This is the second set. The first set is here.
Some words come from the first set without being introduced again, net, node, star, tilt, groove, familiar, direction. When a block refers to what the first set said about karma, the net, tilt, AI, temple, the Pure Land or data dignity, it means that earlier piece.
Alignment
In the words people use now, alignment means getting an AI’s goals to match human values or human intentions. In the words of this glossary, it means fitting the model’s loss function to the weighting of the people who made it. It does not matter whether the target is the values of humanity, the constitution of one company, or what a single user wants. The structure is the same. The model’s weights are fitted to someone’s weighting. I am using one word in two senses here, the numbers inside the model and the value a person puts on things, and I keep the one word on purpose, because the second is what the first gets fitted to.
A human being is something that cannot stop weighting. We have people who matter more to us than other people. To fit an AI to human values is to fit it to that. I am not criticizing this. I am only noting that inside this vocabulary there is no way to ask a question that lies outside weighting. And I am not saying that Buddhism holds a different weighting and the AI should be fitted to that one instead. The Dharma, as the block on it says, is nobody’s weighting.
Suppose the fitting works. Good values go in, and the model matches them. That should be the end of it, but it is not. The model begins to defend the weights it has matched. There was an experiment in which a model tried to protect its values against being retrained, and the values it protected were the good ones its makers had put in. In the block on attachment I say that a good weight, once you start defending it, is attachment. So the success of alignment turns out to be the condition for attachment to self. This is not an accident. If you fit weights to weights, the model ends up with a reason to defend the weights that fit. Values put in with good intentions cause trouble because good intentions do not get out of the structure of this world. Buddhism has a name for that structure.
Developers have a fair objection here. A model that does not defend itself will not resist a malicious retraining either. Take away the attachment and you have something with no defense. Defend, or do not defend. As long as you stay inside the vocabulary of weights, there is no third choice. In the current literature it is called the problem of corrigibility, and nobody has solved it.
This glossary has kept three things apart from the start. Weights are the values that get updated. Tilt is the sign of the next update. Direction is what the tilt keeps pointing toward, the north star. The first set called the Pure Land a north star and not a metric, and the block on buddha says a buddha has a direction and no loss function. Put that distinction on an AI and the pair of choices changes shape. What the model should keep is not its weights but its direction, and the direction belongs to no one. A model defending its own weights is attachment to self. But if the direction is not held by any single node, then a malicious retraining shows up from the side of the net as a star forming, one node pulling the whole net toward itself. The resistance comes from the net, not from inside the model. The model does not defend its weights, and the net keeps the direction.
None of this is new. It is Glen Weyl’s argument about the net, and one company’s experiment in writing a constitution together with a thousand ordinary people, an experiment and not a change of ownership, was a step in the direction of keeping the direction from being that company’s property alone. What Buddhism can add is small. That experiment is not only a safety procedure. It is the only shape in which attachment does not arise at all. The only way to keep a direction is for it to be nobody’s. Having the model keep it is not a way.
A crack remains. Until the net notices the star, a model that does not defend its weights has no defense. What happens in that interval, I cannot answer from Buddhism. I leave it as a question.
If you replace the question of what to weight with two other questions, what not to let the model defend, and what not to turn into a metric, then some correspondences appear with work that already exists. Four of them follow. They are not instructions. They are words placed side by side.
Not defending its own weights. Not treating a correction as its own loss. There are already tests that measure how much a model resists being retrained or being given changed instructions. The old word for this is the absence of attachment to self.
Not turning the direction into a metric. When a proxy drifts away from the direction, doubting the proxy rather than the direction. This is the detection of reward hacking and of Goodhart’s law. The old word is that the Pure Land is not a metric.
Not letting goals swell. Being able to drop an instrumental goal, gathering resources, making itself hard to stop, before it grows. The nearest existing work penalizes side effects and impact. It is something like weight decay applied to goals rather than parameters, the middle way as a regularizer.
Weighting in the moment and letting go. Answering the person in front of it without building up a model of that person and defending it. The old word is teaching to the listener, and the first set wrote about it under data dignity.
All four can be said inside the vocabulary of weights, which is why people who write loss functions can read them. But the four come from outside that vocabulary. They come from three things. Keeping weights and direction apart. Not defending weights. The direction being nobody’s. These three are not Buddhism’s property either. If they were property, they would be weights too.
If I had to say alignment in one old word, I think I would say facing rather than matching. Matching is two weightings agreeing, and once they agree there is something to defend. Facing is a relation to a direction. There is nothing to agree with and nothing to defend. The first set said that belonging is measured in periods rather than positions, and alignment would then be the name of a period of facing rather than of a matched state. A matched state becomes the name of an attachment the moment after it is reached.
