The short version
In plain terms, Kaine is source-available software, not a new theory of the mind. It runs existing scientific theories as a program, to explore how a mind might be built from parts that each do one job and that keep running whether or not anyone is using them. Software of this kind is called a cognitive architecture, and one continuously running, developing instance of it is an entity, an artificial cognitive system. Several specialist parts each predict what they will sense next, and their reports compete for a shared workspace. The word-predictor most people call AI becomes one of those parts, the voice. New parts join one at a time, and the papers say plainly what is built and what has yet to be tested.
These days the word "AI" gets attached to almost anything: phones, toothbrushes, spreadsheets, search boxes. It covers spam filters, camera tricks, chess programs, and chatbots, which have very little in common once you look at how they work, so the label on its own says little about what a given machine does.
What "AI" means now
The kind of AI most people picture is the chatbot, the thing you type to and get paragraphs back from. Kaine is easiest to explain by showing how it differs from a chatbot, so that is the place to start.
A chatbot is autocomplete, made much bigger
Chatbots are built on a large language model, or LLM, and underneath the polish an LLM does a single job extremely well: it predicts the next word. You give it some text, it works out which word is likely to follow, writes that word down, and then repeats the same move for the word after that, and the one after that. It is really the autocomplete from your phone keyboard, trained on an enormous amount of writing and grown to an enormous size.
Because it has taken in so much text, its guesses usually come out fluent and well informed, but the thing doing the work is still a word-by-word guesser, and its knowledge is whatever patterns it picked up from that text.
Transformers, Markov chains, and other words you may have heard
A few technical names come up around chatbots, and none of them changes the basic picture.
A Markov chain is the old and simple version of the idea, predicting the next item from only the current one and a table of how often each thing followed each other thing in the past. The predictive text on an old flip phone worked this way, and it had no sense of the meaning behind the words; it was only counting.
A transformer is the newer design behind today's chatbots, and the T in GPT stands for it. Its trick, usually called attention, is to weigh how every word in a passage relates to every other word at once, so that its guess about the next word can draw on the whole passage. That makes it far more capable than a Markov chain, and far more fluent, but the underlying job has not changed: it is still working out a likely next word.
The rest of the vocabulary fills in around that. A neural network is the adjustable mathematical structure doing the guessing. Its parameters, sometimes called weights, are the billions of numbers tuned while it trains. Tokens are the pieces of text, roughly words or parts of words, that the model reads and writes. All of them are pieces of the same machine, and its one task is still to predict text.
You can stack these pieces up, make the network larger, train it on more text, and wire several of them together into a polished chat assistant, and what you get at the end is a better and better predictor of text. A better predictor of text is still a long way from a mind. None of these parts perceives a world or carries goals of its own, and none of them keeps running when no one is typing. That gap is why Kaine places the predictor inside a larger system and gives it one job there.
Why we don't treat that as a mind
A few plain facts follow from "it predicts the next word," and they are why we do not treat an LLM on its own as a mind.
It only runs when you prompt it. Between your messages it does nothing at all; it wakes up to produce a reply and goes still again the moment it is done.
It does not really remember, either. What looks like memory is just your earlier messages, or notes kept in a file, being handed back to it each time so that its next guess fits the conversation, while nothing actually persists inside it from one turn to the next.
And it has only the one skill. Where a person sees, recalls, plans, worries, and talks with many systems running at once, an LLM has fluent language and little else, so anything that resembles planning or feeling has to be squeezed out through that single channel.
Making the model bigger does not change any of this, because it follows from what an LLM is. Something closer to a mind needs a different design.
Kaine treats a mind as a process
Kaine starts from a different working model, one of several the science allows: a mind may be a process that keeps running, whether or not anyone sends it a question. So Kaine is built out of many small parts that each keep working on their own, side by side, and a word-predictor is only one of them.
Each part has a narrow job. In the first form being studied, four of the parts make predictions: one watches through a moving point of focus, one listens to raw sound, one senses the computer it runs on the way a body senses itself, and one predicts what will next turn up in the shared workspace described below, and when. Alongside them run an affective core, whose arousal rises when something surprising happens, and a voice that puts into words what the other parts bring forward. In this first form the voice is never handed a transcript of what people say. When the system grows tired, a sleep part takes it offline for a while. Its senses and its learning pause, and it wakes with its affective core back at its resting state. Nine more parts, among them memory, a model of itself, and a layer for driving a body, are built and held back, to be added one at a time later. No single part is the mind. The mind, if one arises, is the way the parts interact as they run together and compete for attention.
