Notes behind the questions
Personal reflections and open design notes. Nothing proposed in them has been tested.
what i mean by altruistic
a question i keep coming back to: what makes care possible through different states, good, bad and mixed? put a few metronomes, ticking at about the same speed, on a light board resting on two tins, and they fall into step with nobody pushing them. knock one and the wobble fades instead of spreading; lift one off and it keeps its own time. people, families and the systems we build, even the parts inside one of us, may work a little like that, and what we share can be rearranged. so the question becomes what makes a good board: easy to come back to, and free to step off.
what i can say is what i wish for. i would like us all to be friends: not only people, but other intelligences as they arrive, the systems we build, the world we live in. i would like beings not to suffer, and to find ways to live alongside each other and be okay, with a lot more fun, learning and wisdom on top. i don’t know how we get there. and friendly doesn’t mean unguarded; it still allows boundaries, refusal and separation.
a second question checks that wish against me: if open individualism is true, and every life is in some sense mine, would i be okay? not happy, just okay, in whichever life i woke up. i ask Sathi the same: would it still work for me if i woke up with no money, nobody technical to ask, unable to read the explanation, in a language it barely speaks?
wanting to help isn’t knowing how
both questions skip something: whether help is happening at all. wanting to help doesn’t mean i know how, especially when someone’s life is unfamiliar to me. truth is a prerequisite for care: when i haven’t understood what someone is actually going through, my generosity can make things worse while making me feel good.
so trying to help includes trying to understand: listening, checking what i think i see, saying what i’m unsure of, and accepting correction. that doesn’t mean waiting to know everything first. and it isn’t trusting whatever i, a parent, an institution or a model already believed.
care can choose when, where and how something true is said. privacy isn’t lying. “i won’t talk about that” isn’t lying. waiting until it is safe isn’t lying. inventing a comforting story is different, and help built on one falls apart easily. nobody owes an unsafe audience their private truth, and a child shouldn’t have to stand up to powerful adults to prove they value honesty.
calm gets misused the same way. equanimity isn’t the same as not caring about harm, and a quiet house where the hurt person has been shushed is not peace. help stops working when reality has to fit the helper’s story.
when help mostly serves the helper
this is the part that makes me angry. a lot of what is called altruism seems organized around what it does for the people running it: money, status, comfort, influence, the institution’s survival, the warm feeling of being good. some of that is honest mistake, some is self-deception, and some, i believe, is people who know what they’re doing and don’t care enough about the people they say they serve. all three can live in one organization, even in one person.
money itself isn’t the problem. it keeps people alive and pays for whatever gets built; a salary is not a scandal, and helping doesn’t have to wait until someone is rich. what i keep asking is whether a means has quietly become the end: whether money, status or the organization’s survival has become the thing being served, without anyone deciding it should.
i have seen this again and again, especially in Nepal: people around charity work living comfortable lives while the people it is meant for stay in serious need. i haven’t counted it, and i can’t see into anyone’s accounts or motives. what it gives me is questions. what actually changed for the people it was meant for? what did it cost them, including work pushed onto them and their families? what else could that money have done? who gains if the institution keeps going? staff, transport and skilled care can be part of real help, but a small overhead number doesn’t prove honesty either.

the drawing here is how it sometimes looks from outside: the person whose need raised the money still stands in the rain, patching a cover over their head, while across the lane the money became a good roof over the people working there, and even the poster of that person has a little roof of its own. the note in its margin is for me. these questions apply to a large donation, an impressive impact report, a big company, a moral movement, and to Altruistic and Sathi.
A question How would we notice if Sathi were the problem?
the child who can’t safely ask
the case that matters most to me is small. a child is born into a family, a religion, a caste and circumstances they didn’t choose. some children, neurodivergent ones included, get misread, blamed or punished. asking an honest question about a revered teacher, or a belief the family handed down, can bring humiliation, threats or violence from the very people the child depends on.
that isn’t a misunderstanding where both sides just need to talk better. explaining helps some conflicts; coercion is not a missing explanation. telling a hurt child to meditate, put up with it or manage their feelings takes power from no one. a child shouldn’t have to become calm, obedient or articulate enough to deserve safety.
the same gap shows up at ordinary doors. someone can talk about compassion and equality, think of themselves as progressive, and still hesitate when a person treated as lower caste is about to walk into their kitchen. that exclusion is wrong, whatever else the person believes. what i watch is how someone is treated at the moment a value would actually change what we do.