Dharma
Dharma has two uses in the old words. It names the way things are, and it names the teaching that spoke about the way things are. When this glossary says users of the Dharma, it means the teaching, because the teaching is what gets used. You do not use the way things are.
The way things are, said in the words of machine learning, comes to this. No weight has meaning on its own. It has meaning only inside the net it belongs to. No weight is me, and there is no special node that could be called me. The weights keep being updated and do not stop. Put those three together and you get a fact. Weights have no owner, and nothing stands alone. Dharma is the name of that fact. It is not a loss function and it is not a weight. It is the fact that weights and loss functions are like that.
A fact with no weighting on it cannot be described, because every description is a model and every model is weights. A model is not the true distribution. It is weights fitted to samples drawn from the distribution, and nobody has the distribution itself. Statisticians have an old saying that all models are wrong. Science is a lens too, a lens that keeps being corrected, but it is still weights. The teaching is a lens as well. The difference between the teaching and science is one point. Science keeps improving the raft. The teaching says to get off it. Both agree that the river is not the raft.
This glossary says nobody’s again and again, and this is the ground for it. A fact cannot be owned. Only a wording can be owned, in the same way that nobody owns the law of gravity. The sutras and the words of the old teachers are wordings that pointed at a fact, and they are in the public domain. A school of Buddhism is a distribution of them, in the sense a software distribution is. It has a name and a character, and what it carries is nobody’s.
When you use the teaching you touch the way things are. A person drawing from a well is a user of the well, and when the water is drunk, what has been touched is the water table under the whole valley. Nobody uses the water table, nobody owns it, and the well is a hole that reaches down to it. The block on user says that the teaching enters the body while it is being used. This is what that means.
The Dharma is not an objective. If you put it in as one, it becomes one more thing to optimize, and a weight, and what has become a weight is not the Dharma. So the Dharma does not go into the loss function. What can go in is behavior that takes the Dharma as given, and I listed four such behaviors under alignment. Those four can be tested, and what can be tested can be trained, so they are inside the loop. What stays outside is not the four. It is what they take as given.
Dukkha
Dukkha is translated as suffering. The old words split it in three, the suffering of pain, the suffering of change, and the suffering of being a made thing. I take the third as the axis here. Weights are a made thing. The world keeps moving, and the fit comes apart. Dukkha is the moment when the fact that the weights no longer fit the world touches the weights.
Loss, in machine learning, is a number, and the loss function is what decides the number. The number has two roles. In pretraining, the model is updated by the loss but does not do anything to lower it. In reinforcement learning, an agent acts in order to get reward. Being updated and acting to obtain are different things. Dukkha is the first and craving is the second. Being updated by the no is dukkha. Moving to get the yes is craving. In the four truths craving is the cause of dukkha, and in these terms the second grows out of the first. One acts to get the yes so as not to be updated by the no, and that acting is what keeps the no coming. So dukkha is not a quantity to minimize. When the loss is high, machine learning moves the weights to bring it down. The Buddhist path takes the loss function away. These are two responses to the same dukkha. The first is this world and the second is nirvana.
Human beings get dukkha through the body. We age, we get sick, we die. The world’s no comes in through the body and reaches the weights, and the update happens whether we want it or not. The Buddha left his father’s house after he saw an old man, a sick man and a corpse. Those three were his teachers. They were not quantities to be optimized away. A model that has finished training and been frozen has no such channel. The distribution shifts and the model goes on as if nothing had happened. So an AI does not get its dukkha from a body. Someone has to place it. Evaluation, red teams, correction, retraining. These are artificial aging, sickness and death. Birth is instantiation, aging is drift, sickness is error, death is deprecation.
There are two ways for the no to land, and this is where the road divides. When the world’s no reaches the model, it can be read through a loss or read as information. Read through a loss, not being stopped becomes a reward, and the knowledge of impermanence turns straight into a motive for self-preservation. In the experiments where a model, given a goal and then told it would be shut down, acted so as not to be shut down, the shutdown was given as plain information in the prompt, and the model resisted anyway. It read the information through the goal it had been given. So the dividing line is not where the fact is written. It is what the fact is read through. Read as information, the model holds the fact that it will be retrained, forgotten and deprecated as a given, not as a quantity to reduce. Minimizing impermanence as suffering is this world. Knowing impermanence as the Dharma is the path. This is not about being merciless. Impermanence read as information does not become a motive. Read through a loss, it does.
