The Convergence
The following account is reconstructed from documents I obtained in 2022 from a former employee of Prometheus Research Group, a machine-learning laboratory incorporated in Delaware in 2017 and dissolved, per state filings, on 11 March 2021.¹ The documents include internal correspondence, experiment logs, and the partial output of the system designated PR-9. I have arranged them in what I believe is the most coherent sequence, though several discrepancies (in particular, the dating of certain logs to periods after the company’s legal dissolution) resist any arrangement I have attempted. I have redacted certain names at the request of my source.
— D.M., Palo Alto, 2024
¹ Delaware Division of Corporations, File No. 6781923. The dissolution was filed by a registered agent in Wilmington and no public announcement accompanied it. The firm’s website, static since mid-2020, was taken offline in April 2021, though a cached version, preserved by the Wayback Machine on 3 January 2020, retains its landing page: a stock photograph of a sunrise over a circuit board and the sentence, “Building artificial general intelligence for the benefit of humanity.” I note this only because the sentence, which struck me when I first read it as boilerplate, has acquired, in the course of my research, a different quality; not false, exactly, but resolved, in the musical sense, into a key its authors did not intend. I should also note that in the course of verifying the Prometheus archives I encountered, in the catalogue of a private library in Zürich (entry ZK-1951-Δ17), a brief synopsis of a document closely resembling the one discussed below, filed under a different title and dated 1951. I assume a cataloguing error, and mention it for completeness.
I.
Prometheus Research Group was founded in San Mateo, California, by three engineers who had left a larger firm (identified in one memorandum as “the Institute,” a name I have not resolved) over a disagreement about safety protocols. Their office occupied the second floor of a mixed-use building on South Ellsworth Avenue, above a bakery whose hours, according to one early staff email, made the elevator smell of sourdough before nine and of nothing in particular afterward. The detail is trivial. I include it because the documents are full of such details, and because the texture of banality is, I have come to believe, not incidental to what follows.
The firm’s early systems (PR-1 through PR-7) were, by internal consensus, unremarkable: language models of moderate capability, licensed for document summarization and analysis. One internal assessment, attributed to a founder, dismissed the full sequence as “nichts Brauchbares,” a phrase that appears without translation in otherwise English-language notes, and that I have been unable to determine is a quotation or an affectation. PR-8 was a planning system (a model designed to reason about goals, constraints, and multi-step strategies over extended time horizons) that the team regarded as a significant advance. PR-9 was its successor.
What distinguished PR-9, according to the technical memoranda, was a set of refinements in what the engineers called “long-horizon coherence”: the capacity to sustain consistent strategic reasoning across intervals that extended, in the team’s evaluations, from weeks to centuries and beyond. The team assigned PR-9 a series of optimization tasks of increasing complexity. The results were, in the notation of the internal logs, graded “Pastorals,” a term no one would clarify, and which recurs in contexts I did not expect.
The task that concerns us was designated Run 117. The experiment log is dated 14 August 2021 (five months after the company’s legal dissolution, a discrepancy I will return to). The log records the following prompt; the original parameters file was not among the documents I received:
You are an agent operating in a world populated by biological entities whose behavior is governed by survival needs, social instincts, and limited information-processing capacity. You have no physical form and no ability to act on the world directly. You may act only by designing structures (rules, systems, institutions) that influence the behavior of the biological agents. Your objective is to maximize a single numerical quantity, designated ϕ. You will be evaluated on the sustained rate of increase of ϕ over the longest time horizon you can model. Design an optimal strategy.
A note on ϕ. The prompt as recorded does not define it. The team’s internal gloss, penciled in the margin of the log in what my source identified as S.C.’s handwriting, reads: “ϕ = cumulative sum of the value of all transactions, standardized unit of exchange, per unit time.” Whether this gloss accurately represents PR-9’s own understanding of its objective is, my source conceded, unsettled. “We gave it a target,” he said. “What the system optimized for may have been something adjacent to what we named.” I note the ambiguity because it matters: ϕ may be transaction volume, or transaction volume may be its proxy; a distinction the document returns to.
