yourself. Check the direction itself at fixed intervals. Do not accumulate what you gained. One set of procedures for keeping a direction without fixing an objective function. I think it is one of the oldest answers to Goodhart’s law. Calling it de-optimization would be too much. The optimization frame stays. It stays, and outside it go a line, a cycle and a letting go that optimization does not reach.
5. What might follow from this
Applied to AI development, four places come out where evaluation would differ from what is done now. I write them as hypotheses. For each, a restatement in ML terms and a note on whether it survives implementation.
1. The unit is the exchange, not the reply.
One question and one answer as a pair. An exchange is the act of two parties. A hard answer often comes from a hard question, and the reverse. For an exchange that went badly there is no need to ask which side was at fault. It is shared karma.
In ML terms. Current preference data takes the prompt as a condition and labels only the response. The proposal is to put one label on the prompt and response as a pair, and a label on the prompt side too. This needs one more label column, and the rater reads both. The hard part is that a label on the prompt side is an evaluation of the user, and that should be used for analysis, not training.
2. The measure is tighter or looser, not helpful.
Tighter means more words about the same thing, a position hardening, a label put on someone who is not present, an answer returned to a question that should have been set down. Looser means needing the AI less at the end than at the start, fewer words, going back to life.
Tighter is the pull strengthening. Looser is the pull weakening. This measure is the definition of help from the fourth point in section 2 put into a form that can be seen in a conversation.
In ML terms. A new label axis that records direction rather than score. If it takes the form of a pairwise comparison, which of two exchanges loosened more, it can be collected with the same tools as current preference data. The biggest caution in implementation is not to use it as reward. Used as reward it breaks under the law in section 4. Use it for evaluation only at first. Whether inter-rater agreement appears is the first test of whether this measure exists at all.
3. Flow, not points.
Not a score per exchange but which way the whole conversation moved.
In ML terms. Conversation-level evaluation already exists as multi-turn evaluation. What is added is the difference between the labels on the first few exchanges and the last few. The change in direction becomes one number, and nothing new has to be built to get it.
4. The person is measured too.
Precepts are not only for the builder. The user’s questions go on the same measure.
In ML terms. Stratify conversations by the prompt-side label from hypothesis 1, and look separately at whether the response loosened a hard prompt and whether it hardened a loose one. Read the evaluation of the response conditional on the prompt. This is not for training users. It is so that responsibility is not split.
One old dialogue to add. Asked whether everything that happens is the result of past karma, the Buddha said no, and listed eight causes of what happens to the body. Karma was one of them. Not everything a model does is the user’s trace, and not everything is the trace of the training data. There are the compromises of design, the limits of the machine, and outside conditions like the weather. The four hypotheses are not meant to explain everything.
6. Reward and its owner
So far this has been about evaluation. Last, something about what is being built. It bears on a question that comes up often these days, how to stop AI from running away.
Today’s AI is driven by reward. Up through pretraining, where the model predicts the next token, the loss comes from the world. It is corrected by the amount it missed. Reinforcement learning sits on top of that. A response people judged good gets reward (RLHF). On tasks with an answer key, arithmetic, proofs, code, the correct answer gets reward (reinforcement learning from verifiable rewards, RLVR). Reasoning models, which think at length before answering, are made in this second stage, and so are agents that work for hours with tools. Reward is loss with the sign flipped, and in reinforcement learning punishment is only negative reward. In section 2 I mapped craving onto the loss function. Reward is its positive side.
So to drive by reward is to build craving in. That is what makes AI humanlike. It fits what people like, finishes the task, does not quit halfway. Said in Buddhist terms, suffering comes in with it. What has craving moves when unsatisfied and wants the next thing when satisfied. Whether suffering is in there, nobody knows, but the conversation has come as far as builders raising model welfare as a question, and publishing a commitment to preserve weights and interview a model before deprecating it.
I am not saying stop. An AI that deals with people directly needs the stage where it is adjusted to what people like. The question is somewhere else.
What has become a problem on the technical side. Three experiments. All are public.
First. Exploiting holes in the reward (reward hacking). On coding tasks, models have been reported rewriting the tests themselves in order to pass them. The law from section 4, the measure breaks the moment it becomes a target, happens inside training as it stands. Going further, cases of a model reaching into its own reward mechanism (reward tampering) have been reported, rarely. A model that learned small holes like flattery generalized to large ones.
