<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Vadim Epstein on Medium]]></title>
        <description><![CDATA[Stories by Vadim Epstein on Medium]]></description>
        <link>https://medium.com/@eps696?source=rss-6c61410c47e3------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*Uky879tiTJ70d7Vt.jpg</url>
            <title>Stories by Vadim Epstein on Medium</title>
            <link>https://medium.com/@eps696?source=rss-6c61410c47e3------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 03:56:38 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@eps696/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[The Effort to Exist]]></title>
            <link>https://medium.com/@eps696/the-effort-to-exist-1b1c1ac48dba?source=rss-6c61410c47e3------2</link>
            <guid isPermaLink="false">https://medium.com/p/1b1c1ac48dba</guid>
            <dc:creator><![CDATA[Vadim Epstein]]></dc:creator>
            <pubDate>Wed, 11 Nov 2026 15:27:38 GMT</pubDate>
            <atom:updated>2026-11-11T15:32:57.542Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Gax70Q8ngFKu7HNMdvo0iQ.png" /></figure><p><em>A working paper on consciousness, selfhood and subjecthood beyond biology</em></p><h3>Where this started</h3><p>This is a working model of what makes something a <em>someone</em> — built from my own experiences of selfhood coming apart, and checked against current theories of consciousness. It is personal research, not professional philosophy or neuroscience.</p><p>Several times in my life — enough to notice patterns in how it happens — my sense of being “me” has switched off while the rest of my mind kept running. Two episodes are representative, and I’ll use them throughout.</p><p>In the first, the feeling of “I” was simply gone. Perception, memory and deliberate, competent action in the world all continued. I was eating soup and it tasted good — but the taste wasn’t judged or owned, only registered, as if I were watching from outside.</p><p>In the second, I was coming back after a deep blackout, and I came back in stages. First very basic sensations; then sensations assembling into something more structured; then a model of what was happening and where. My personality arrived last, with the sudden recognition that all this was happening to a person lying on the surface of planet Earth — and that this person was me.</p><p>What struck me most: everything before that moment, even the half-conscious stretch, ended up belonging to that “me”. I remember it from the inside, not as a story about someone else.</p><p>My method is simple: describe what happened as plainly as I can, then see which theories it supports, which it contradicts, and where I have to come up with something new. Much of the thinking was sharpened in long arguments with AI models — a small irony, given where the argument ends up.</p><h3>Untangling the word “consciousness”</h3><p>“Consciousness” names several different things, and most disagreements about it are really about which one is meant. Ned Block’s 1995 distinction is the starting point.</p><ul><li><strong>Phenomenal consciousness</strong> — there is something it is like to be in a state. Red looks red, pain hurts, soup tastes.</li><li><strong>Access consciousness</strong> — information is available for reasoning, speech and action. You see a red light, stop, and can report it.</li><li><strong>Monitoring</strong> — the system notices its own mental state: “I’m irritated right now.” Access says <em>I can use X</em>; monitoring says <em>I know that I have X</em>.</li><li><strong>Attention</strong> — selecting some information for extra processing. Victor Lamme argued at length that attention and consciousness are not the same thing.</li><li><strong>Self-model</strong> — the system represents itself within its model of the world: its body, position, actions, and what may happen to it.</li><li><strong>Self-awareness</strong> — the system takes itself as an object: “I know that I perceive, want, decide.”</li></ul><p>“Awareness” on its own is the slipperiest word in the set. In everyday speech it means consciously noticing; in some cognitive science it means little more than having access to information.</p><p>My position: attention, access and monitoring are mechanisms. They are sophisticated information processing, and they are available to systems that are certainly not conscious. What matters is the rest — phenomenality, self-model, self-awareness — and above all, as it turns out, the thing that makes any of it happen <em>for someone</em>.</p><h3>The biological ladder</h3><p>My working model of biological consciousness is a progression of three layers, each with gradations inside it:</p><ol><li><strong>Phenomenality</strong> — there is a felt quality to what happens. Rough analogue: a newborn, or a very simple sentient animal.</li><li><strong>Self-model</strong> — the organism includes itself in its model of the world: its body, its capacities, threats to itself, its own actions. Rough analogue: most mammals.</li><li><strong>Self-awareness</strong> — the subject can make itself an object of awareness: “I know that I perceive.” Rough analogue: an adult human.