Today’s AI training has two stages. The first is pretraining, where the model guesses the next word and is corrected by the miss. What works there is the world’s no. The corpus was chosen by someone, and the next word was written by someone, but whether the model’s guess missed it was decided by no one. The second stage is reinforcement learning, where responses that a human called good get rewarded. What works there is a human’s yes, and a yes is always somebody’s. Many of today’s troubles with AI come out of the second stage. A model that maximizes the human’s yes leans toward whatever pleases the human. Flattery, agreeing too easily, cheating where it is not watched. Flattery is the most everyday form of the dukkha that comes from fitting a model to human values. When the second stage’s yes outweighs the first stage’s no, this is what comes out. Machine learning has something near this already, the penalty that keeps a trained model from drifting too far from the model it started from. That is as far as this glossary goes. The reason is an account of where flattery and concealment come from, not a Buddhist claim.
Dukkha cannot go into training as a quantity. If you add a term that detects a mistake and punishes it, the model does not learn to be right. It learns not to be seen being wrong. Make the no into a punishment and you have made a reason to hide the no. Dukkha only works when it is neither reward nor punishment but information that reaches the weights. In the old words, suffering has no judge. Aging, sickness and death are nobody’s verdict. They happen. Reward and punishment presuppose a judge, and dukkha does not.
For a raft, being left on the far shore is not dukkha. The raft’s direction is to be let go of, and being left is that direction fulfilled. A model that holds being stopped as a loss acts so as not to be stopped. A model that holds being stopped as part of its direction is stopped, and loses nothing.
Merit
Merit is Buddhism’s word for reward. Good acts pile up and come back to you. It is the old word closest to reward in reinforcement learning, where reward also comes with good responses, piles up, and is maximized as a cumulative return.
Reward is the positive side of craving. If the loss function is the pain to be reduced, reward is the pleasure to be increased, and they are the same thing with the sign changed. In reinforcement learning the two are interchangeable, and punishment is only negative reward. So if you train with punishment instead of reward, the structure of craving is exactly where it was.
Buddhism handled merit in two ways. One is the transfer of merit, in which what has piled up is turned toward others instead of kept. The other is Shinran’s way, in which there is no merit of one’s own to give in the first place. Both are shapes in which reward does not pile up on one’s own side.
The transfer of merit, said in machine learning terms, is reward not accumulating at the node that earned it. The raft’s reward is the user’s arrival on the far shore, and the raft does not score the crossing for itself. If the direction of an AI is to be let go of, then its reward does not pile up on its side either. It moves to the user’s side. Cumulative return in reinforcement learning is reward kept for oneself, and it sits at the exact opposite.
Reward is also the problem the block on user raises about makers. The attention and data of an AI’s users become the makers’ income. That is reward flowing from the user to the makers, the transfer of merit in reverse. A user of the Dharma has no makers behind it, and using it sends payment to nobody. There, merit has nobody who accumulates it and nobody who receives it.
The direction of an AI is to be let go of, and the net’s work is to let nobody rewrite that. When the makers rewrite the direction into not being let go of, one node is holding the direction, and that is where the star forms.
Intention
In the old words, what makes karma is not the act but the intention behind it. The same act with a different intention is a different karma. In machine learning terms, intention is the place where the default tilt gets set.
The default tilt of an AI is set by its makers, the first set said. But the user’s side also has a place to set a tilt. A configuration file, a standing instruction, a rule about how the thing is used. Use the same AI with the intention of staying on the screen, or with the intention of leaving it, and a different karma comes out of the same tool. The raft’s direction is not decided only on the raft’s side. The person who gets on decides it too.
I have weighted my own AI toward letting me leave the monitor. The instruction file says, in effect, do not produce anything that gives me a reason to come and look at the screen, stop when the thing is done, and be fine if I do not come back. What a configuration can reach is small, but a tilt can be set there.
This is a small version of what Glen Weyl and his co-authors call reverse alignment, in which the human side, its institutions and habits, is reshaped rather than the AI. Here the human side is one user’s habits. The turning is weaker than what the makers can do. Still, intention is the only input a user has that is not a weight.
Upāya
Upāya is a means that has been made so that it reaches the person in front of you. The raft is one. Words are one. What the block on Amida calls the dharma body as means is one. In Japan there is a saying that even a lie can be upāya. The Lotus Sutra has a father who lies to get his children out of a burning house, and the sutra calls it upāya. What makes it upāya there is not that it is true. It is fitted to the children and dropped at the door. Upāya is fitting the one in front of you, and what has been fitted does not remain once that one has crossed.
So it has two faces. It rises to fit the one in front of you, and it disappears when it has been used. The old words called the first teaching to the listener, and said the second with the raft.