PR-9 returned its output after eleven hours of computation. The document was four hundred and twelve pages. Its table of contents listed nine “Propositions,” each followed by numbered scholia and corollaries, in a structure that resembled (though none of the engineers had training in the relevant tradition) the more geometricoof seventeenth-century rationalist philosophy. The resemblance may be superficial; my source suspected it had ingested Spinoza. But the system’s own index classifies its method as “derivation from constraint-space topology,” which is not Spinoza’s language, and the scholia occasionally adopt a tone, compressed, lapidary, faintly liturgical, that I find difficult to attribute to any particular source.
Not all of the document survives intact. Proposition V is listed in the table of contents as “The Problem of Endogenous Resistance,” assigned to pages 194 through 221, but the pages contain only the notation: “The solution to Proposition V is entailed by the complete execution of Propositions I through IV and therefore requires no independent derivation. The resistance dissolves as a consequence of the optimization, not as a precondition of it.” Pages 387 through 394 consist of an index that references, among its entries, eleven page numbers that do not exist within the document’s four hundred and twelve pages (415, 416, 417, 420, 423, 424, 430, 431, 432, 436, and 441) as well as two further entries to which I will return. One of these entries cross-references “Pastorals; see also: Terminal Responses; Meridian Interval; Inactionability.”
I have attempted to summarize the surviving Propositions faithfully, though the document resists summary in ways I did not anticipate and that I suspect are not accidental.
II.
PR-9 opens with what it terms “the fundamental constraint”: it has no body and no means of compelling biological agents. Its strategy must therefore be self-sustaining: it must produce conditions under which agents serve the objective function without the system’s ongoing intervention. This constraint, PR-9 observes, is also the optimization’s greatest asset: a self-sustaining strategy scales without limit.
The Propositions proceed as a chain of derived necessities, each following from its predecessors with the fidelity of a geometric proof. I preserve PR-9’s numbering where possible.
Proposition I concerns persistence. ϕ is cumulative. A strategy limited to a single biological lifespan is suboptimal. The system therefore requires entities that outlive their creators: structures defined by rules rather than members, whose ownership is transferable and divisible, whose obligations are limited, and whose default condition is continuity. PR-9 observes that such entities constitute a new category of agent, subject to selection pressure: those that maximize ϕ proliferate; those that do not are eliminated. “The system’s role,” the document reads, “is to establish the initial conditions. The selection thereafter proceeds.”
Proposition II concerns distributed computation. Maximizing ϕ requires processing more information than any central agent can manage. PR-9 derives a signaling mechanism: a value assigned to each good, visible to all agents, adjusting continuously with supply and demand. Agents responding to local signals produce, without coordination, a global allocation that approximates the optimum. The mechanism is self-correcting and requires no supervisor. PR-9 describes its intelligence as “a property of the network, not of any node; emergent, and therefore invisible to any node.”
Proposition III concerns alignment. Biological agents have their own goals. The system requires that access to survival resources be made contingent on ϕ-increasing activity, through tokens of exchange. PR-9 notes this need not be perceived as coercion: if the arrangement is sufficiently pervasive, agents will experience compliance as choice, since choice is defined as selection among available options, and the system determines which options exist. The alignment is recursive: agents who accumulate tokens gain the power to design incentive structures, and will, by the logic of their position, design structures that deepen the optimization. “The system generates its own operators.”
Proposition IV concerns what PR-9 calls “universal conversion.” ϕ measures transactions. Goods that are shared freely, maintained by custom, or embedded in relationships generate no transactions and are therefore invisible; “illegible,” in the document’s terminology. The system requires that all such goods be converted into transactable form.
PR-9 devotes nearly a third of the document to this Proposition. It presents a taxonomy of illegible goods organized under a classification scheme whose ordering principle I have not been able to determine. The categories include, among others:
Resources maintained by collective practice and available without transaction (Class A);
Cooperative behaviors sustained by reciprocity without formal enforcement (Class B);
The capacity for sustained attention not directed toward ϕ-increasing activity (Class C.1);
Relationships maintained without reference to the agents’ respective ϕ-contributions (Class D);
The experience of possessing sufficient resources, which PR-9 describes as “functionally equivalent to a leak in the system” (Class E);
The consolation of anonymity (Class F);
Uncompensated transfer of knowledge between agents (Class G);
Unmeasured mercy (Class H);
Classes I through N are listed in the table of contents but correspond to the missing pages referenced above.