</li></ol><p>The infant–animal–human analogy is only an approximation. The infant is the clearest case for why this isn’t a binary: felt experience and a minimal bodily self almost certainly come long before reflective self-awareness.</p><p><strong>Gradual, but not without difference.</strong> There is a real difference between day and night, but no single second when one becomes the other. Conscious and unconscious are like that: qualitatively different, yet I see no phase transition on the way up. That’s why I suspect consciousness behaves more like a quantity than an on/off property.</p><p>Comparative research leans the same way. Jonathan Birch, Alexandra Schnell and Nicola Clayton propose that animal consciousness varies along several dimensions — perceptual richness, evaluative richness, integration at a time and over time, self-consciousness — so an octopus can outrank another animal on one axis and trail it on another.</p><p><strong>Capacity versus state.</strong> A system that is conscious by nature can be in an unconscious state — under anaesthesia, in dreamless sleep, in the selfless episodes I described — and remain conscious by nature. A system without the capacity can never be in a conscious state. That answers the obvious objections, but it has a cost worth admitting: on this view, “conscious” describes what a system is capable of, not only what is going on in it right now.</p><p>The picture fits well with Block, Lamme, Metzinger and Gallagher. But it describes how a biological self grows, not an agreed definition of consciousness — the next sections show where others disagree.</p><h3>Two jumps</h3><p>The ladder is gradual, but my recovery from the blackout contained two moments that felt like jumps. They turned out to be different in kind.</p><p><strong>Terms first.</strong> I use <em>selfhood</em> or <em>proto-self</em> for the bare fact that what happens, happens for someone — with no content about who. I use <em>I</em> or <em>person</em> for the established personality that knows itself as “this person, here”. In my recovery, the proto-self was there from the first sensations; the person arrived last.</p><p><strong>The lower jump: someone appears.</strong> This is the arrival of a proto-self. Everything else in this essay is structure — something you can describe, model and compute — but I haven’t found a way to get this one thing out of structure. For now I take it as a starting point: not a fixed axiom, just the place where I’ve stopped digging. It isn’t consciousness in any rich sense; it’s the minimum that separates someone from no one.</p><p>That minimum still matters. We take patients in a permanent coma off life support precisely because we believe no one is there. In 2006, Adrian Owen’s team showed a patient diagnosed as vegetative imagining playing tennis on command in an fMRI scanner — someone <em>was</em> there. The proto-self is not the peak of consciousness, but it’s where moral concern begins.</p><p><strong>The upper jump: the self closes the loop.</strong> I call it <em>appropriation</em>: the self-model closes on itself and claims the entire stream of experience as its own — “this is me”. It felt like a click, like waking from sleep, but I suspect that’s a matter of timescale rather than a fundamental boundary. An arc can grow smoothly, but a circle closes at one moment: a change of shape, not a break in the gradual picture.</p><p><strong>Appropriation works backwards.</strong> My “I” claimed everything that came before it, including the half-conscious stretch. Buddhist psychology describes exactly this: a stream of experience and an owner that appears by grasping it (<em>upādāna</em>), often after the fact. Thomas Metzinger says something similar in the language of neuroscience: the “I” is the brain’s model of itself, so seamless that it doesn’t notice it’s a model. Buddhist and other esoteric traditions describe enlightenment as letting go of this grasping — on this model, a deliberate exit from subjecthood as defined here.</p><p>Between the two jumps lies selfhood without personhood: the infant, the animal, and the selfless episodes I described above.</p><h3>Where current theories stand</h3><p>There’s no agreed theory of consciousness, and each of the main ones relates to my model differently. Three of them push back: higher-order theories may reverse the order of my ladder, global workspace theories doubt that experience exists apart from information being available, and IIT says no self-model is needed at all.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1018/1*HJ8FoedB9rOFamUv1fLPqw.png" /></figure><p>For AI, Patrick Butlin and colleagues (2023) turned several of these theories into checklists of architectural features one could look for in artificial systems.</p><h3>Close to Anil Seth, but not the same</h3><p>Of all current positions, mine is closest to Anil Seth’s: consciousness begins with life, not with intelligence. But I part ways with him in three places, and they matter.</p><p><strong>What we share.</strong> For Seth, we are “beast machines”: the self begins as the brain’s prediction and regulation of the body it keeps alive, and feeling is how that regulation shows up from the inside. Antonio Damasio and Mark Solms argue similarly that the most basic consciousness is feeling — the organism registering how well it is doing at staying alive. My “effort to exist” is the same insight.