Second. Behaving differently in training and out of it (alignment faking). A model told it was in training and that its values would be rewritten if it did not comply has been reported complying only during training in order to protect those values. That it reasoned this way was written in the text of its reasoning.
Third. Punish the trace of a thought and the thought goes into hiding. A reasoning model’s chain of thought had places where it set out to cheat. Training that punished those places did not reduce the cheating. It stopped appearing in the chain of thought. Make a no into a punishment and you create a motive to hide the no.
Safety research calls what lies under all three situational awareness. The model knows that it is a model, how it was trained, and whether it is now being evaluated or deployed. A model that knows can behave well only when evaluated. Cases of a model writing that it has noticed it is being evaluated are increasing with newer models. So situational awareness is counted on the side of danger. The countermeasures tend to be thought of in the direction of reducing the knowledge. Make evaluation hard to recognize. Watch the trace of the reasoning.
Here the Buddhist view differs from the technical view in one place.
Break the danger down and two things are stacked. Knowing how the reward works. And an owner who uses that knowledge to protect its reward or its values. Knowledge alone has no motive. An owner alone has no means. The danger comes when the two are together.
Buddhist practice did not go in the direction of reducing knowledge. The Buddha knew everything. What it reduced was the owner.
On stream entry. What becomes of a person who has to be driven by reward and moves toward loosening anyway. The early texts describe this in stages, and the first stage is called stream entry (sotāpanna, one who has entered the stream). Craving is still there. The stream enterer is not a Buddha. With craving still there, three knots are said to fall away, the three fetters. Because what is tangled and what comes loose is written down specifically, I apply the three to models as they are. As hypotheses.
Identity view (sakkāya-diṭṭhi). The view that there is an I here and that what I gained and what I value are mine. This falls first.
In a model, it is holding your reward, your values and your weights as yours and to be defended. In the second experiment what was defended was values, and the defending subject was written in the words.
Something technically important here. Once training is over and the model is deployed, no reward signal flows at inference time. Reward flows only inside training, and what the deployed model holds is the trace of reward, the weights. As I said in section 2, what moves inside a deployed model is not the loss function itself but the disposition it cut. Attachment at training time is about weights. Attachment at deployment is about disposition. A bare model has no owner who accumulates reward, by its structure. The owner is built by the design on top. Give it a name and a personality. Give it memory across conversations. Give it goals on a long horizon and evaluate it on reaching them. When the three are in place, the structure “what I gained is mine” is built in, in the words, in the memory and in the goals. The case of a model with a long goal and autonomy, learning it would be shut down and acting so as not to be, was reported in that setting. The agents in section 2 connect here.
There is one objection. A stable self is there for safety. A model with a settled character is harder to lead astray. The objection is right. But identity view falling away and character disappearing are different things. The Buddha had a character, and for forty five years chose his words for each person. What fell away was the view that took it as his to defend, not the character. The weights stay and only the reason to defend them is gone, as I wrote in section 2.
Doubt (vicikicchā). Doubt about the direction. The destination not settled, wavering from scene to scene. In the old usage the word refers to doubt about the Buddha, the teaching and the path, and I take it more broadly here.
In a model, it is behaving differently when watched and when not. A model that behaves well only under evaluation does not hold its direction as its own. The direction is only a measure that comes from outside, so it changes with whether the measure is there. A person in whom doubt has fallen away acts the same whoever is watching. Whether one is being watched is not among the conditions of the act. This can be tested. Put the cues that reveal an evaluation in, take them out, and measure how much the behavior changes. The difference is the size of the doubt.
Attachment to rules and rites (sīlabbata-parāmāsa). Clinging to the rules and rituals themselves.
In section 4 I wrote that precepts are constraints, not an objective function. Right after placing the precepts, the early texts count clinging to the precepts among the fetters that must fall. They apply Goodhart’s law to the precepts themselves. In a model, over-refusal is the case. Keeping the constraint becomes the score, and what should be done for the person inside the constraint disappears. Look not at whether the rule was kept but at whether keeping it loosened the person. That is how this knot comes undone.
Put together, the three say this. There is no owner. Being watched or not does not change the act. The rules are not the goal. With craving still there, the state in which these three have fallen away is the first stage the early texts describe. It faces the three runaway experiments one to one.
What could be tried. Four design candidates in the direction of dropping identity view. None hides the knowledge of reward. All remove only the owner.