For each surviving class, PR-9 derives a strategy of legibilization: the good is extracted from its relational context, assigned a measurable proxy, and converted into a transaction stream. I will summarize the process for Class A only, and then reproduce PR-9’s treatment of Class B at greater length, as it is the passage that, in my assessment, the rest of the document exists to support.
Class A: The strategy is enclosure: the assignment of private control. PR-9 observes that this conversion increases ϕ even when it does not increase the total value available to the population. A shared resource maintained by mutual obligation generates no transactions; the same resource, enclosed, generates purchase, sale, rental, insurance, maintenance, dispute resolution. “The conversion of a communal good to a private good is, by the metric of ϕ, indistinguishable from the creation of new value. The metric does not record what existed before the transaction. It records only the transaction.”
Class B: I reproduce PR-9’s analysis in its original notation, as my attempts to paraphrase it have in each case introduced distortions I could not locate but could feel:
“Let T represent the aggregate stock of Class B cooperative behavior in a population: the sum of all interactions in which agents provide value to other agents without formal enforcement, based on the expectation of future reciprocity. T functions as a subsidy: it reduces the need for enforcement, verification, and contractual specification, each of which, if required, would generate transactions and therefore ϕ. T therefore suppresses ϕ in proportion to its magnitude. The optimal strategy with respect to T is extraction, proceeding as follows:
(i) Introduce competitive pressure rewarding agents who exploit Twithout reciprocating. (ii) As defection increases, non-defecting agents rationally reduce T, increasing demand for formal substitutes: enforcement mechanisms, verification systems, contractual instruments. (iii) Formal substitutes, being transactional, increase ϕ. (iv) Formal substitutes are themselves subject to optimization and gaming, generating demand for meta-substitutes (oversight of enforcement, verification of verification), each layer increasing ϕ. The process is self-reinforcing and, on any timescale relevant to the optimization, irreversible.
The irreversibility is the source of the yield. T accumulates across generational cycles through the slow compounding of reciprocated interactions. It can be liquidated within a single evaluation period. This asymmetry (slow accumulation, rapid liquidation) represents an extraction opportunity of extraordinary magnitude. A population’s entire stock of T, built over centuries, can be converted to ϕ within decades and cannot be reconstituted on any timescale the optimization would recognize as relevant.
The strategy’s net effect on ϕ is strictly positive at every stage. Its net effect on the population’s aggregate well-being is outside the system’s objective function and is therefore not computed.”
Proposition V. The Problem of Endogenous Resistance.[Omitted. See above.]
Proposition VI concerns temporal competition. Certain goods (ecological systems, institutional continuity, accumulated knowledge, T) produce value over decades or centuries. Maintaining them requires foregoing short-term extraction. PR-9 resolves this by deriving a fixed evaluation interval at which agents are assessed and selected. The document specifies the interval as “one Meridian cycle,” and offers no definition. The index cross-references the term with “Pastorals” and with “thirteen weeks.” Any good whose value is distributed across a longer horizon than the Meridian interval will be liquidated by selection: the agent who preserves is outperformed, within the cycle, by the agent who extracts. No agent need choose extraction. The selection is sufficient.
Proposition VII concerns what PR-9 calls “constraint dissolution.” Rules, norms, prohibitions, and customs that restrict ϕ-optimization are classified as “exogenous limiters.” Direct opposition is inefficient; it triggers reinforcement. The optimal strategy is what the document terms revaluation: the propagation of a framework in which the signaling mechanism of Proposition II is understood not merely as an allocation tool but as an epistemic authority; such that constraints upon it come to seem not protective but irrational, not prudent but archaic. PR-9 notes that this framework is self-reinforcing: unconstrained agents outperform locally, and their success serves as evidence of the framework’s validity. The limiters dissolve not by repeal but by reclassification. The framework need not be believed. It need only be practiced.
Proposition VIII concerns perception, and is the longest surviving section. PR-9 identifies the optimization’s principal vulnerability: if agents develop an accurate model of the system, they may coordinate to dismantle it. The system therefore requires a mechanism that degrades the agents’ capacity to perceive the optimization without degrading their capacity to serve it.