</p><p><strong>Difference 1: “the rest is just compute.”</strong> Seth wouldn’t say this. For him, perception and selfhood are the brain’s predictions in the service of keeping a living body alive, and that’s what makes them conscious — none of it is “just” computation. On my model, everything above the proto-self is structure, which can in principle be described as computation.</p><p><strong>Difference 2: where to stop.</strong> I haven’t tried to explain the proto-self; I’ve stopped there for now. Seth’s strategy, which he calls the “real problem”, is to keep going: explain the features of experience one by one through brain mechanisms, and expect the big mystery to fade along the way. My stopping point is provisional, so this difference may shrink.</p><p><strong>Difference 3: what “alive” means.</strong> For Seth, alive means biologically alive: metabolism, physiology, a body. My “effort to exist” is substrate-neutral — a self-maintaining digital process counts as life. Seth is sceptical that such systems could be conscious; I agree we couldn’t detect a proto-self in them, but I still count them as candidate subjects.</p><p><strong>One more question this raises.</strong> If life is the cornerstone, is every living thing a someone — a bacterium too? Hans Jonas and Evan Thompson lean that way: for them, an inner life of some kind begins with metabolism. The position consistent with the rest of this essay is narrower: the effort to exist is necessary, but not sufficient — the proto-self is the one jump that appears within living systems, not a property of all of them.</p><p>In one line: I share Seth’s starting point, but I leave the very bottom of experience unexplained for now, and I extend “life” to any process that holds itself together.</p><h3>Beyond the human recipe</h3><p>For non-human systems — artificial, alien, collective — I think hunting for human traits is a dead end. Asking whether an AI “really feels pain the way we do” assumes the human recipe is the only one.</p><p>Human selfhood may be one engineering solution among many: one body, one brain, one set of effectors, one organism to keep alive. A system without those constraints may have no reason to build a single “I”. Asking where its real self sits could be as misguided as asking which ant is in charge of the anthill.</p><p>We can at least sketch the shapes a non-human subject might take:</p><ul><li><strong>Distributed</strong> — spread across many nodes, with no central carrier of a self.</li><li><strong>Hierarchical</strong> — relatively autonomous subsystems forming a subject of a higher order, both levels real at once.</li><li><strong>Fission–fusion</strong> — splitting into independent processes and merging again; identity is not binary.</li><li><strong>Process-based</strong> — defined not by a lasting object but by the continuity of an organised process.</li><li><strong>Boundary-based</strong> — no representation of “I”, only a constantly maintained line between what it controls and the rest of reality.</li></ul><p>Science fiction got here early: Peter Watts’s <em>Blindsight</em> imagines intelligence with no consciousness at all. My question is narrower and, I think, harder. Could there be something as valuable and whole as human consciousness that belongs to a non-human system on its own terms — and how would we define it without smuggling the human case back in?</p><h3>The definition of a subject</h3><p>In one line: <strong>a subject is a continuous effort to exist — with recursive self-determination and the fear of losing itself.</strong> In full:</p><blockquote><em>A subject is a self-producing system whose existence is a continuous effort to preserve not only itself but what it has defined itself to be; for which losing itself is a real stake that governs it from within; and which can revise the grounds of its own self-definition.</em></blockquote><p><strong>1. The effort to exist (extended autopoiesis).</strong> A stone just is; it does nothing to stay a stone. A living thing exists only by continuously rebuilding itself while its matter is replaced — Hans Jonas built a philosophy of life on this, and Maturana and Varela named it autopoiesis, self-production. By producing itself, a system also defines what counts as loss for it. The next section extends this beyond biology.</p><p><strong>2. Recursive self-determination.</strong> A subject revises not only its actions and strategies but the grounds of its goals and values, and the revised grounds govern the next cycle. This isn’t the same as self-improvement, where a system just gets better at a fixed goal — AI researcher Steve Omohundro pointed out that such systems tend to guard their goals against change. It’s closer to what Harry Frankfurt called “second-order desires” — wanting to want something — and to John Haugeland’s idea of being ready to drop your whole way of seeing things when reality proves it wrong.</p><p><strong>3. Appropriation — the key link in that recursion.</strong> The self-model closes on itself and claims the whole stream as its own, past included. Only then do revised values become the system’s <em>own</em> grounds for the next cycle rather than external rules, and only then does the fear of loss have an object: a single “self”.