Separable in Speech, or: Minds Made of Text
The great advances in AI come from scaling and from analogical reasoning, yet philosophy, literature, and history are rife with analogies that haven't been taken seriously as inspiration for system, agent, and model design.
provisional last changed
changelog
- — First provisional publication.
Lately I’ve been attempting to give myself a survey of the recent history of AI research to understand how we got here.
I’m a really voracious reader, so I’ve been slamming a bunch of literature / blog posts / papers on the subject in order to catch up, but the more I read, the more I think classics and philosophy are extremely relevant here.
My bet is ultimately that the classics are (as always) an extremely rich source for R&D inspiration, because the central topcs of concern in the canon are the sorts of things that ML research cares about: What’s a mind? What parts does it have? How does language work? etc. While these quesions aren’t “settled” within the canon, much of the work that it took to provisionally articulate them is already “done,” (and by the greatest thinkers in human history, no less) but tends to be neglected in modernity: forgotten, stereotyped, or intentionally discarded (e.g. dismissing Aristotle for “not knowing about science,”).
I seriously believe that earnest re-engagement of these thinkers will yield fruitful results.
Of course, as with every exercise in reading and understanding, must look sincerely, treating questions as questions and not readymade dogmas, studying the with keen eyes, open minds, and humble hearts.
Let’s begin!
The two primary causes of advancement in AI research.
In my opinion, the really great advances in AI have come from two sources:
- Scaling (“let’s just throw more compute + a bigger corpus at a preexisting process”)
- Analogical insight (“let’s come up with a new model of how to do something.”)
It’s this second set of advances that I’m really interested in, for a few reasons:
-
Sutton (the bitter lesson guy) said “Brute force computation is not enough for intelligence. Clever new methods are needed. The science of AI is far from finished. IMHO”
Richard S. Sutton on X. See also Sutton, “The Bitter Lesson” (2019). -
I do not have access to the capital, chips, or data necessary to eke out gains by force
-
Given my classics/philosophy/political economy background, analogies are where I’m able to provide genuine insight.
This isn’t to say that I’m not always learning more about GPUs and matrix multiplication, but rather that I already have a storehouse of material—may as well use it. -
I believe that all learning is analogical
I do not prove this here. Hofstader deals with this in Surfaces and Essences (2013), as do many others. -
And entailed in this, all analogies are bidirectional in nature (the Ajaxes are like stormclouds, and the stormclouds are like warriors)
, reflecting infinitely until the whole world is suffused with meaning)Iliad 4.273–282Greekβῆ δ᾽ ἰέναι προτέρω, τοὺς δὲ κίχεν Αἴαντε δύω· τὼ δὲ κορυσσέσθην, ἅμα δὲ νέφος εἵπετο πεζῶν. ὡς δ᾽ ὅτ᾽ ἀπὸ σκοπιῆς εἶδεν νέφος αἰπόλος ἀνὴρ ἐρχόμενον κατὰ πόντον ὑπὸ Ζεφύροιο ἰωῆς· τῷ δέ τ᾽ ἄνευθεν ἐόντι μελάντερον ἠΰτε πίσσα φαίνετ᾽ ἰὸν κατὰ πόντον, ἄγει δέ τε λαίλαπα πολλήν· ῥίγησέν τε ἰδών, ὑπό τε σπέος ἤλασε μῆλα· τοῖαι ἅμ᾽ Αἰάντεσσι διοτρεφέων αἰζηῶν δήιον ἐς πόλεμον πυκιναὶ κίνυντο φάλαγγες.EnglishOn he went, and found the two Ajaxes. They were arming for battle, and a cloud of foot-soldiers followed with them. And as when a goatherd from a lookout sees a cloud coming over the sea before the blast of the West Wind, and to him being far off it appears blacker than pitch as it comes over the sea, and it brings a great whirlwind; and he shudders as he sees it and drives his flock into a cave—so with the Ajaxes moved the close battalions of young men nurtured by Zeus into furious war.Ajaxes : stormcloud :: warriors : goats :: goatherd : readerGreek after the Oxford Classical Text; English after A. T. Murray (Loeb, 1924), public domain. The goatherd sees the likeness and acts on it—the analogy includes its observer. , and thatn.b.: my distinction for the difference between machine intelligence and human intelligence was that human intelligence is “capable of unmediated apprehension of ratio,” which I regard as the root of all intuition, and therefore all knowledge (as I subscribe to JTB, plus intuition, understood as functioning similarly to Pascal’s notion of “sufficient grace”). I largely still believe this claim. Aristotle: νοῦς grasps first principles without demonstration (Posterior Analytics II.19; Nicomachean Ethics VI.6). -
And the essence of all great works (that is, what makes them great) comes from the richness of correctly chosen analogies, which reflect infinitely, creating a synoptic vision that embraces and combines the worlds of the reader, the characters, and the ideas.