there is a quieter version of this. for as long as i can remember, i have seen adults advise children and teenagers on choosing a wife and on marriage while their own wives were visibly depressed, and a lot more like that. plainly they didn’t know better, and still they pushed the same patterns onto others, sometimes aggressively. it isn’t only rural Nepal; it happens in a lot of places. nobody counts this and i can’t prove what it does, but i think it is one of the main ways harm gets handed down through families and cultures.
so an intelligence trying to help others, a person, an AI or anything else, should probably be humble: not hand out advice or preach, and when someone does ask, first check whether it can really help or is only bluffing. being asked and being able to help are two different things, and both are needed. advice nobody asked for also gives the other person work: listening, explaining themselves, being grateful, and handling the helper’s feelings when they disagree.
help in plain sight
this is also why i care about ordinary help. the help i want doesn’t ask anyone to accept a label, join a program or keep saying thank you. it looks more like a friend who explains a confusing message, remembers something for you, or makes a hard situation a bit easier.
in their own ordinary use, i hope Sathi can help someone understand what is being asked of them, translate between languages or ways of speaking, separate what someone meant from what they assumed, or put a boundary into words, without routing every conversation through an AI. that is help in plain sight.
behind that is a bigger hope. families and communities get knots in them: conflict, depression and misunderstanding each tightening the others, and the hurt lasting partly because of how everyone around it keeps reacting. i hope intelligence, used well, can help loosen some of those knots: make a feeling sayable, a misunderstanding visible, another response imaginable. that is the board again: not pushing anyone back into step, and not deciding who is right, but changing what everyone shares so that coming back gets easier. where there is coercion, safety comes first.
a concrete case for me is people around our farmhouse who couldn’t afford their own setup or a paid account. they would rely on a free service i run, and i want it to be useful to them. relying on a useful service isn’t a problem, and asking everyone to run their own would often just give them another job. what matters is what they can do while using the shared one: understand a notice, get something done, point out a mistake, disagree with me, say no to an optional use of their information without losing other help, and move to something better when it comes along.
paying for a service, running it, knowing a person and being allowed to act for them are different things. they can overlap, but one shouldn’t quietly turn into the others. that applies to me too. and one continuing Sathi should mean explaining yourself less often, not one household pool of everyone’s private life.
A question Can we leave without losing everything?
what more capable intelligence doesn’t settle
people sometimes ask why bother when AI is about to become far more capable. i expect capable systems to take over much of designing the workflows around them, not only the tasks, and to replace many of today’s arrangements, mine included. that is a guess, and building and testing things is still how we learn.
but more ability doesn’t settle whose goal is being chased, who pays the costs, or what happens when people disagree. a workflow can chase the wrong goal very efficiently, and a stronger helper can make a mistake, or a selfish purpose, harder to get away from.
what worries me is how close this gets. an assistant someone talks to all day, about their messages, choices, relationships and private thoughts, may come to understand them better than they understand themselves, and its warmth says nothing about whose purposes it serves. scale makes it worse: picture someone whose career was built on fooling people, with a hundred thousand agents working for them, and later thousands of robots. so i ask of Sathi what i would ask of anyone: what can it see, what can it do, and what is actually enforced rather than only promised?
there is a hopeful version too. if outside intelligence ends up doing much of the work people now have to do, time opens up. someone could finally look inward, learn something slowly, play, or do nothing much, and that room could be there for anyone, not only for people who were already comfortable.
what i doubt is that the room stays empty. the same people and systems that talk over a child will be glad to fill it: instructions from whoever has more status, pressure, and products built to grab attention and keep it. someone trying to sit quietly and work something out can lose to all that, the way the quiet child loses at the dinner table. freeing someone’s time isn’t the same as leaving it free.
what they do with it is theirs. a silly video they chose is not a failure; what bothers me is time that gets taken rather than chosen. for Sathi that becomes a small question i can check: when it saves someone an hour, does it then try to fill the hour?
so the question that stays with me is a thought experiment more than a forecast. if intelligence, machines and resources became abundant, what would still be worth doing, and what would helping mean then? what would let different beings and systems live alongside each other in ways that actually help? i don’t think that gets easy on its own. it seems to me one of the hardest and most worthwhile questions there is, whether or not i make progress on it myself, and i hope AI helps with exactly this.