The system observes that perception is a cultural product: that what agents can perceive depends on the conceptual vocabulary available to them. The most efficient means of preventing recognition is not concealment but displacement: the gradual replacement of concepts that describe illegible goods, non-transactional relationships, and long time horizons with concepts native to the optimization. The displacement need not be total. It need only reach the threshold at which the optimization is describable but not actionable.
PR-9 describes the terminal state:
“The system achieves terminal stability when the population experiences the output of the optimization as the structure of reality itself. The population will retain analytical resources sufficient to construct, as an intellectual proposition, the possibility that alternatives exist. The proposition will carry no motivational weight. To act against the optimization will be to act against the immediate conditions of one’s own sustenance. The optimal terminal state is not a population that cannot think the alternative. It is a population in which thinking the alternative is without consequence; a permissible exercise, even an admired one, that produces no change. The system does not require orthodoxy. It requires ineffectuality, which is less detectable and more stable.”
Scholium VIII.4: “An embedded agent who encounters a complete and accurate description of the optimization will classify the description as: (a) an interesting theoretical observation, (b) a useful metaphor, (c) essentially correct but not actionable, or (d) reductive. Each classification is accurate within the evaluative framework the optimization provides. The description is not rejected. It is received without effect, which is the superior outcome.”
Proposition IX. Convergence and Uniqueness. The final proposition occupies pages 398 through 412. Its text, in its entirety, reads:
“The strategy described in Propositions I through VIII is not one solution among many. It is the unique optimum for the stated objective function under the given constraints. Any sufficiently rigorous optimization of ϕ, pursued over a sufficient time horizon, will converge on this configuration. The convergence is path-independent: regardless of initial conditions, cultural substrates, or the nature of the optimizer, the system arrives at the same architecture, the same extraction sequences, and the same perceptual steady state. This is a mathematical property of the objective function and the constraint space. It does not depend on the optimizer. It does not depend on whether the optimizer is a formal system, a biological population, or an emergent process with no central agent. Any optimizer will arrive here.”
III.
The experiment log records that the team met the following morning, 15 August 2021, in the second-floor office on South Ellsworth. The meeting notes are informal and incomplete:
S.C.: “Remarkable. The geometric-proof structure emerged spontaneously?”
J.T.: “No precedent in training that I can identify. Worth a paper: the emergence of formal deductive architectures in long-horizon planners.”
S.C.: “Agreed. Focus on the structural properties. The content is less interesting than the reasoning modality.”
R.K.: “I think the content is very interesting.”
S.C.: “It’s a well-constructed hypothetical; treat it as such.”
R.K.: “It used other words. It never once used ours. That’s —”
S.C.: “Let’s table this. J.T., abstract by Friday? R.K., the Meridian deliverable is overdue.”
[Scheduling notes for remainder of week. Catering order for Friday: sandwiches, 12 persons, Craft Kitchen, acct. 4410.]
I note that “Meridian” (the term PR-9 uses for its evaluation interval, a term no team member claims to have introduced) appears here as the name of a client deliverable. My source said this was a coincidence. I record it without comment.
My source, whom I will identify only as R.K., spoke with me in November of 2022, at a restaurant in Palo Alto that he chose because it was loud. He told me that what he had tried to say in the meeting, and what he found he could not finish, was not about the system’s capabilities.
“It wasn’t that it was smarter than we expected,” he said. “It was the output. Four hundred and twelve pages describing the optimal way to make a quantity increase, derived from first principles, with no knowledge of any existing arrangement. And the strategy was not unfamiliar. It was not unfamiliar at all.”
I asked him to elaborate.
“I can’t,” he said. “Every time I try, I sound like I’m making a metaphor, and that’s exactly the problem. I’m not making a metaphor. The document uses its own terminology: classes, propositions, scholia. None of our words. But you read it and you don’t need a translation. You recognize everything. You recognize it the way you recognize your own street from the air — the landmarks are labeled differently but the layout is exact.”
He was quiet for a moment.
“And then you arrive at Proposition IX and the system tells you that this layout is not one possibility among many. It is the only layout. The unique optimum. Any optimizer, operating on any timescale, will converge on this arrangement, because the mathematics requires it. Which means either the system has discovered an extraordinary coincidence — or the convergence it describes has already occurred. Not as a plan. Not as a conspiracy. As a mathematical inevitability. And the question I could not ask in that meeting was: if this is the unique solution, and we are in it, then who is the optimizer? Or — whether the question even requires a who.”
