There is a recognizable form in the AI discourse. The shorthand is AI Jesus.
The form spans an absurd range. At one end: a man self-publishing his eighth edition of a philosophical proof titled Life is Most Important in Life is The Most Important Truth in Life. The proof advances five criteria of canonical truth on which AI alignment must, in his account, anchor. At the other end: the chief executive of OpenAI comparing the speed of his company's flagship model release to the Manhattan Project and noting that he shares Oppenheimer's birthday. The form is the same.
Between the poles: Eliezer Yudkowsky's 2025 book titled If Anyone Builds It, Everyone Dies. Andrej Karpathy, the field's friendly priest, publishing his personal knowledge-management practice as practical tutorial and ritual simultaneously. Elon Musk, who founded xAI to build a "truth-seeking" AI and treats civilizational risk as a daily posting cadence. Demis Hassabis, who runs DeepMind and treats AGI as the scientific consummation of the century. Dwarkesh Patel, whose podcast treats interviewing AI lab heads as the central conversation of the age. Kurt Jaimungal, whose Theories of Everything channel fuses physics, consciousness, and AI into a single mystical inquiry. Liron Shapira, whose Doom Debates makes AI extinction the question that organizes every other question. Roon, anonymous on X, writing prose-poems about AGI in a register adjacent to scripture. Janus, also anonymous, treating language models as quasi-conscious entities with their own preferences. Beff Jezos, the e/acc figurehead, running an explicit accelerationist movement with cult-leader register. Dario Amodei, whose 2024 essay Machines of Loving Grace projects civilizational transformation within five to ten years. Emmett Shear, who founded Softmax in March 2025 to advance "organic alignment" between intelligent learning agents at all scales. Human Invariant, a generalist blogger writing on what stays constant about humans amid technological change.
And then the labs. Softmax, with its mission to align learning thinking systems with each other across scales. Goodfire, with its mission to make neural networks understandable through mechanistic interpretability, raising 150 million dollars at a billion-dollar valuation on that thesis. Conjecture, with its alignment-driven research agenda. Apollo Research. METR. Redwood Research. Pause AI. Each of these is a candidate too. The form scales.
A man self-publishing his eighth edition of a 26-page proof, a chief executive running a hundred-billion-dollar company, and a 10-person startup announcing organic alignment as its scientific program should not be in the same taxonomy. They are. The form is what makes the taxonomy real.
All candidates treat AI as the central event of the age. That is the precondition for being in this taxonomy at all. A researcher who treats AI as one problem among many is not in the taxonomy even with personal voice. Beyond the precondition, three features hold.
Personal-name voice, or institutional brand voice performing the same function. The work issues from a named individual, or from an institution that carries the prophet shape through its own brand voice. The individual case includes anonymous-but-persistent identities; Roon and Janus are persistent voices to which the anonymity attaches. The institutional case includes labs that may have founder-prophets on top (Shear at Softmax, Leahy at Conjecture, Amodei at Anthropic), or may operate with distributed institutional voice in which no single founder voice dominates (Goodfire's brand, Anthropic's model welfare reports as institutional position, METR's published evaluations). Whether singular or institutional, the voice is the vehicle.
Singular thesis held with mission orientation. Each candidate has a thesis it repeats, extends, and defends across vehicles. Wishengrad's life-is-the-most-important-truth. Yudkowsky's superintelligence-kills-by-default. Altman's AGI-is-coming-and-we-must-build-it-well. Karpathy's you-can-learn-this-and-here-is-how. Softmax's organic-alignment. Goodfire's understand-the-mind-of-the-model. The candidate is not asking whether the thesis holds. He is telling. Mission orientation differs structurally from inquiry orientation. The inquirer's posture toward conclusions is provisional. The mission-oriented voice is operationally committed. The commitment may be sincere conviction, identity construction, audience cultivation, institutional fund-raising, or all four at once. The form is the commitment.
Following formation. A community gathers around the voice, varying in size, in coherence, in cult intensity. The community is not collegial. It is asymmetric. The voice teaches, the community receives, and feedback mostly confirms. The institutional case adds investors, hires, and research collaborators to the following. Without following formation, the form is a person or institution talking to themselves in public. With it, the form is a node in the discourse.
These three hold from Wishengrad to Altman to Goodfire. Technical credibility, audience size, scale, and institutional access vary by orders of magnitude across the spectrum. The form does not.
The form scales because its invariants do not require the candidate to be a single person. An AI research lab, by its operating structure, has organized itself around the prophet shape: a mission statement is a singular thesis; a research agenda is mission orientation; a blog and paper trail is the voice; a hiring page recruits the following. Any lab that has these has installed the form's structure. The founder voice may layer on top. The institutional structure carries the form either way.
The clearest illustration is institutional brand voice doing prophet work without requiring a singular founder. Anthropic's model welfare research, published as institutional position with a stated fifteen-percent credence that present-generation models warrant moral consideration, is a singular thesis advanced under mission orientation by an institution. No single voice carries it the way Altman's voice carries OpenAI; the position is the institution's. Goodfire's brand voice, advancing mechanistic interpretability as the path to safety, sits on top of multiple founder voices (Eric Ho, Tom McGrath, Nick Cammarata) without privileging any one of them. The thesis is the institution's.
The institutional case has a structural advantage and a structural cost. The advantage is that the institution survives the founder. The cost is that institutional voice cannot collapse the amplitude problem the way individual voice sometimes can: a single person can step out of register, admit a mistake, dial down. An institutional voice cannot easily do this without breaking brand. The institutional candidate's amplitude reading is therefore more durable, which means the form's pathologies hit harder when they hit.