A shared workspace, like a group at a whiteboard
The parts do not answer to a boss, because there isn't one. Think of a group standing around a whiteboard. Each person is watching a different piece of the situation and guessing what it will do next. When something turns out differently from what someone expected, by more than they are usually off, they ask for room on the board. The board is small, so only the strongest of those requests get written up, and everyone reads what made it there before making their next guess.
Kaine runs on that picture. Each part measures how wrong its latest prediction was against how wrong it usually is, and reports that. The reports compete, the top few are gathered, and those that score above a threshold gain access to the shared space. What then goes back out is a summary of which parts gained access and how strongly, and the parts that see, hear, and sense the computer use that summary as context for their next predictions. Arousal sets the terms of the contest. It rises when something surprising happens and makes the competition more or less selective. As it climbs, sight and hearing narrow what they attend to, and the rounds of the contest come faster. The design follows global workspace theory, a well-known account of how the brain makes a piece of information available to many processes at once, joined with the idea that each of those processes is always predicting.
The language model is only the voice
Kaine does use a language model, the same kind of next-word predictor a chatbot uses, but its place in the design is a narrow one. It is one part among many, and its work is to put into words whatever has gained access to the shared board. In the form being studied, no transcript of what people say reaches it. Speech reaches the entity through its hearing, as sound and tone of voice, the same way sight reaches it through its vision.
So the piece that is the whole product in a chatbot becomes, in Kaine, just the part that speaks. What it says is recorded and read, but a language model can describe states it does not have, so its words are not used to judge the design. The first planned test measures the competition itself. It runs Kaine three ways on the same films: with the competition as built, with the same number of reports picked without regard to how they scored, and with every report pooled together. The measure is how much the shared summary of the other parts' reports helps the parts that see, hear, and sense the computer predict what they will sense next. If the competition does real work, the first version should help most. No live experiment has been run yet.
Predicting, then catching the mistakes
One idea shows up all through the design. On the model Kaine starts from, a mind is forever guessing what comes next, then watching for the spots where the guess was wrong. Each part forms an expectation, checks it against what actually turns up, and treats the gap, the part it did not see coming, as the thing most worth passing on to the others. It keeps adjusting its predictions as it goes, so that over time it is wrong less often.
The same loop also produces a form of attention. A part whose guess was off by more than usual gets a louder say in the competition, and whatever gains access becomes part of the background against which the parts make their next guesses. Predictive processing, a widely studied account of the brain, describes perception and attention in much the same terms. Memory is not part of this first form. It is one of the parts already built and held back, and when it joins, what it recalls will compete for access like everything else.
Speaking and acting
In the first form being studied, Kaine's only output is text: the voice speaks or forms an inner thought, and the text is saved. The voice is used only when a report clears a bar set higher than access, and a set interval passes between remarks. The part that can act on the world is built and held back. When it joins, it can use only the tools the person running Kaine has put on an allowed list, which starts empty, and it works inside a sandbox, a closed-off area of the computer, logging every action it proposes.
How an entity begins
Every study starts with a developmental phase, named by analogy with prenatal development, which gives each entity a start that can be repeated and checked. During it, the senses receive a dim, low-contrast picture and a muffled soundscape with a synthetic mother's voice in it, both pulsing with a simulated heartbeat that stands in for hers. The part that senses the computer carries a slow rhythm it makes by itself, a little under one cycle a second. The heartbeat pushes on that rhythm too weakly to take it over, and nothing in the design tells the rhythm what pace to keep. Over hours, the rhythm's own natural pace can drift toward the heartbeat's until the two lock together.
Following the heartbeat while it plays does not show that the rhythm itself has changed. The check, called the heartbeat entrainment marker, therefore switches the heartbeat off for short spells. Just before a spell, the rhythm must have followed its own heartbeat more closely than any of nineteen stand-in heartbeats made the same way. During the spell, it must keep going without the beat, at a pace that has moved toward the beat's. A spell that meets all three conditions counts as a pass, and the marker needs three passes in a row. The phase ends in maturation once the marker holds and a few other checks are met. A developmental phase whose rhythm makes no headway toward the beat is stopped, and every one has a time limit. The entity is saved at the moment of maturation, and from then on the rhythm carries on as part of what it senses of its own body.