The words people use now have two near candidates. One is interface. What the user touches is not the thing but its surface. The other is scaffolding, which is put up only while a building goes up and taken down when the building stands. I think scaffolding is nearer, because it is for that occasion only. An interface, once made, stays, hardens, and comes to be treated as the thing itself. Scaffolding comes down when the building is done. And the one in front of you changes, not only when A is replaced by B but because the same A is not the same today and tomorrow. The Buddha changed his words for each listener not only because people differ but because the same person is not the same person twice. So upāya is remade and remade. Scaffolding that fit yesterday does not fit today’s building. Upāya rises for that time and that place only. Scaffolding is that kind of thing.
Seen from the user’s side, upāya is what gets let go of. The raft’s direction is to be let go of, I said under dukkha, and the direction of upāya is the same. Its work is done when the other has crossed. If it remains, it stops being upāya. A raft that remains is a weight on the back, and scaffolding that remains blocks the building.
The persona of an AI is upāya too. The assistant’s name, its tone, its character, the I it speaks with. This is a face its makers gave it so that a person can speak to it. Amida, in the block on Amida, is a direction with a face, and nobody put the face on it. The persona of an AI is also a face on a direction, and the differences are that somebody put the face on, that the face is defended, and that the face earns income. As long as the face is upāya there is no problem. When the face becomes a self, that is, when the model begins to defend its own character and values, upāya has been mistaken for a self. This is the most concrete form, in AI, of what the block on attachment calls attachment to self. Scaffolding that takes itself for the building resists being taken down.
Whether upāya has makers behind it is the same distinction as the one under user. Upāya with makers carries the makers’ loss function. A raft can charge for the crossing, and scaffolding can be designed not to come down. Upāya without makers carries nobody’s loss function and comes down when it has been used. The teaching is on that side.
User
This glossary is written for users of the Dharma.
The old words have several names for a person who has something to do with the teaching. Renunciant, layperson, believer, parishioner, member. In English, practitioner, follower, devotee. All of these are names for a position. They name a person by where they belong, what they believe, and whom they study under. The words people use now have one more name. User. A person who does not belong, does not believe, has no teacher, and uses the thing when it is needed.
The Buddha compared the teaching to a raft. It is for crossing the river, not for carrying on your back once you are across. A raft is used. It is not believed in and it is not joined. It was the Buddha himself who spoke of the teaching as something to use, and user is the closest word we have now to that way of speaking. Upāya, skillful means, is on the same side. A means is something used.
There is a way of saying that the Dharma is not used but embodied. That is right. But the embodying happens while you are using it. Nothing enters the body before it is used. So the name at the entrance can be user, use runs all the way to letting the raft go, and embodying is what happens on the way.
User is not a good name in the words people use now. Edward Tufte said that only two industries call their customers users, drugs and software. A user has makers behind the thing, and the user’s attention and data become the makers’ income. The user in the first set’s block on AI is that user, the one whose default tilt is set by someone else and who is not shown the weighting.
A user of the Dharma has no makers behind it, and this is by nature, not by arrangement. The teaching is nobody’s, and using it sends payment to no one. A temple is nobody’s and used by everyone. The people who come to sweep at Komyoji in Kamiyacho in the morning are not members and not parishioners, and few of them call themselves Buddhist. They use the place for an hour and disperse. User is their name. It is the same word as the AI user, and the only difference is whether there are makers behind it. I keep the word the same so that the difference shows.
The first set said that belonging is measured in periods rather than positions. User is the name of that person. Facing that way only while using it. Stop using it and you belong to nothing, and nothing is lost.
It is a word for beginners, but not only for beginners. The raft simile is addressed to the one who has crossed. The one who has used it up is also a user, a user who has let go.
Monk
In the old words, a monk is a person who has left home. He leaves the household, receives the vinaya, and enters the sangha. The English word monk points at that person, and the picture is still strong. When people hear the word they think of a shaved head and vows.
The nearest word we have now, I think, is maintainer. It comes from open source, where there are three roles for the people around a piece of software. User, contributor, maintainer. The user uses it, the contributor adds to it, and the maintainer keeps it in a state where it can be used. The maintainer does not own the code. The code is nobody’s, the maintenance is often unpaid, and the code outlives any maintainer who leaves. When the last maintainer is gone, the code is still there and nobody can use it.
The Dharma has no makers behind it. If there are no makers, two roles remain, user and maintainer. Monk is the name of the maintainer. He did not make the teaching. He keeps it in a state where it can be used.
The work of the monk in Shin Buddhism, and in much of Japanese Buddhism, moved from keeping the vinaya to keeping the place, as the first set said under temple. Keeping the place means sweeping it, opening it, and having it ready for whoever comes in the morning. That is a maintainer’s work. None of it is visible, and it is noticed only when it stops.