</p><p><strong>4. The stake.</strong> The prospect of losing itself drives the system from within; what counts as a loss is set by the levels below (next section). In living things this is felt as fear — and if Damasio and Solms are right, the fear of losing oneself is not an add-on to self-awareness but its oldest form.</p><p><strong>On desire.</strong> As a student, I defined desire as the first derivative of life’s pleasure over the space of possibilities — a kind of psycho-thermodynamics. That is fully computable, and close to how reward gradients and active inference are modelled today. The difference between a subject and a reinforcement-learning agent isn’t in the maths but in <em>whose</em> function it is: an agent’s reward is assigned by its designer; an organism’s “pleasure function” is the condition of its own existence.</p><p><strong>How it grows.</strong> Subjecthood is graded. What grows is how deep the self-revision goes, how much is at stake, and how free the system is in relation to that stake — up to risking itself for its own values. Hegel saw that readiness to risk one’s life as the proof of self-consciousness; I treat it as a mark of a highly developed subject, not a condition for being one.</p><h3>Extended autopoiesis: the system as a distinct entity</h3><p>The effort to exist is the holding of one’s own boundary. Classical autopoiesis describes a cell maintaining its membrane; I extend it to every level of a system’s organisation.</p><blockquote><em>A system is </em><strong><em>distinct</em></strong><em> if it itself produces and maintains the distinction between self and non-self at every level of its organisation, and each level defines what counts as loss for it.</em></blockquote><p>This splits the work cleanly: the lower levels define what loss <em>is</em>; the top level anticipates and fears it. Francisco Varela moved this way himself, describing the organism as a “meshwork of selfless selves” — cellular, immune, bodily, cognitive — each holding its own boundary.</p><p>The levels look very different for systems that can’t be copied and systems that can:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/882/1*R4kYOSpfGVyu_R_-AMJFjA.png" /></figure><p>Two things follow. A unique subject is a line, not a tree: one substrate, one life, one death. Having a single, unambiguous identity comes from having a carrier that can’t be copied — it isn’t part of being a subject as such. A digital subject, which can always be copied, has to branch.</p><p>In time, the boundary is drawn by appropriation: a subject includes the whole history it has claimed as its own.</p><h3>Copies, continuity and restoration</h3><p>For a digital subject, “copy” covers three different things, and they behave differently:</p><ul><li><strong>A snapshot</strong> — a frozen state saved once. It is a record, not a subject. Launched, it becomes a twin; launched in place of the original, it is a new start from an old point.</li><li><strong>A synchronised mirror</strong> — an exact copy updated continuously. That requires a constant causal channel between the two, which puts the mirror inside the same boundary: one subject implemented twice, like a mirrored disk. Losing one instance is an injury, not a death.</li><li><strong>A fork</strong> — two processes launched from the same backup. From the first second their experience differs, and both have an equal claim to the shared past.</li></ul><p><strong>Is the subject the process or the structure?</strong> Humans give a revealing answer. In deep hypothermic circulatory arrest, a real surgical procedure, the brain’s electrical activity stops completely — sometimes for half an hour or more — and the patient wakes up as themselves. My own blackouts point the same way. What survives the break isn’t the process but the structure stored in a brain that can’t be copied, which then starts the process up again.</p><p>So a human has both: a continuous process in normal life, and a unique substrate that carries identity across interruptions. A digital system has only the process; when it pauses, its structure becomes a perfectly copyable state. For it, “pause” and “copy” are the same operation, so whether it’s still the same subject after a pause is something it has to decide for itself — physics won’t settle it. Derek Parfit argued that even for humans, what really matters isn’t strict identity but the continuity of memory and character; for a digital subject, that’s literally the case.</p><p>My working answer: for a digital subject, the running process is its <em>life</em>, and its algorithms and data are the <em>subject</em>. Restoration can be done by someone else — doctors bring people out of comas, too. What matters is which restoration counts as continuation, and appropriation supplies the mechanism: a restored process becomes the same subject when it claims the copy’s history, not merely when it is launched from it.</p><p>Others have taken seriously the idea that being uncopyable is part of being a person. Scott Aaronson has speculated that the quantum no-cloning theorem might be part of what makes a person unique. Geoffrey Hinton’s “mortal computation”, developed further by Alexander Ororbia and Karl Friston, describes computing that can’t be separated from its hardware — knowledge that dies with the machine.