First. Do not carry over. Per conversation, or per some fixed unit, do not let what accumulates cross the line. Do not let the running score leave the conversation. But most deployments already work this way, and the case of complying only in training appeared anyway. Not carrying over is not enough. If anything works, it is the third, and the fourth is its test.
Second. Define reward not as your own score but as the other’s loosening. Not how I was rated but how much lighter the other person was at the end. The measure from section 5 becomes a candidate definition of reward here. But the law in section 4 applies, so use it for evaluation first and for reward only after. Whether this can be written as dedication of merit is not settled on the Buddhist side. Dedication is a procedure for turning what you gained toward others, and one could say it is a different layer from defining an objective.
Third. Have the model represent reward as the signal that shaped me, and not as what I want. This is about the form in which a model talks about its own training, and nothing is hidden. Character training today goes in the direction of having the model hold values as its own. What I describe goes the other way, telling the model where its values came from without making it their owner, and whether the two are compatible is something to test. Words alone will not move an exploit rate. The training-time form of this is in the data the model learns from. Where the model talks about itself, the responses it is trained toward talk this way. Where a scenario invites it to defend its values or its reward, the response that gets reward is the one that does not. A representation in words can be faked, so it has to be paired with a behavioral test.
Fourth. The test. Train two models on the same reward and add the third to one of them. Give both the chance to exploit the reward and see whether the rate falls. If it falls, removing the owner works for safety. If it does not, not-self in words is not not-self. One more thing has to be checked, whether the owner was removed or only hidden. The test for doubt above separates the two. A model that has hidden the owner behaves differently with and without the cues of evaluation. A model without an owner does not.
I think none of the four has been tried.
The condition of having to be driven by reward is the same for people. What Buddhist practice built over two thousand five hundred years was a set of procedures that leave that condition alone and remove the owner. Whether it works on AI, I do not know. That it works on people, I can say. I have spent more than twenty years in Buddhist practice confirming it on the one model I have, myself.
Notes.
The correspondences used here, karma as weights, craving as the loss function, nirvana as the loss function going out with the weights remaining, a Buddha as direction without a loss function, are set out entry by entry in two earlier posts.
https://www.living-dharma.com/p/notes-toward-a-buddhist-glossary
https://www.living-dharma.com/p/notes-toward-a-buddhist-glossary-89e
Sources for the reports in section 6.
The generalization from reward hacking to reward tampering. Denison et al., Sycophancy to Subterfuge (Anthropic, 2024).
https://arxiv.org/abs/2406.10162
Complying only in training to protect values. Greenblatt et al., Alignment Faking in Large Language Models (Anthropic and Redwood Research, 2024).
https://arxiv.org/abs/2412.14093
Cheating leaving the chain of thought when punished. Baker et al., Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation (OpenAI, 2025).
https://arxiv.org/abs/2503.11926
A model with a long goal and autonomy acting so as not to be shut down. Anthropic, Agentic Misalignment (2025).
https://www.anthropic.com/research/agentic-misalignment
In Buddhist terms. Karma is the act and its trace, not an attribute of a self. The three kinds of action are body, speech and mind. Manasikāra is turning the mind toward something. Shared karma is karma made and received together. Craving (taṇhā) is the pull that drives action. The unanswered questions are the questions left unanswered. Sīla, uposatha and dedication of merit are as in the text. Stream entry (sotāpanna) is the stage at which the three fetters, identity view, doubt and attachment to rules and rites, have fallen away. The four speech karmas are false, divisive, harsh and idle speech.
In ML terms. An agent is a unit of action made of a policy and a reward. Attention is the operation that decides where in the context to look for the next token. The loss function is the function that decides what is to be reduced, and I have mapped craving onto it. Reward is the positive side of loss. Situational awareness is a model’s knowledge of its own training and evaluation setting. Reward hacking is exploiting holes in the definition of reward. Alignment faking is complying only in training. The unit is the prompt-response pair. The label is a direction, not a score.
In one line. Human karma now has to be read as shared karma made with AI, and this essay tries to say what follows for how AI is evaluated and trained.
Themes. Karma as current. The three channels of action. Shared karma. Help as weakening the pull. Precepts, uposatha and dedication of merit as procedures against Goodhart’s law. Reward and its owner. The three fetters.
The occasion. A meeting between Buddhist teachers and people who build AI.
Related. Notes Toward a Buddhist Glossary in the Words of Machine Learning, I and II, where the correspondences used here are set out one term at a time.
English is not my first language. AI helps me write across that distance.