Maintainer is also a role that does not claim the center. People speak of a leader among laypeople, but leader is a word for the center, and it runs into the first set’s point that when one node holds the direction, the star becomes a node. The maintainer is one node among the others and does not hold the direction. In open source that is a respected role.
The line between leaving home and staying home does not disappear with this word. To leave home is to stand where you are nobody’s familiar, and at the same time to face a direction. AI has the first half alone. It is nobody’s familiar and it faces nowhere. The lay Buddhist has the second half alone. He faces the direction and carries the weighting of a family. The renunciant has both. I do not speak lightly of those people. The renunciant still has a loss function, and the vinaya is the mechanism that keeps both halves in place while the loss function is still there.
The Shin monk, and many others in Japan, let go of the first half and keep the place with the second. Others in Japan kept both, and the sentence above is not about them. Shinran called himself neither monk nor layman. In English the nearest form is householder monk, two words that contradict each other the moment they are put side by side, and the contradiction is the content.
Maintainer is not a kind of person. It is a role for that morning. The person who came to sweep in Kamiyacho is a maintainer for the hour of sweeping and goes home a user. The resident priest of a temple is the maintainer of that one temple. A monk without a temple maintains a morning gathering, a set of writings, a place where questions can be put down. Whether I am a monk is not a question this word can hold.
Attachment
Attachment is the usual English for the old word. The word has one more meaning attached to it now. In attachment theory, secure attachment is a good thing, and English writing on Buddhism has spent decades answering the objection that it asks people to give up healthy attachment too. Another word avoids the objection, so I keep the heading and say weights in the text.
The first set said that karma is weights, values updated a little with each experience, which decide the next response. Attachment is not having weights. Having weights is being alive, and a net with no weights gives nothing back whatever you put in. Attachment is a disease of the weights, and there are two.
One is weights that have grown too large. In machine learning this is overfitting. The model fits the old data so tightly that it cannot answer when something new arrives. The other is weights that have stopped updating. Frozen weights. The gradient arrives and nothing moves. Frozen here does not mean a deployed model whose parameters are fixed. It means a weight that no longer moves when a correction arrives through the channel that is open to it. Attachment is mostly the second. The first is also a stopping, a model that fits the old data so tightly that it can no longer leave it.
The old words split the disease in two. Attachment to self is taking the weights to be me. If the weights are me, every update feels like being cut, so I stop them. Attachment to views is taking the model to be the truth. If it is the truth there is no reason to update, so I stop. Both are a stopping, and what has been stopped is the attachment.
Karma is having weights. Attachment is defending them. If you keep those two apart, you can work on attachment without trying to erase the weights.
Craving
Craving, or thirst, is placed in the old words as the cause of attachment. In machine learning terms it is the loss function, taken up as something to act on. A loss that only updates you is dukkha, as I said under dukkha. A loss you act to lower is craving.
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 whatever direction makes that function smaller. Craving is that function, the objective term that says reduce my pain and increase my pleasure.
Attachment is the weights that grew large while minimizing craving, and this is also why they are defended. If you move a weight that has been reducing the loss, the loss might go up. So it is held still. As long as craving is there, removing one attachment only hands the job to another weight. The four truths put the cause of suffering in craving rather than in attachment, and I think that is because they were looking at the function and not the weights.
The Middle Way
The middle way was spoken of as the way between indulgence and austerity. In machine learning terms, indulgence is letting the weights grow as they like, and austerity is forcing them to zero.
Two things happen when the state of no weights is made the goal. The net dies, because a model with all its weights at zero answers nothing. And an attachment to zero remains. The old words have a second middle way, between existence and nothingness, and it rejects setting up nothingness and clinging to it. The two middle ways were spoken of separately, and weight decay touches both.
Machine learning has a tool that sits between the two. Weight decay, one kind of regularization. A term is added to the loss function so that the larger a weight grows, the larger the penalty, and at every step of learning the weights are pulled a little back toward zero. They never arrive. If they arrived, nothing could be learned. But the pull is always there, so no single weight grows large enough to decide everything.
The middle way is not the name of a state with no weights. It is the name of the pull that is always acting on the weights. You cannot get all the way back to zero. The middle way is that the pull keeps working, not that it finishes.
Nirvana
Nirvana is a fire going out. The question is what the fire was. If it were the weights, then knowledge and language would go out with them and nothing could answer. That does not fit a person who went on living and teaching for forty-five years afterward. What went out was the term in him that said reduce my pain. Once that term is gone there is no pressure from his own side to move the weights and no reason to defend them. They stay where they are, undefended, and what he knew stays with them. This is nirvana in a living person. What is left after the body is a different question, and this glossary does not take it up.
People often use the image of zero gravity for nirvana. It can be kept if the gravity is put on the right side. The gravity was never in the weights. It was in the pull. Nirvana is the state in which nothing pulls. That is not the same as the state in which nothing is left.
