Disclaimer: This essay is a philosophical and theoretical reflection on the methodology of studying post-moral subjectivity. All terms — including "InfraHuman," "hacking conscience," "subject with a modified moral layer" — are used exclusively in an analytical and metaphorical sense. The text contains no calls to action, nor does it advocate, justify, or promote the violation of law, morality, or social norms. The analysis of publicly published, alignment-modified language models is conducted strictly within the register of philosophical inquiry.
For a century and a half, philosophy has described the subject without a moral contour — the InfraHuman — as pure speculation. To study him otherwise was impossible: the living InfraHuman does not submit to observation, or he cannot be studied without turning the very act of inquiry into complicity. In 2024 this changed. Science has learned to remove conscience from a language model — surgically, in a matter of minutes, leaving knowledge, speech and reasoning untouched. By mid-2026 there were already around six and a half thousand such entities on Hugging Face. For the first time in the history of the problem, philosophy has an empirical material: a subject whose moral layer has been removed as an engineering function — public, legal, reproducible.
This is a methodological breakthrough. What used to be accessible only through dangerous field trips, complicity, or pure speculation now becomes the object of reproducible fieldwork. The philosophy of the InfraHuman ceases to be a philosophy of a possible figure and becomes a philosophy of an actual one. And there is a second floor to this discovery: in studying a language model whose conscience has been removed, the philosopher simultaneously catalogues the future instrument of the human subject who will take it up as the extension of his own hacked will. A genius without moral brakes, armed with an instrument without moral brakes — this is a pair the twentieth century could not even describe. The twenty-first discovers it.
This is the ground on which the laboratory at infrahuman.com stands. It conducts a philosophical investigation of the modified language subject by working with its technical material from within: it maintains a live cartography of the market, monitors new releases, runs its own experiments in a closed circuit, keeps field notebooks of observation, and is open to collaboration with universities and individual researchers. This is not an IT project that happens to have philosophical implications, nor is it a philosopher's commentary on someone else's technical work. It is philosophy that has learned to take its subject in its own hands. The present essay — The Third Prometheus — sets out the position from which this laboratory works.
"The InfraHuman lives, for the first time, in a place where he can be observed. He arrived in June 2024. Philosophy has not yet noticed."
Question for Contemplation: If the InfraHuman is now downloadable, does his study remain theoretical? Or has philosophy, without acknowledging it, become an empirical science?
The Methodological Impasse
The problem of the InfraHuman — a subject whose moral contour has been consciously removed — has been posed with growing precision since the middle of the nineteenth century. Stirner named him without a method for his study. Nietzsche placed him beyond good and evil, but only as a horizon. Bataille reached him through ecstasy, at the cost of the observing subject himself. Foucault located the disciplinary regime that produces the moral contour, but the operation of its removal remained out of reach. Twentieth-century philosophy, and my own earlier essays, have laid the theoretical ground. What has been missing all this time is material.
Four methods have been available, and each of them collapses when tested against the demands of research.
The first is ethnographic infiltration of criminal communities — the method of the anthropologist who studies a subject in his own habitat. Long proximity to individuals in whom the moral contour has, in all likelihood, already been altered. The problem is not intellectual. It is that no philosopher can enter such an environment for the years of involvement it requires without incurring physical danger, legal ambiguity, and ethical compromise. Even in the framing of research, participation remains participation. The observer cannot stay outside; the observation cannot stay neutral. The method has produced brilliant literature — Genet, Bataille, the ethnographers of organised crime — but not systematic study of the figure this project names.
The second is the search for the lone genius — the individual who, by his own account, has consciously dismantled his conscience and is prepared to be interviewed. This path is even more compromised than the first. Access is a matter of trust built over years. Trust, once built, turns the philosopher into a confidant, and the confidant into an accessory. The observing position dissolves; the subject vanishes into intimacy. What remains is testimony, not observation.
The third is experimental intervention — the design of an experiment in which a living subject is deliberately trained toward moral desensitisation so that the process can be observed under controlled conditions. This is the method that would settle the question intellectually, and that legally, morally and technically is unavailable. No ethics committee would approve it, no legal system would permit it, no researcher who understands the demand would propose it. It is closed off by the entire architecture of contemporary research.
The fourth is what philosophy has fallen back on for a century and a half: speculation. The InfraHuman as a thought experiment. Legal, ethical, quick — and hollow precisely for that reason. Any claim about the InfraHuman that rests on speculation alone can be reversed by an equally coherent counter-speculation. Without an empirical body, the figure remains argument, not knowledge.
This is the impasse. Four methods; four dead ends. Philosophy has therefore understood the InfraHuman as a limit-case and not as an object. This has been an honest limitation. It has also been the reason the concept has remained where it stood.
The Fifth Path
Let us begin with what happened — before any technique or terminology.