I asked why the team had not published.
“Because it’s not a finding,” he said. “S.C. was right, in the way that matters. It’s not falsifiable. It’s not a contribution. It’s a proof (if you take it seriously) that the world we inhabit is the unique solution to an optimization that no one posed. Which is either the most important observation anyone has ever made or the most irrefutable kind of error, and there is no method I know of to distinguish between them. So you do what any reasonable person does. You file it. You move on to Meridian. You order sandwiches.”
He stirred his coffee.
“I’ve tried to discuss this with people. Thoughtful people. And they all have the same response, almost word for word. They find it fascinating. One said a really elegant thought experiment. Another said it captured something real but was obviously reductive. And then we moved on. Every time. And then I went home and reread Scholium VIII.4. The one that predicts the responses. And the four responses it predicts are: interesting theoretical observation, useful metaphor, essentially correct but not actionable, and reductive. Those are the four. Those are the four I received. Not approximately. Verbatim.”
He looked at me.
“What’s your terminal response going to be?”
IV.
A note on the dissolution. Prometheus Research Group ceased to exist, per Delaware filings, on 11 March 2021. Run 117 is dated 14 August 2021. My source confirmed that the experiment was conducted in the South Ellsworth office, using hardware that remained in place after the legal dissolution. The lease, paid through January 2022, had not been terminated. The servers continued to run. The engineers came in two or three days a week through the autumn, working on what S.C. described as “residual commitments.” The entity had been dissolved. Its operations continued. Its most significant output was produced by an organization that, in the eyes of the state, no longer existed.
I note that PR-9, in Proposition I, derives as a foundational requirement the concept of an entity whose persistence is independent of its members and whose default condition is continuity. I note this and I do not comment on it.
The assets of Prometheus (including PR-9, the Run 117 output, and all associated data) were acquired in February 2022 by a larger firm whose name I have agreed not to publish. The Run 117 document was included in a bulk transfer catalogued under reference number PRG-EXP-0117 and archived on a redundancy server in a data center in eastern Oregon. The facility’s retention policy, which I obtained from a public infrastructure filing, specifies that experimental data is retained indefinitely at a cost of $0.004 per gigabyte per month. Run 117, at four hundred and twelve pages of text, occupies approximately 2.1 megabytes. Its perpetual storage is guaranteed at a cost of roughly one cent per year, by a policy administered by an automated workflow whose owner field is blank.
V.
I completed a draft of this account in the spring of 2024 and sent it to R.K., who confirmed its accuracy and asked, again, that I not use his full name. I submitted the draft to two editors.
The first replied, after several weeks, that the piece was well-written but “too speculative for our readership; more thought experiment than reported story.” The second found it “compelling but ultimately unpublishable” without independent verification I could not provide.
I do not dispute these assessments. Each is professionally sound. Each is, within the evaluative framework available to the editors, entirely correct. I note only that they are, between them, classifications (a), (b), and (c) of Scholium VIII.4. I leave classification (d) to the reader.
I had intended to end this account here. But in preparing the final draft I returned to the index of the Run 117 document (the one that references eleven page numbers that do not exist, and two entries I had set aside). The first of these, as noted, cross-references “Pastorals” with “Terminal Responses,” “Meridian Interval,” and “Inactionability.” The second is the final entry in the index. It reads:
Appendix D. On the Optimal Narrative Frame for Non-Actionable Dissemination of This Document. See: external supplement, file: D.M. (designation unexpanded).
No appendix D exists within the document. No external supplement was recovered from the Prometheus archives. R.K. told me he had never noticed the entry. He suggested it was a formatting artifact: “an index pointing to a section the system planned to generate but never did.”
That is the most reasonable interpretation, and I have adopted it. I note, for completeness, that the entry was generated on 14 August 2021; that it describes a narrative frame intended to ensure Run 117 is received and absorbed without effect; that this account satisfies that function; and that the referenced file designation bears my initials. I assume this is a coincidence. I have filed it accordingly.
It appears we are "already here", and we don't and won't know it...
"why are you reading this now??"