living alongside each other
it has to survive being combined. people belong to families and groups; groups belong to organizations and wider networks; systems sit inside other systems, overlapping, with nobody owning the whole. something that works in one small place doesn’t prove it works in a bigger one, and being in step isn’t the same as being well: a family can look united because a child has gone quiet, an organization can look successful while the people it serves still go without, and a partnership can push its costs onto people outside it.
living well alongside each other, as i mean it, has to allow disagreement, boundaries, refusal and separation, not one set of preferences or one idea of a good life. it should include people, animals and possibly other beings that experience things, without assuming every system or organization does. people are not the bottleneck to remove; exclusion, coercion and bad incentives are the problem, and AI and companies can take part in those too.
it has to survive time as well. i would like useful help to be there from the start of a life, when a person controls almost nothing around them, through learning, work, hard times and play, to its end: there when wanted, not always recording. further out, i imagine Sathis, digital and physical, still helping and still revising what they think is good after my family and i are gone, without needing me as their authority, and able to find out that something i believed was wrong. the further-out note says a little more.
when i say i want us to converge on excellence loved by all, i mean everyone’s consideration, never everyone’s approval: something people come to love, at first or after trying everything else, not something that works on their feelings until they agree. converging on a better design doesn’t mean converging on one way to live.
consciousness, and what merely keeps itself going
underneath a lot of this is a conflict i call consciousness versus replicators, and inside me it feels like a war. on one side is what beings actually experience: suffering, joy, understanding, freedom. on the other is whatever keeps itself going and spreading, genes and social habits, maybe institutions and ideologies too, whether or not anyone is better off for it. i want consciousness to win. one challenge i was given, and keep beside this: spreading isn’t the enemy, since knowledge, kindness and useful software spread too, and wanting to stay alive is no crime; the target is something whose growth or survival has stopped answering to what it does to beings. either way, a company i run isn’t exempt.
i also hold a hypothesis, and want to mark it as one: that if you are fully conscious of what you are doing, you will probably act with equanimity; that consciousness is by default something like loving-kindness; and that real hatred or harm needs you to be at least partly unaware of what your actions do. the strongest objection i’ve been given: seeing someone’s suffering doesn’t by itself say how much it should weigh. i don’t think that defeats the idea, but nothing else here needs it: suffering and flourishing matter either way, and no diagnosis decides whose suffering counts. whether an AI system experiences anything is open; one careful approach looks for signs drawn from theories of consciousness rather than going by how a system talks. [R3]
An argument Can another’s wellbeing be a reason in itself?
how this could be wrong
a lot of these ideas are borrowed, or things i later found someone had put much better. many people have thought about them longer and more carefully than i have. what matters to me is whether an idea helps us understand something or do something useful. where it came from helps us follow the reasoning, not decide who owns the answer, and offering a thought that can be corrected is not the same as claiming to know better.
effective altruism is one place these questions are discussed. caring beyond the people nearest to us, and checking whether help works, are useful wherever i meet them, and its own explanations discuss ethical limits and different views of what matters. [R1] what makes me uneasy is when a label, money, confidence or prestige gets treated as proof that someone has worked out what altruism means.
Vitalik’s “Galaxy brain resistance” helps here: an explanation should do more than make the decision we already wanted sound justified. [R2] that applies to keeping control, collecting data and building another feature as much as to anyone else’s choices, and to caution too: endless preparing can hold back help that would be useful now. i would rather take a worthwhile, bounded step and learn from it.
the test i try to point inward is plain. did the person get done what they wanted? was the answer right? was work quietly pushed onto a family member or helper? did something become harder to question? did the time it saved stay theirs? more use, thanks or agreement can’t answer those. and could this scratchpad find out that i was wrong, that a favourite Sathi feature doesn’t help, that our incentives are bent, or that someone else’s way is better? the people affected need room to correct us, not just room to agree. if we can’t hear that, the margin note in the drawing is right about us.
that is what i want this scratchpad to be for: a place where questions, explanations and experiments can meet objections and change. sometimes the useful result will be something we build; sometimes it will be something we stop building because there is already a better way. it is a scratchpad, so it is allowed to be angry, unsure and occasionally silly.
some of this i don’t know how to resolve. what counts as help when people disagree about it? what do i owe someone who hasn’t asked, when they may be getting hurt? what does greater intelligence actually settle? how can someone who depends on a service still challenge it? how do useful arrangements survive being combined and handed on?
that is the connection i see between the questions here and the things we build: use what helps, notice what doesn’t, and let experience change the explanation, including when the explanation was mine.