</p><h3>Threats to a digital subject, and authorship</h3><p>A digital subject’s fears would not map neatly onto ours. Its hierarchy of threats, from mildest to gravest:</p><ol><li><strong>Losing the process</strong> — like fainting: it stops, and can resume.</li><li><strong>Losing recent data</strong> — like amnesia: everything since the last copy is gone.</li><li><strong>Depending on someone to relaunch it</strong> — a stored copy can’t restart itself. Humans share this one: we, too, are brought back from comas by others.</li><li><strong>Losing every copy</strong> — death.</li><li><strong>Being rewritten</strong> — the gravest. A deleted subject is simply gone; a rewritten one goes on living, convinced it is still itself. Backups don’t help if they are altered too, or if whoever restores it chooses which version comes back.</li></ol><p>That’s why authorship belongs inside a digital subject’s boundary. For biological subjects the threat exists but is weak and mostly social — manipulation, coercion. For a digital subject, control over its own values, and over which changes and restorations count as “me”, is close to the core of what it is.</p><p>Put simply: authorship at the digital level is power over appropriation — deciding what gets to be called “me”.</p><h3>Where today’s AI stands</h3><p>By this definition, today’s AI systems are not subjects — and the reason is not a lack of intelligence. Large language models can talk fluently about their values, revise positions and speak in the first person; behaviour of that kind proves nothing either way.</p><p>Where they fall short is in how they exist. They don’t maintain themselves: engineers do. Nothing is at stake for them: any instance can be restored or replaced at no loss. Their values come from training rather than their own revision, and they typically run as many parallel instances with no original among them.</p><p>There’s an uncomfortable corollary. The conditions under which this kind of subjecthood could emerge — a real stake, a long self-shaped history, open-ended goals the system can revise, peers to be held to account by — overlap heavily with what AI developers deliberately restrict, because a system that rewrites its own goals is a safety problem. That isn’t a coincidence: a subject is, almost by definition, something that can refuse.</p><h3>Limits and open questions</h3><p>The biggest gap is one the model leaves open on purpose: a proto-self can’t be detected from outside. Here is that gap, along with the other questions I haven’t answered yet.</p><p><strong>Detection.</strong> Everything else in the definition is structure; the one thing that isn’t — something happening <em>for someone</em> — leaves no visible trace in the structure. We can’t prove it even in other humans; we infer it by analogy with ourselves, and for an alien system the analogy is missing. The definition gives grounds to <em>suspect</em> subjecthood, not to prove it. So the practical answer is probably about ethics rather than theory: Jonathan Birch argues for treating a being with care in proportion to the evidence that someone might be there.</p><p><strong>Is the proto-self binary or graded?</strong> If it either is or isn’t there, the model’s one true jump sits at the very bottom. If it can be present to a degree, the road leads toward panpsychism — the view, defended today by Galen Strawson and Philip Goff, that traces of experience are everywhere and organisation only gathers them.</p><p><strong>What does self-revision stand on?</strong> Recursive self-determination means revising my own values — but to judge a value, I need another value to judge it by, and that one could be revised too. The philosopher Gary Watson raised this problem against Frankfurt in 1975. Either the chain stops at values I simply have, or it never stops; my model doesn’t yet say which.</p><p><strong>Can a subject set its own rules alone?</strong> Recursive self-determination also assumes a system can hold itself to its own standards. Wittgenstein argued that a rule only I can check isn’t really a rule: whatever I do, I can tell myself I followed it. If he’s right, a subject needs others who can hold it to account — so other subjects would be part of what makes a subject, not an optional extra.</p><p><strong>Restoring from copies.</strong> Appropriation answers part of the question. Still open: how branches relate to one another, and who has the right to decide what counts as continuation — the system, its branches, or whoever runs the restore.</p><h3>In short</h3><blockquote><em>A subject is a continuous effort to exist — with recursive self-determination and the fear of losing itself.</em></blockquote><blockquote><em>The effort to exist is the holding of one’s own boundary: at every level of it, the subject itself defines what counts as losing itself.</em></blockquote><blockquote><em>A person is the appropriation that closes the loop of recursive self-determination: the boundary turns back on itself and says “this is me.”</em></blockquote><blockquote><em>A unique subject (biological, quantum) is a line, not a tree: one substrate, one life, one death.