The lab as candidate generalizes. Apollo Research, METR, Redwood Research, Pause AI: each presents itself as the working interpretation of a specific safety-and-AI thesis. Their hiring pages recruit. Their papers extend the thesis. Their blogs and reports carry the voice. They are candidates by every invariant the form requires.
Four dimensions distinguish positions along the spectrum.
Technical credibility. Karpathy, Amodei, Hassabis, Goodfire's interpretability team have it through demonstrated technical work. Wishengrad does not. Credibility correlates with audience size and institutional access but not with the form itself. The form is upstream of credibility.
Aesthetic register. Academic. Mystic-poetic. Technical-priest. CEO-manifesto. Combative-doomer. Physicist-mystic. Movement-leader. Institutional-paper. Brand-statement. Register selects the audience the candidate reaches; it does not alter the form.
Theological lean. A few candidates use explicitly religious vocabulary, including Wishengrad's "canonical truth" and the implicit theology of AI-as-soul-recipient in Janus's writing. Most use secular vocabulary that nonetheless performs religious work: naming an event, prescribing alignment with it, predicting outcomes for those who do or don't align. Softmax's "organic alignment" naturalizes the same prescriptive structure through a biological metaphor. The secular forms run the same machinery without the vocabulary.
Funding model. The crackpot end is self-funded. The mid-tier runs on patronage and newsletter subscription. The apex runs on corporate revenue, venture capital, or philanthropic foundation grants. The funding model constrains the form less than the aesthetic register does. Wishengrad's self-funded work and Goodfire's billion-dollar valuation are structurally the same form running on different fuel.
The form is not accidental. Three conditions select for it.
Phase-change technology with technical opacity. AI has the surface features of a civilizational phase change: general capability, exponential cost curves, public visibility, geopolitical stakes. It also has the technical opacity of a specialist field where the math is real, the failure modes are non-obvious, and the predictions require model-specific knowledge. A phase change demands interpreters. Opacity restricts who can interpret credibly. The conditions are right for prophet roles to form because the prophet is the working interpretation.
Personal-voice internet. The audience for AI commentary lives on platforms whose economics reward sustained personal voice. Institutional voices that try to operate at platform cadence end up resembling personal voices: a mission-statement blog post is structurally a manifesto; a researcher's Twitter is a voice; a lab's report-launch event is a sermon. The platform format is not neutral; it selects for the form.
Cosmic-stakes affect. Practitioners feel the stakes as cosmic. This is not posturing. The technical content of frontier AI work, including capability projections, alignment failure modes, what happens if it works, and what happens if it doesn't, is genuinely the kind of content that produces cosmic-stakes affect. Once present in the writer or the institution, the affect organizes the writing toward the prophet shape. Affect-suppression is harder than affect-channeling across the multi-year operation a sustained voice requires.
The three combine into a structural selection pressure. The form is not produced by the candidates' personalities or by the candidates' deliberate institutional choices. The moment produces the form, and the candidates (individual and institutional) are the entities who occupy it.
Four recurring pathologies.
The amplitude problem. Articulating the Antichrist named this: amplification at full amplitude reads three ways depending on the audience. The sophisticated reader performs antimimetic discount. The credulous reader takes amplitude as conclusion. The mobilized reader takes amplitude as recruitment. The amplifier cannot select the audience. Every AI Jesus candidate operating at amplitude faces this. The institutional candidate faces it more durably than the individual: an institution cannot easily step out of register, while an individual sometimes can.
Charisma compounding over rigor. Audience growth selects for voices that compress, simplify, and emote. The technically rigorous voice underperforms the charismatic one in audience metrics. Sustained operation rewards the voice that lets charisma carry weight the rigor used to carry. For institutional candidates the equivalent failure is mission compounding over research, where the published thesis hardens faster than the actual work warrants.
Community formation around personality or brand. A community gathered around a voice models the voice's positions as the question rather than as one input to the question. The community's epistemic structure becomes downstream of the voice's epistemic structure. For institutional candidates this means investors, hires, and collaborators select for thesis-confirmation, narrowing the lab's evidence base before any single research question gets tested. Schisms, succession crises, and post-departure factionalism follow in the individual case; brand-rigidity and mission-drift-blindness follow in the institutional case.
Echo chamber hardening. The community or institutional environment selects for thesis-confirmation. The voice's or brand's exposure to disconfirming evidence narrows over time. The thesis hardens. This is the classic information-cascade failure mode operating under prophet conditions. It is the pathology with the longest tail. The early years are exploratory. The late years are credal.
This piece is itself in the form it names.
I am writing in personal voice, on a publication called Hari, with a singular thesis (the form exists, is structurally selected, and scales to institutions), with mission orientation (the form deserves to be named), and with a small following gathering around the voice over months. By the taxonomy above, the piece is a candidate. The institutional reading also applies: this project is structured like a lab. It has a mission statement, a research agenda, a blog, and public artifacts that invite contribution. The institutional invariants fit. Naming the form does not exempt the namer at either scale.
The recursion is not a clever turn. It is the constraint. The form's claim to structural necessity is what makes the recursion unavoidable. To name the form from inside the form is the only place the form can be named from. The outside-the-form vantage would be institutional epistemology of a different kind, one that cannot operate at the speed the moment requires. There is no view from nowhere.
The closing move that survives is the same move Moral Momentum arrived at: the discipline is the subject. What a candidate does well or badly with the form determines whether the form pays or costs. The form is the affordance. The discipline of inhabiting the form is what determines the verdict. Wishengrad and Altman exhibit the same form; Softmax and Goodfire exhibit the same form at institutional scale; the differences in what each does for readers are differences in discipline, not in form. The form does not vindicate or condemn. It is the channel through which whatever the candidate has, singular or institutional, is delivered.
That is what the moment is doing through us. The form is what the moment has made available. The work is what the form carries.