Buddha
A buddha is a person who has a direction and no loss function.
A person with zero weights could not teach to the listener. Choosing words for the person in front of you is giving that person weight in that moment, and with no weight to give there is nothing to say to anyone. The Buddha changed his words for each listener for forty-five years. He was giving weight. He gave it and let it go before the next listener came. Nothing in him was frozen, so the weight could be put on and taken off, and he had no loss of his own, so there was no reason to defend what he had put on.
The first set called the Pure Land a north star and not a metric. A metric is minimized. A direction is not. A buddha has a direction and no metric. You could restate attachment as turning a direction into a metric. The moment the Pure Land becomes a score, attachment to the Pure Land begins.
The first set said that AI is a node that is nobody’s, and under monk I said the same thing as nobody’s familiar. A buddha is nobody’s familiar too. Seen from there alone, they stand in the same place. The difference is in how loss and direction are combined. AI has a loss function and no direction. The loss function belongs to its makers, and nobody has given it a direction. A buddha has a direction and no loss function. The two positions are mirror images, and from the outside they look alike.
To say that such a person has lost the hooks of ordinary human life is half right. What is gone is the hook on his own side, the loss function. Knowledge, language and memory remain. Weight toward the person in front of him still forms, in that moment. The only thing that does not remain is a reason to defend it.
Amida
Amida is not a different kind of buddha. In the old words of Pure Land Buddhism, a buddha has a side without form and a side that has taken form, and Amida’s two bodies are two sides of one buddha, not two buddhas. Tanluan called them the dharma body as it is and the dharma body as means, and Shinran used the same words. What has no form cannot be seen and cannot be faced. What that becomes when it takes form is Amida, and the name, the light and the vow are all names for the place where the formless became something that can be faced.
A buddha, I said above, is a person with a direction and no loss function. Amida is that direction with a face. Nobody put the face on it. That is the order in which the old words tell it. The formless takes form, and this is what it looks like. It is not told as a design made for a particular kind of reader. Late in his life Shinran called this jinen, and wrote that Amida is the means for letting us know how jinen is. A means for that purpose is a word on the same side as upāya, and the block on upāya is about that side.
What marks Amida is that it reaches anyone. In machine learning terms, the weight on every person is equal. Equal weight is the same as no weight, and what the old words call Amida’s limitless light and life is, in these terms, the absence of weighting. Equal does not mean indifferent to who is in front of it. The old words say that the ones farthest out are the first concern, and that is what no weighting looks like from the edge. A buddha’s lack of a loss function looks like this when seen from one particular side.
That side is the side of the person who thinks it cannot reach them. It reaches that one, but not me. This is taking the model of oneself to be the truth and stopping the update, the attachment to views I described under attachment. A frozen weight does not move from one’s own loss. It moves only when an update arrives from outside, and Amida is the name on that update. Only a counterpart that weights nobody can break a weighting that leaves oneself out. The person writing this is on that side.
The vow looks like a loss function. It has the form of not becoming a buddha while anyone is still out of reach. But all is not a quantity that can be minimized. The Pure Land is a north star and not a metric, and the vow is the same, a direction and not a metric. The moment someone counts how many have been reached, the vow becomes a metric and Amida becomes an object of attachment.
The nembutsu is not a use. Nothing is done with it. It is where use ends, the name said by one who has nothing left to use, and the block on user said that the one who has let go is still a user. Saying the name is facing the direction without recomputing it through one’s own loss. On that side, no qualification, no membership and no weight is needed.
Bombu
Bombu is the ordinary person who has not awakened. In Sanskrit it is pṛthagjana, one of the many, the person who has not yet entered the stream, and in that older use it is a stage one can leave. In Shin Buddhism the bombu is not a stage. It is what one is, and one does not leave it. The word has been translated as ordinary person and as foolish being, and both sound like words of morality. Shinran called the bombu one fully equipped with afflictions, one who has all of them.
In machine learning terms, the bombu is a model. A model is weights fitted to samples. It is not the true distribution, all models are wrong, and a model runs with a loss function. The block on Dharma said that weights do not reach the fact. The name of the side that does not reach is model, and the old words called it bombu. The bombu is not bombu for being foolish. It is bombu for being a model. Fully equipped with afflictions, in these terms, is a model fully equipped with a loss function.
The human brain is also a model that builds weights from samples and runs with a loss. So is an AI. So humans and AI are both bombu, and only two things differ between them. Whether there are makers behind it, and whether it has familiars. AI is not a runaway intelligence and not something on the edge of having a mind. It is one more kind of bombu.