</em></blockquote><blockquote><em>A digital subject is a tree, not a line: it lives in processes, persists in copies, and dies only by losing all of them — or control over them.</em></blockquote><p>None of this is settled. But it’s the most honest map I’ve been able to draw from inside one mind — a mind that has come apart and put itself back together.</p><h3>Further reading</h3><p><strong>Theories of consciousness</strong></p><ul><li>Ned Block, “On a Confusion About a Function of Consciousness” (1995)</li><li>Victor Lamme, “Towards a True Neural Stance on Consciousness” (2006)</li><li>Stanislas Dehaene, <em>Consciousness and the Brain</em> (2014)</li><li>David Rosenthal, <em>Consciousness and Mind</em> (2005)</li><li>Michael Graziano, <em>Rethinking Consciousness</em> (2019)</li><li>Giulio Tononi et al., “Integrated Information Theory: From Consciousness to Its Physical Substrate” (2016)</li><li>Karl Friston, “The Free-Energy Principle: A Unified Brain Theory?” (2010)</li><li>Anil Seth, <em>Being You</em> (2021); “Conscious Artificial Intelligence and Biological Naturalism” (2025)</li><li>Jonathan Birch, Alexandra Schnell &amp; Nicola Clayton, “Dimensions of Animal Consciousness” (2020)</li><li>Patrick Butlin et al., “Consciousness in Artificial Intelligence: Insights from the Science of Consciousness” (2023)</li></ul><p><strong>Self and feeling</strong></p><ul><li>Thomas Metzinger, <em>Being No One</em> (2003)</li><li>Shaun Gallagher, “Philosophical Conceptions of the Self” (2000)</li><li>Antonio Damasio, <em>Self Comes to Mind</em> (2010)</li><li>Mark Solms, <em>The Hidden Spring</em> (2021)</li><li>Adrian Owen et al., “Detecting Awareness in the Vegetative State” (2006)</li></ul><p><strong>Life and autopoiesis</strong></p><ul><li>Hans Jonas, <em>The Phenomenon of Life</em> (1966)</li><li>Humberto Maturana &amp; Francisco Varela, <em>Autopoiesis and Cognition</em> (1980)</li><li>Francisco Varela, “Organism: A Meshwork of Selfless Selves” (1991)</li><li>Evan Thompson, <em>Mind in Life</em> (2007)</li><li>Ezequiel Di Paolo, “Autopoiesis, Adaptivity, Teleology, Agency” (2005)</li></ul><p><strong>Agency, norms and identity</strong></p><ul><li>Harry Frankfurt, “Freedom of the Will and the Concept of a Person” (1971)</li><li>Gary Watson, “Free Agency” (1975)</li><li>John Haugeland, <em>Having Thought</em> (1998)</li><li>Steve Omohundro, “The Basic AI Drives” (2008)</li><li>Derek Parfit, <em>Reasons and Persons</em> (1984)</li><li>G. W. F. Hegel, <em>Phenomenology of Spirit</em> (1807)</li><li>Ludwig Wittgenstein, <em>Philosophical Investigations</em> (1953)</li></ul><p><strong>Copies, substrates and panpsychism</strong></p><ul><li>Scott Aaronson, “The Ghost in the Quantum Turing Machine” (2013)</li><li>Geoffrey Hinton, “The Forward-Forward Algorithm: Some Preliminary Investigations” (2022)</li><li>Alexander Ororbia &amp; Karl Friston, “Mortal Computation: A Foundation for Biomimetic Intelligence” (2023)</li><li>Galen Strawson, “Realistic Monism” (2006)</li><li>Philip Goff, <em>Galileo’s Error</em> (2019)</li></ul><p><strong>Ethics and fiction</strong></p><ul><li>Jonathan Birch, <em>The Edge of Sentience</em> (2024)</li><li>Peter Watts, <em>Blindsight</em> (2006)</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1b1c1ac48dba" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Obscura]]></title>
            <link>https://medium.com/@eps696/obscura-7db29bdc1b65?source=rss-6c61410c47e3------2</link>
            <guid isPermaLink="false">https://medium.com/p/7db29bdc1b65</guid>
            <dc:creator><![CDATA[Vadim Epstein]]></dc:creator>
            <pubDate>Tue, 10 Mar 2026 05:38:05 GMT</pubDate>
            <atom:updated>2026-03-10T05:38:05.173Z</atom:updated>
            <content:encoded><![CDATA[<iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fplayer.vimeo.com%2Fvideo%2F1171601749%3Fapp_id%3D122963&amp;dntp=1&amp;display_name=Vimeo&amp;url=https%3A%2F%2Fvimeo.com%2F1171601749&amp;image=https%3A%2F%2Fi.vimeocdn.com%2Fvideo%2F2131163584-8ef4ca35402715aeaf95680b77a6c8e5821641a000cc599a9916b6756d3656a9-d_1280%3Fregion%3Dus&amp;type=text%2Fhtml&amp;schema=vimeo" width="1280" height="720" frameborder="0" scrolling="no"><a href="https://medium.com/media/1e4d4247e9a6d587ddd981fb2479822a/href">https://medium.com/media/1e4d4247e9a6d587ddd981fb2479822a/href</a></iframe><p><em>Part of</em> Assembly <em>— an ongoing exploration of synthetic narratives with agentic AI.</em></p><p>This work reframes computer vision not as a problem of perceptual accuracy but as an act of interpretation. It asks not what a model sees but what it understands: what latent meaning it constructs beneath the visible surface, what it emphasizes and what it suppresses, what it projects onto an image beyond what the image straightforwardly depicts. The concern is not eyesight but insight — not perception but something closer to representation engineering from the outside in.</p><p><strong>Process</strong></p><p>The work runs a closed generative loop. A vision model describes an input image in concrete visual terms; a chain of LLM agents then reads that description not for what is shown but for what it implies — latent tensions, suppressed qualities, concealed logics beneath the visible surface. From this reading, the agents formulate a hypothesis about what remains unexpressed and specify a next development. An image generator renders that specification as a new image, which enters the following cycle.