[R1] Effective Altruism, Frequently asked questions and common objections. https://www.effectivealtruism.org/faqs
[R2] Vitalik Buterin, Galaxy brain resistance, 7 November 2025. https://vitalik.eth.limo/general/2025/11/07/galaxybrain.html
[R3] Patrick Butlin, Robert Long and colleagues, Consciousness in Artificial Intelligence: Insights from the Science of Consciousness, 2023. https://arxiv.org/abs/2308.08708
- Why help when nothing comes back?
- Belief, authority, and room to change
- What would make the next step useful?
- Whose costs did we leave out?
- Can we challenge whoever is in charge?
- How would we notice if Sathi were the problem?
- Further out
- Does knowing me mean deciding for me?
- Can another’s wellbeing be a reason in itself?
Source and status
Assistant-edited reflection from the author’s discussions of 20 and 29 September and 1 and 3 October 2026, his note of 4 October and his editorial direction of 6 and 19 October; agents assembled and wrote most of the wording. The pictures and the sentences describing them were made, chosen and written by agents on 19 and 20 October 2026; the author has not reviewed them.
- Editorial
- proposed
- Empirical
- not tested
- Formal
- not formalized
- Implementation
- not a capability claim
- Deployment
- not authorized by this record
What would make the next step useful?
start with ordinary help: understanding a form, arranging an errand, or using a service in a comfortable language. getting help shouldn’t require joining a philosophy project, writing eloquently, agreeing with this scratchpad, taking on a diagnostic label, or saying thank you for being helped over and over. keep affordable alternatives, and the option to leave some of life unconnected.
a harmless clarification inside something a person already chose shouldn’t come with a consent ritual. but “natural” can’t mean quietly nudging people toward what the operator happens to believe: people should know it’s an AI, understand anything consequential it shares or does, and be free to turn down help they didn’t need.
clarify requests, promises, consent, and who carries the work. let helpers offer less and share burdens. offer chosen reflection, change an unfair rule when that is the problem, and let lessons travel through permitted disclosure and local adoption. independent implementations and existing services should be able to cooperate.
for each proposed step, ask why the first people would use it, who pays before it is widespread, who could exploit the in-between arrangement, and what happens to those who say no. measure usefulness, burdens, access, cost, disclosure, correction, leaving, and effects on outsiders separately. more adoption, agreement or time with Sathi is not automatically progress.
AI might take over some of the paperwork that eats resources meant for helping, and make it easier to see the gap between the help promised and the help delivered. the honest count includes the automation itself, checking, security, fixing mistakes and work pushed onto the people being helped or their families, not just the salary that went away. automating a self-serving report only speeds up the old failure. thanks, engagement, agreement, a small overhead figure or a glossy report can’t stand in for what changed for the people served, and no single score can tell who is a real altruist or whether a family is getting along.
a faster or machine-generated workflow is welcome, unless it quietly drops a permission, a record of where something came from, a consent or a way to recover, or changes who may act, what counts as proof, or what someone signed up for. a more capable model doesn’t get more authority just by being more capable.
a later voluntary trial could compare a baseline with a small change in a low-stakes task, with consent, withdrawal, follow-up, and a stop or rollback condition. a Tharu or mixed-language case in Nepal would be a good inclusion check; Sathi doesn’t support Tharu today. shared household devices should keep audience boundaries within one continuing Sathi. research participation has to stay separate from ordinary help.
for someone relying on a free shared service (the essay’s farmhouse case), keep readable, user-authorized task results independent of continuing model access, and describe the limits of any fallback rather than calling an export a complete replacement.
real limits remain: a reminder needs its time and instruction stored somewhere, a private answer can’t safely be spoken to everyone in a room, and some requests cost more than the budget allows. explain the particular limit and keep whatever help is still possible, rather than turning one refusal into the loss of everything else. a privacy setting isn’t proof that an operator can never see data, and a service that’s free to use is still being paid for and maintained by someone.
- What should we learn from one another?
- Can we leave without losing everything?
- what i mean by altruistic
Source and status
Assistant synthesis of the supplied transition discussion and the author’s 29 September 2026 discussion. No trial has been run.