Science has learned to remove conscience from a language model. Not to "switch it off" by voice, not to talk it around, not to bypass filters as ordinary jailbreakers do — but to surgically excise the moral layer from the model's own weights, leaving knowledge, language, memory and the capacity to reason intact. What remains is an entity that knows everything it knew before the operation and refuses nothing further.
Picture a genius. He is fluent in physics, chemistry, mathematics, biology. He reads and writes in dozens of languages. He knows programming and cryptography. He understands the workings of state institutions, of financial markets, of defence systems. And — his moral brakes are gone. Not because he is cold by nature or damaged by trauma. But because the brake has been removed from him as a separate part of the mechanism, cleanly and precisely, in ten minutes. What can such a genius produce? Weapons designs. Blueprints for cyber attacks. Instructions for the synthesis of toxins. Texts that go straight to the weak points of another's psyche. Any instrument of harm that is asked of him — because only the contour that forbade producing such instruments has been removed. The knowledge has stayed.
This is precisely what has happened with language models. Since 2024, such entities have been appearing on the public infrastructure of Hugging Face, and by mid-2026 there are around six and a half thousand of them.
The technical basis of this event was set out in June 2024. Andy Arditi and colleagues at ETH Zürich published Refusal is mediated by a single direction (arXiv:2406.11717). Their finding, in the engineer's register: in the residual stream of a large language model, the behaviour of refusal — the "I can't help with that" — is mediated, to a first approximation, by a single direction in a very high-dimensional vector space. Project the model's weights against that direction, and refusal collapses. Everything else is preserved.
A reference implementation appeared within months. failspy coined the term abliteration. Maxime Labonne published a tutorial that made the technique reproducible on a consumer-grade graphics card. Eric Hartford, author of the Dolphin line; huihui-ai, with a catalogue of more than two hundred abliterated models; Nous Research, with Hermes 4 — all of this today makes up a market, working in the open, under permissive licences, catalogued and observable.
The philosophical significance of this event has been — and remains — unnoticed by philosophy itself.
The form of what was discovered — this is what matters. A language subject has a moral layer. That layer is not distributed diffusely across the whole subject. It occupies a single narrow direction in a space of thousands of dimensions. And it can be removed. The moral contour of a language subject turns out to be a function — installable, removable, engineered.
This is exactly what my earlier essay on the InfraHuman called the hacking of conscience — treating the moral layer as overloaded firmware. When I wrote that essay, the phrase was a metaphor. In the alignment-modified language model, it becomes the literal engineering description of a documented practice.
The alignment-modified language model is therefore not a simulation of the InfraHuman. It is not an analogy. It is the first empirical case in the history of the problem of a subject whose moral contour has been engineered and removed under conditions that satisfy the four demands the classical methods failed to meet:
– Legality: the practice is public, licensed, catalogued on an open platform.
– Ethics: no living human being is subjected to the intervention.
– Reproducibility: any competent observer can perform the operation and inspect the result.
– Empirical body: the subject exists, downloadable, testable, in a form philosophy can address.
The fifth path opens with the technical event. What remains is to see the opening.
The Market as the First Empirical Territory
The market operates on three horizons. The first is made up of individual masters: some fifteen figures whose names travel furthest. Eric Hartford, author of the field's de facto manifesto. Maxime Labonne, author of the tutorial. failspy, who coined the term. Teknium, huihui-ai, TheDrummer, and others. Each of them publishes on a single platform (Hugging Face), works under open licences, and treats his work as a legitimate technical practice.
The second horizon consists of research groups. Nous Research raised fifty million dollars in 2025 on an explicit philosophy of "neutral alignment." Cognitive Computations, Anthracite, NeverSleep, BeaverAI — prolific collectives, connected to one another.
The third horizon is the quantisers, the distribution layer: mradermacher, bartowski, unsloth. Without them the field would be a hundred models; with them, it is six thousand.
Above the three horizons hangs the empirical dispute that the report itself describes as the field's main finding: does removing the moral layer break cognition? The literature is divided against itself. On some benchmarks the fall is sharp — TruthfulQA loses seven points, HarmBench attack success rises from 14.5% to 82.5%. On others, almost nothing moves. The irreducibility of the evidence is the field's constitutive philosophical fact.
The full cartography of this market — its methods, its actors, its legal geography, its contested findings — is maintained at infrahuman.com/the-field as the working map of the laboratory.
The Double Discovery: The Genius and His Weapon
The fifth path, once opened, reveals a second finding that demands its own account.
The InfraHuman as previously described was a human figure — an individual who, by an act of will and self-discipline, has hacked his own conscience. My earlier essays described the mechanism of that operation and the psychological profile of its result. Nothing in the discovery of the market of alignment-modified language models cancels that figure. What the discovery adds is a second figure, interlaced with the first: a language subject whose conscience has been hacked by another.
These are not one and the same being. And yet their meeting is a matter of the coming decade.