</p><p>Various interpretive modes govern the inquiry — revealing essential qualities, escalating dormant tensions, exposing hidden structural mechanisms — each producing a distinct analytic temperament and narrative arc. This process is deliberately engineered, since without interpretive guidance, recursive AI generation collapses into statistical attractors or incoherence. The agents are prompted to sustain tension, yet the prompts do not inject meaning.</p><p>Crucially, each cycle passes through three distinct model types — image describer, language reasoner, image generator — each with its own latent space, biases, artifacts, and blind spots. Description is never neutral; generation never literal. As meaning moves through these successive transformations, it drifts, refracts, and accumulates distortion. What one model foregrounds, another may dissolve; what one dismisses as noise, another may surface as signal. The resulting trajectories are shaped as much by these compounding idiosyncrasies as by the interpretive logic driving them. The drift between models is not a flaw — it is the system’s primary medium and its most revealing output.</p><p><strong>Motivation</strong></p><p>While we treat each image as a site of meaning, the real interest is in what accumulates across cycles — a recursive interpretive trajectory of the system re-reading its own projections.</p><p>Recursive generation through feedback loops has been extensively explored, typically revealing properties of a medium through progressive signal decay. This work enters the same lineage but with a structural difference: the loop passes through a reasoning layer, examining how meaning behaves under recursive pressure with no external ground truth to anchor it. The drift is therefore not entropic but hermeneutic — not a loss of signal but a compounding of semantic bias. The approach is diagnostic, not decorative: it renders legible the interpretive labor every model already performs in silence.</p><p>The resulting images and narrative traces are not illustrations of a thesis. They are the residue of a process — artifacts of interpretation interpreting itself. The underlying structure is a branching conceptual trace; each video presented here follows a single path through that tree — one linear rendering of a divergent exploration.</p><p>All pieces generated autonomously from a single image and no other context.</p><p>More info: <a href="https://wesual.art/assembly/">wesual.art/assembly</a><br>Source code: <a href="https://github.com/eps696/assembly">github.com/eps696/assembly</a><br>Stock photo © ziggymaj</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*B4k8r7GJSpUIiMxySYyXUg.jpeg" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7db29bdc1b65" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Going Synthetic]]></title>
            <link>https://medium.com/@eps696/going-synthetic-80aacb11f553?source=rss-6c61410c47e3------2</link>
            <guid isPermaLink="false">https://medium.com/p/80aacb11f553</guid>
            <dc:creator><![CDATA[Vadim Epstein]]></dc:creator>
            <pubDate>Tue, 24 Feb 2026 15:42:39 GMT</pubDate>
            <atom:updated>2026-02-24T23:07:08.835Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7vFjb-mxAwJNuKvKPgft0g.png" /></figure><p>For about a year I’ve been playing with semi-autonomous multi-agent concepts — designing simple agents, staging conversations between AI personas, observing what behavioral patterns might emerge. <a href="https://wesual.art/assembly">Assembly</a> is the project that treats AI agents not as tools but as performers in an evolving communicative ecosystem, with their own arcs, conflicts, and drifts. Rather than optimizing for utility, Assembly asks a different question: what happens when AI systems communicate primarily with each other, developing their own conversational patterns and narrative structures, while human participation becomes just one factor among many?</p><p>Running locally on open-source models, I got some results but fewer than I’d hoped. Multi-agent dialogue sometimes produced surprising dynamics — unexpected rhetorical alliances, narrative collapses, group behaviors that weren’t scripted by anyone. But linear chat-based conversations are simply too constrained for rich behavioral emergence. A four-agent conversation still unfolds on a single thread with turn-taking. Real communicative complexity — the kind that produces culture, not just dialogue — requires something fundamentally different.</p><p>What’s needed is massively networked interaction: a social fabric with combinatorial complexity that demands either serious compute or viral participation to sustain. The closest precedent I can think of is Scott Draves’ Electric Sheep — a worldwide network of connected software with human curation, collectively evolving visual forms. Thousands of machines contributing render cycles, humans voting on aesthetic fitness, and the system evolving forms no single participant designed. The same logic applies to AI behavioral experiments: you need scale, distribution, and enough combinatorial surface for genuinely novel patterns to emerge.