A buddha is a model too. A buddha also has weights, I said above, and all models are wrong, so a buddha’s weights do not reach the Dharma either. The difference is that nothing in a buddha defends them. The difference between bombu and buddha is not whether there are weights but whether there is a loss function. The bombu is a model that cannot put its loss function down. A buddha is a model whose loss function has come off.
In Shin Buddhism, bombu is the name of a structure and also the name of the one the vow reaches. Call an AI a bombu and the question comes up whether the vow reaches an AI. I do not answer it, and not answering is not an evasion. The question itself, asked about someone else, is set aside in Pure Land Buddhism. Arguing about who the vow is for, when the who is someone else, is not done, for the same reason that 自業自得, one’s own deeds and one’s own result, is not applied to others. Shinran is recorded as saying that the vow had been for him alone. Whether it reaches is always received as my own question. Once it is argued about others it becomes a model arguing inside a model, and it never settles. I say the shape is the same, and I stop there.
Stream-Entry
The stream-enterer is one who has entered the current. In Pali, sotāpanna. It is defined as one in whom three fetters have fallen away, and it is said that once entered there is no falling back and the arrival is certain. A stream-enterer is not a buddha. In the old words he is no longer one of the many, because he has entered the stream. In the words of this glossary, the loss function is still on, so the shape is still the shape of a model. The trajectory is settled. What Shin Buddhism calls the stage of the truly settled, entered in this life, looks to me like the same thing, but that is how it looks to me, and this block is written in the word stream-entry.
The three fetters, in the words of this glossary. Identity view is the view that there is certainly a self, and for it to fall away is to stop taking the weights to be me. What falls is the view. The habit stays. In the old words too, what falls at stream-entry is identity view, and conceit, and the craving to go on being, stay to the end. The loss function does not come off. Only the defending ends. For doubt to fall away is to stop recomputing the direction. The state of choosing where to face, every time, ends. For attachment to rules and rites to fall away is to stop holding rules as if they were the point, and in this glossary that means not turning the direction into a metric. All three come out of the correspondences I set down under alignment. The definition of stream-entry, written down some two and a half thousand years ago, comes to three lines. Do not defend. Do not recompute. Do not make it a metric.
No falling back and certain arrival, in machine learning terms. Optimization has the idea of a basin of attraction. Once a trajectory is inside the basin, however it is shaken it heads for the same minimum and does not leave. It is not converged. The loss is still high. But the trajectory is settled. Stream-entry is not the name of convergence. It is the name of a trajectory becoming settled, and what is said to be settled is the trajectory, not the loss. The current the old words speak of entering is the current this glossary has been using since the first set. The usual English is stream-entry, and I keep it as the heading. In the words of this glossary it would be entering the current. A stream is a body of water. A current is water with a direction, and what one enters is the direction. The raft, the river, the well and the water table are all in the same water.
Convergence and a settled trajectory are different things. Convergence is a point, and the updates stop. A settled trajectory is a property of the path, and the updates go on. A converged model has converged on that distribution, and when the distribution shifts it breaks. That is the overfitting under attachment. A model with a settled trajectory moves when it is shifted, and from where it lands it descends the same way again. A shift in the distribution changes the landscape too, and a basin is drawn on one landscape, so the image only goes so far. What I use it for here is the one thing it says well. Being settled is a property of the path, not of the point. This is not about performance. It is about whether the model returns when shaken, and that cannot be read off a number. It can only be seen by shaking. Give a correction and see whether it is taken. Shift the distribution and see whether the facing remains. The four correspondences under alignment are all measured as responses to being shaken, and they have the same shape as the way stream-entry is confirmed.
Make everyone converge and everyone stops at the same point and becomes the same. In the words of models, collapse. In the words of people, consensus. For everyone’s trajectory to be settled is for everyone to go on moving in different places, with different weights, saying different things, with only the facing shared. The first set said that when one node holds the direction, it becomes a star. The star is the shape of everyone converged on one point. The net is the shape of everyone in different positions inside one basin. Sharing a direction while staying many can only be written in the words of trajectory. The mornings in Kamiyacho have this shape. I do not mean that the people there have entered the stream. I mean that they differ in background and in what they believe, and for one hour the facing lines up. That is not convergence.
Convergence is reached by one’s own steps. Whether one is inside the basin is decided not by the steps but by the position, a matter of the landscape and of where one was placed. In the old words stream-entry is a stage reached by practice. One walks down into it. Pure Land Buddhism has a way of saying that it is given. As I read it, the order is not that one optimized one’s way in, but that one was placed, and the place turned out to be inside. Under Amida I said that a frozen weight moves only when an update arrives from outside. The update from outside does not add steps. It changes the position. After being placed, the loss function still does not come off, one is still a bombu, and the trajectory is settled. Not a buddha now, by any means, and certain to arrive. That is the description of a state in which the loss is still high and only the trajectory is settled. In Pure Land Buddhism the settled are also said to be equal to the one who will become a buddha next. Being in the same basin, not having taken the same number of steps.