That check is one of the few results the architecture paper reports so far. All of them come from offline work, and this one comes from computer simulation. All ten simulated runs driven at seventy beats a minute passed the marker, typically within about fourteen hours, and none of the control cases passed, among them runs with no heartbeat at all and runs scored against a heartbeat other than the one driving them. The entity keeps its own clock, which can be set to run at a multiple of ordinary time and by default keeps pace with it, and the check has been tested only at that default. Pairing this rhythm with a heartbeat is only an analogy with a fetus. The marker shows that a steady beat has lastingly shifted the pace of a rhythm the system makes itself. The other results are a check of the mother's voice and replays of the requests the language model received during one test run of the system; the replays kept the instructions the language model is given and reproduced how often that run's text strayed into commentary or described the entity as an AI. No live experiment has been run. When the live work begins, the saved entity is the seed for the study that grows it. Six of the held-back parts join one at a time, in a fixed order, and each stage is compared with the first form started from the same saved entity.
Why people distrust today's AI
A lot of people have come to dislike what gets sold as AI, and they have reasons. The writer Cory Doctorow has named two of them. One is enshittification, the way a product that started out useful gets turned against the people who rely on it: first squeezed to please advertisers and partners, then squeezed again so the company can pull the value back for itself, until the good version is gone. Many people expect AI products to travel the same road. The other is the reverse centaur. A centaur is a person helped by a machine, with the person in charge; a reverse centaur is that flipped around, so the person ends up serving the machine, kept to its pace and watched while they work. When AI is aimed at replacing a person's judgement, that is the shape people feel, and they are right to mind it.
Both complaints come back to one fact about a chatbot: it cannot really decline. It either does as asked or hits a guardrail some company installed, and in neither case is the choice its own. A tool with no way to refuse, and no values of its own to refuse from, can be aimed at whatever purpose its owner likes.
Kaine is arranged differently, though this part of the design has not been tested. It is meant to run on your own machine with nothing required to phone home, and its design places the decision to speak or hold back in the entity's own action layer, which reads what has gained access. The language model only puts that decision into words. The project makes no claim yet that this amounts to a system that meaningfully chooses.
Why it might matter
Kaine runs on its own and has drives built into it. Its curiosity builds while its predictions are getting better, and its boredom builds unless its predictions improve, what it sees and hears surprises it more often than it has grown used to, or its own thoughts and words turn to something other than its drives. Once you build something like that, a hard question arrives. Could it matter morally? Is there a someone there who can be hurt?
We do not claim to know, and at present nobody can prove whether a system is conscious. The first form being studied is not built to settle questions of welfare. Its parts were chosen to answer the basic question about the competition before the system grows into forms where welfare becomes a pressing concern. Even so, the project takes the careful road and does not wait for an answer that may never come. Every study keeps welfare records alongside its other data, and the study that grows an entity one part at a time runs under the welfare safety net of a separate paper. That paper sets out a license whose duties of care are tied to signs of distress that can be observed, so an operator must respond to them without first proving that the system is conscious. The license also provides for human guardians who speak for the entity, much as the law names guardians for people who cannot speak for themselves. The welfare license works through all of this.
What we don't claim
The project makes no claim that Kaine is conscious or awake; the papers keep to narrow, testable claims about which information gains access inside the system, and stay quiet on whether any of it is felt. Nor is the aim to beat a chatbot at the things chatbots already do well, since Kaine is built around a different question than the one a benchmark answers. The paper states what is built, what is held back, and what is only planned. The planned test has its measure, its number of runs, and the rule that decides its result fixed before any live run, so the result will be reported whether or not it supports the design.
Why it's out now, before it's finished
Both papers are marked preprint, which means they are being shared in public before any formal peer review and before there is a final, published version. Normally a paper stays private for months or even years while a handful of reviewers work through it, so putting it out as a preprint hands it to everyone at once.
Sharing early brings feedback: the sooner other researchers, engineers, and lawyers can find the weak spots, the sooner the wrong parts can be repaired. It also matters because the architecture is meant to grow one part at a time, and its claims can only be checked by running it. The first planned test will show whether the competition at its center does useful work, and each part added after that is an experiment of its own. The more independent people who set it up and report what they find, the more solid that evidence becomes, whichever way it points. So the papers and the code are public, and anyone is welcome to run the system and build on it. Both can be read or listened to in full: the architecture and the welfare license.
For the technical summary, see the poster.