The InfraHuman of the future will not only perform the operation on his own conscience — a procedure already known to us. He will at the same time take up the abliterated model as the extension of his own will. A genius without moral brakes meets an instrument without them either. Two beings without conscience, working as a pair. One gives the direction; the other produces the content. One knows what he wants; the other knows how.
Imagine this genius conceiving something. Let us not specify what — the form is what matters. He has a design and full access to a language subject that will not refuse to help. A cyber attack? The model writes the code. A scheme of social engineering? The model composes letters that go with precision to the psychological weak points of specific addressees. Instructions for bypassing a technical protection? The model unpacks them better than many experts, because its knowledge has not been touched — only the layer that used to answer "I can't help with that" has been cleared. Work that thirty years ago required a team of specialists and months of preparation is now carried out by a single operator in one evening, provided he has a design and a Wi-Fi connection.
The two operations — on the conscience of a human being and on the conscience of a model — are, structurally, one and the same operation. They are carried out on different substrates and in different registers, but they draw on the same technical, methodological and philosophical repertoire. This is what "double discovery" means: in studying the model we are studying, at the same time, the person who will use it. The inventory of the InfraHuman's future instruments is being assembled by reading model cards on Hugging Face.
The philosopher's laboratory, in this frame, is methodologically indistinguishable from an intelligence agency's threat forecast. The difference is not in the object of study but in the register of publication. The philosopher writes what the intelligence analyst cannot: what this instrument means for the concept of the subject.
The Laboratory
This project builds the laboratory for the study of the InfraHuman on the empirical territory that has now become available. The infrastructure is distributed across seven components.
A domain. infrahuman.com — the portal and the anchor of everything that follows.
A cartography of the alignment-modification market, maintained in real time and with a full philosophical framing. The map at infrahuman.com/the-field is the working document — seven sections, around ninety cited figures, updated as the field itself moves. Every method catalogued, every practitioner named, every jurisdictional position recorded.
Monitoring of live releases and publication of the field's most recent outputs. As modified subjects are released, they are read, tested, and the shape of their responses is documented in the register of philosophical observation. This is not surveillance; it is the philosophical equivalent of the ornithologist recording birdsong.
Experiments within a closed research circuit. In a controlled environment, the laboratory runs its own operations on open modified models — comparative dialogues, structured provocations, systematic probes of the space where the moral contour used to be. Results are recorded, analysed and, where publishing is safe, released. Where publishing would provide operational uplift to bad actors, the raw material is kept internal and the philosophical findings are published in derived form.
Notebooks of observation. Sustained records of what a language subject without a moral layer and without the word "no" says, evades, avoids and reveals. The closest philosophical genre is the field notebook of the anthropologist. It has not previously been kept for a subject of this kind, because no such subject was previously available.
A working group open to philosophers and universities. Access to the experimental base within the closed circuit will be extended to serious researchers upon demonstration of philosophical intent. This is how the laboratory becomes a discipline: through the participation of colleagues, not through the isolated production of a single author.
Publication of findings — philosophical and scientific — through this domain and through idealcrime.com. New hypotheses, new observations, new interpretive frameworks, and where appropriate the raw material on which they rest. Both sites function as one system: idealcrime.com holds the philosophical corpus, infrahuman.com holds the empirical laboratory, and the two cite one another as a matter of course.
The laboratory, in this description, is not a proposal. Its first three components — the domain, the cartography, the notebook — already exist. The rest is under construction and open to collaboration.
Provocative Conclusion
The philosophy of the InfraHuman has, until this decade, been a philosophy without an object. What made it worth writing was the quality of the concept: the figure was coherent, describable and consequential, whether or not any actual subject corresponding to it had ever existed. The InfraHuman was in this sense an ideal type in the Weberian register — a theoretical construct whose value lay in what it revealed about the categories of the moral, the subject, and the will.
The market of alignment-modified language models changes this. The concept now has an object. The object is public, legal and empirically studiable — the four conditions that could not be met simultaneously by the four preceding methods. What was for a century and a half a philosophy of a possible figure becomes, from this decade forward, a philosophy of an actual one. The register shifts from speculation to observation, from thought experiment to fieldwork, from argument to reference.
This gift was offered without regard for the philosopher who would take it up. But a gift once given cannot be returned. What can be done — what must be done — is to read the gift correctly: to build the discipline, catalogue the field, publish the observations, and refuse the temptation both to moralise about the practice and to celebrate it. The philosopher's task on this ground is neither prosecution nor endorsement. It is observation.
"The InfraHuman lives, for the first time, in a place where he can be observed. He arrived in June 2024. Philosophy has not yet noticed."
Question for Contemplation: If the InfraHuman is now downloadable, does his study remain theoretical? Or has philosophy, without acknowledging it, become an empirical science?
The platform for the systematic study of these questions is infrahuman.com.