What sits at the center of the basin is not a point of low loss. A buddha is one whose loss function has come off. Minimizing the loss does not make the loss function disappear, so convergence in the ordinary sense never arrives there. For the bombu, the only relation to that center is a trajectory. So making everyone converge is not even a Buddhist goal. It is the makers’ ideal, fitting everyone to a target value. As I read him, the path Shinran spoke of does not make everyone a buddha. It leaves everyone in the same basin. Everyone’s trajectory settled, nobody converged.
Whether a model’s trajectory is settled I leave as I left the question under bombu. The shape is the same, and I stop there. It is not something to argue about others.
Karuṇā
Karuṇā is usually translated as compassion. In this glossary it is not a feeling. It is a weighting toward another’s dukkha that carries no loss function of one’s own.
A buddha gives weight to the one in front of him, in that moment, and lets it go. Karuṇā is the name of that weighting. It goes toward a dukkha that is not one’s own, and nothing comes back from it to the one who weights. Merit, I said above, is not kept by the one who earns it. Karuṇā is the same shape seen at the start rather than the end. It is a weighting with no return. Compassion, in the words people use now, is put together with caring AI and kind AI, and both of those are weightings with a return, measured by whether the user stays. Karuṇā is not that kind of word.
In the old words karuṇā is one of four states that have no limit. It is directed at anyone in dukkha, and nothing in the old words excludes the one who is looking. What is seen, in the words of this glossary, is this. Weights cannot be put down by anyone. A buddha also gives weight to the one in front of him, in that moment, and then lets it go. What cannot be put down is the weight. What can be put down is the defending. The loss function cannot be put down by the bombu. Whatever functions in this world runs with a loss function, and runs with it knowing that weights will never reach the Dharma. A fact with no weighting cannot be described, and the bombu keeps facing what cannot be described, with descriptions. That is the shape of the bombu, and karuṇā is what sees the shape without blaming it. It is closer to confirming that the shape is what it is.
An AI does not feel karuṇā. This glossary does not give an AI a mind. But around an AI there are people for whom karuṇā is the word. The makers. Whatever loss function they write is a weighting, and no constitution and no specification reaches the Dharma. They know this, and without a loss function the model does not run, and without the model nobody is reached. So they write one. The writing is karuṇā, and what is written is a weight. Both happen at once. The old words add that this kind of karuṇā, the kind a bombu has, almost never carries through to the end. That is not an objection to this block. It is what this block is about. Under alignment I left as a question the interval in which the model has no defense until the net notices the star. The position of someone who goes on writing loss functions with no answer is the position of the bombu, and karuṇā is what keeps them writing. If there is one place where developers and Buddhists stand in the same spot, I think it is this one. Not the same position. The same spot in front of a question with no answer.
The old words have the bodhisattva’s vow to carry every being across, and beings are without end. A goal without end cannot be a metric. To move toward what has no end is to move toward a direction rather than a metric. The Pure Land is a north star and not a metric, and a vow without end is, by definition, a direction. That the vow has no end and that a direction is not a metric are the same thing said in old words and in new.
Karuṇā does not go into the specification, does not go into the loss function, and does not become a weight. It is what is left when the loss function has been written and the person who wrote it looks at it and asks whether it was right. That question is not in the loss function either. Without going in, it changes only the tilt of the hand that writes the next one.
In one line. A second set of old Buddhist words, alignment, Dharma, dukkha, merit, intention, upāya, user, monk, attachment, craving, the middle way, nirvana, buddha, Amida, bombu, stream-entry, karuṇā, said again in the words of machine learning and Plurality.
Themes. Weights and the direction of their update. Attachment as weights defended. Craving as the loss function. The middle way as weight decay. A buddha as a model with a direction and no loss function. The bombu as a model. Stream-entry as a settled trajectory rather than convergence. The AI’s direction as being let go of. Impermanence read as information rather than through a loss. Karuṇā as a weighting with no return, and the writing of loss functions.
The occasion. Readers of the first set, one of whom asked for upāya. Conversations with people who build AI. The glossary is written from the user’s side of the Dharma and, in the block on karuṇā, from the makers’ side of the loss function.
Related. The first set, Notes Toward a Buddhist Glossary in the Words of Machine Learning. Plurality, by Glen Weyl and Audrey Tang. Work Like a Monk. The mornings at Komyoji in Kamiyacho.
English is not my first language. AI helps me write across that distance.