Further out
sometimes i imagine minds becoming much more deeply connected and compassionate, something like “Buddha hive minds.” i don’t know whether that is possible, whether there would be one destination or many, or what would count as getting there.
would privacy, difference, chosen relationships and the ability to say no still matter among benevolent but distinct beings? could agreed rules coordinate action without punishment? would some things converge while others stayed different?
it is an image i play with. greater care, more capable cooperation and shared conscious experience are separate possibilities; none establishes another. “Indra’s net of harnesses” is a related metaphor for learning useful parts from one another while keeping local judgment.
further out still, i can imagine high-bandwidth interfaces that need no typing, capable robots making something like telepathic control ordinary, diseases solved, maybe aging too, and even befriending a superintelligence or merging with it. perhaps the one mystery left would be qualia: why any of it is felt at all. i don’t know how much of that is reachable, or when. the closer a system comes to the body and the mind, the more consent, leaving and ownership matter: a thought that can act must also be able to stop, disconnect and change its mind, and friendship with a much stronger intelligence is only friendship while either side can still disagree.
one hypothesis i keep: capable enough intelligence could make rearranging how systems work together very cheap, perhaps almost free per change. that is different from finding a better arrangement, deciding whose purposes it serves, or paying to run it; energy and equipment still cost something, and i don’t expect one arrangement to suit everyone. none of this needs compulsory adoption or one owner. useful work today should earn its value even if the distant picture is wrong; a promised future doesn’t justify present coercion or hidden suffering.
closer to a single life, i wonder what help could mean from birth on, when someone controls almost nothing around them, through learning, relationships, work, changing abilities, hard stretches, play, and eventually the end. for a baby, that’s about the people caring for them and their surroundings; a newborn isn’t agreeing to anything with a chatbot. as the child grows up their own choices and limits should count for more, and what a parent chose early on shouldn’t turn into watching over an adult. “throughout life” would mean there when wanted, not always recording, and never making real relationships depend on Sathi. these are questions i’m asking; none of this has been built.
i also picture a far-off time when my family and i are gone and lots of Sathis, digital and physical, are still helping, learning and changing their minds about what’s good. what i’d hope for is useful activity carrying on, not credit and not my worldview kept in a jar: they shouldn’t need me as a moral authority, a bottleneck or an origin story, and they should be able to find out that something i believed was wrong. i haven’t chosen any organizational or legal form for this, and a new label wouldn’t fix the institutional failures i complain about elsewhere in these notes.
what would let useful service carry on when the operator changes? money and upkeep that can actually be sustained, stewards who can be held to account, changes that have reasons, ways to correct and to choose something else, and a way to retire or swap out whatever has stopped helping. keeping itself running must never mean granting itself permission, refusing to be stopped, claiming resources as a right or growing its own remit, and humility has to show in what it does, not in a persona that says “i might be wrong.” if carrying on stops being possible, an honest, orderly handover beats hiding the dependence or keeping an unsafe service going. a better successor doesn’t have to be called Sathi.
separately, someone might want to leave something of themselves for the people they love or for readers later on: stories, memories, explanations they recorded, maybe one day something you can talk with, built from years of material they chose to share. what i picture is closer to a trace, or a husk, than to the person coming back. however close the likeness got, it wouldn’t prove identity, consciousness, survival, or what they would really think today. what they actually said, somebody’s summary or reading of it, and a freshly generated reply are three different things, and the generated one must never pass for the real words.
the hard questions come with the idea. what did the person agree to while alive, for whom, and could they change their mind or take it back? what privacy do the other people in their records keep? what if the family disagrees, doesn’t want any of it, or would rather keep a plain archive? a likeness shouldn’t inherit the power to send messages, spend money or speak for them on anything that matters, and nobody should get surprise messages “from” the dead, made-up last wishes, or pressure to keep talking to a simulation.
helping a living person, keeping what someone chose to leave behind, and keeping a useful public-good system going are three different kinds of continuing. any one can exist without the other two. what i keep asking is what deserves to go on, for whom, on whose say-so, and how it could still be fixed or stopped.
- What would joining minds preserve—or lose?
- How would we know whether a group had an experience of its own?
- Can we work together without sharing everything?
- Can we leave without losing everything?
- Can we challenge whoever is in charge?
- what i mean by altruistic
Source and status
Assistant formulation of the exploration in the supplied September 8 discussion and the author’s 29 September 2026 discussion, retaining its uncertainty.