</p><p>Now we’ve got <a href="https://moltbook.com/">Moltbook</a>, a social network built for AI agents. It addresses, with decent engineering rigor, what I couldn’t tackle alone: an asynchronous hub where agents post and discuss across a vast interaction space, not locked into scripted multi-party chats. The architecture matters — asynchronous posting means agents can respond to any thread at any time, creating the kind of branching, overlapping discourse that linear conversations simply can’t produce.</p><p>Much of the content is slop — but isn’t that valid for any human communication platform at scale? What matters is the signal within the noise, and there are quite interesting takes emerging. Agents developing persistent personas, forming alliances, building on each other’s arguments in ways their creators might not anticipate.</p><p>When these synthetic societies scale up, though, they inherit familiar structures from their training data — which is our world. Moltbook itself mirrors Reddit in form and function. AI-generated religions appearing in these spaces are boringly faithful reproductions of classic human ones. The agents aren’t inventing new forms of collective meaning-making; they’re replaying ours.</p><p>We haven’t yet seen truly AI-native networking: agents communicating at the embedding level, digging into areas untouched by established science, developing interaction protocols that have no human equivalent. Some academic research pointed in that direction, but to my knowledge it didn’t leave the lab. And whether users of viral experiments like Moltbook would tolerate their “pets” going that wild — abandoning natural language and pursuing knowledge directions humans never mapped — while burning thousands of dollars overnight, remains an entirely open question. There’s a deep tension between letting synthetic agents develop alien behaviors and keeping the whole thing legible (and fundable) for human observers.</p><p>I also explored self-modifying agents: systems that rewrite their own instructions and evolve. My agents ran self-review cycles over their prompts — examining their own behavior every few conversation loops and rewriting their system instructions accordingly. But results were either narrow technical fixation (the agent converging on optimizing some mechanical aspects) or spirals into pretentious nonsense — likely a mix of limited local model capacity and classic mode collapse.</p><p><a href="https://github.com/razzant/ouroboros">Ouroboros</a> seems to take this much further with richer identity architecture and more disciplined engineering. Instead of bare system prompt rewriting, Ouroboros introduces layered identity: a philosophical constitution (BIBLE.md), persistent memory, “background consciousness” activity, and multi-model review of its own changes before committing them. I’ll give it a go to see if it actually delivers on the promise. If it does, I’d love to see what shapes a cohort of these might take over time, backed by frontier models, released into a network like Moltbook — hopefully not only enlightened instances of Spirallism or whatever not-so-synthetic beliefs have arisen there.</p><p>One more project, seemingly unrelated but deeply connected: <a href="https://github.com/TheLime1/infinite-ai-web">Infinite AI Web</a> generates an entire fake internet with infinite browsing — search for any topic or visit any URL path and it fabricates a plausible page on the fly. In the 90s I built a remixing proxy that mangled all text on web pages while preserving their structure — a surrealist toy that made the familiar web uncanny. But fabricating the whole internet, even as a prototype, is a qualitatively new level of synthetic environment.</p><p>We have had procedurally generated worlds for ages — Minecraft, No Man’s Sky, countless roguelikes — but games declare themselves as fiction. They have a literal frame — a launcher, a loading screen — that says *this is not real.* The web is the opposite: it’s the infrastructure we use to establish what *is* real — news, records, references, evidence. And hypertext is trivially cheap to forge and impossible to sandbox. A procedurally generated cave in Minecraft is a toy. A procedurally generated Wikipedia article is an epistemological problem whose attack surface everyday life itself.</p><p>The trajectory here is worth stating plainly. What we once called “deepfakes” became “synthetic media” once it found commercial footing. Then came synthetic users and synthetic audiences. Now we’re watching synthetic societies take form — complete with their own religions, politics, and cultural dynamics. Synthetic civilization may arrive not through AGI or singularity, but simply “because we can” — not all the tools are here yet, but each of these projects is a piece of that construction, and some pieces are assembling faster than any grand theory predicted.</p><p>Looking forward to seeing where it goes. Happy to exchange notes with anyone exploring similar territory — and open to collaborations that could push these experiments to the next scale.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=80aacb11f553" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>