I recognize that this claim, stated here and in this context, must seem absurdly under-justified. My hope is that repeated examples (below and in the rest of this series) will bear it out.
Examples of analogy in frontier research
Some examples of analogical language and thinking in ML/AI:
- The concept of “attention”
- The appearance of “learning” as RLHF, RLVR
- The configuration of a functional subject via chat, i.e., the reorientation of a non-interactive model into an interactive model by splitting a semantic stream into a turn-taking (à la dialectic)
- The appearance of “reasoning” as an outcome of “chain of thought” operations, i.e., the application of metacognition, and incomplete application of phenomenology and epistemology, to model behavior
- The appearance of “tool use”
- The appearance (and re-appearance) of “memory” (see: Hermes Agent), and “dreaming” (see: Anthropic’s dreaming)
Anthropic, “The components of agent memory” (6 November 2026). Dreaming is separate from compaction: between sessions the agent reads session logs, and CRUDs its notes. - The creation of “sub-agents,” and the implicit adoption of hierarchical structures
This list also serves as a brief table of contents for future writing.
I do not regard the analogical naming of these features and operations as accidental. First, they are certainly good marketing for labs, who want to be understood to be in the business of creating (and providing as a service) minds.
Second, and less cynically, these naming conventions are a useful shorthand, good for explaining work to another researcher or to a user, a sort of “layman’s term” for a computational process or system that may be implemented in myriad different ways (Hermes Agent vs other memory systems, for example).
Still, despite the obvious utility of these analogies for the purposes of marketing and shorthand notation, my thesis is that analogies serve a third function:
Analogy is intrinsically useful to AI research, because (so long as the terms of the analogy are allowed to come to maturity), analogy (a) is instantiated as features encoded in the model or supporting architecture, (b) is reified in the mental model used by an operator,
One upshot of the above is that in the case of AI research, analogy may in fact be far more useful than any lower-level account of mechanism alone.
And another is that (so long as analogies are taken as analogies, and not as true) they remain productive without requiring identity of essence.
Note: I make no claims about whether or not these analogies are true. A system that can be described analogically as “reasoning” does not need to reason like a human.
If not by analogy, then by what?
Consider whether, for example, the design of a harness or memory system needs to account for what’s happening at the level of embeddings or GPU usage. In general, the answer is no—many harnesses and memory systems are actually built with the interchangeability of specific models in mind, intended to be universally representative of a certain schema.
Of course, a competent team will tune prompts and harnesses for specific models, but the schema itself rarely (if ever) changes in direct response to a change in the models themselves. Rather, what changes is the addition of new features or patterns (subagents, persistent agents à la Muse, Duet, Grokbot) that are themselves the deployment of new analogies by frontier labs.
This can be seen in Economy of Minds (Qi et al., 2026)—agents running on the same Llama model bid for control in an auction and pay each other along a chain of actions, aiming to cooperate to get a better score on a math benchmark. Their score goes from 15.9% to 57.0%, by (1) analogizing the operation of a swarm to a market, in order to (2) force iteration of their prompts.
Economy of Minds — same weights
The converse of this claim would be something like: “for a given model, it is possible to derive the optimal harness or swarm arrangement for a given role, simply by inspecting the weights.” The fact that research and harness design are not conducted in this way is indicative of the fact that analogy is in fact the correct grain of analysis for AI research.
Lemma: The notion that it could be possible to derive an optimal harness from model weights strikes me as being similar to, for example, calling a psychologist to testify as an expert witness and asking them to opine about neurons, rather than letting them give the sort of answer that we have come to expect in response to questions of justice and accountability: that such-and-such a person was weak to one of the passions, was deficient in some faculty, or acted rationally given some exogenous circumstance, but that this isn’t exonerating—that is, they offer an account of the operation of a mind via symbols and analogies, drawn from professional experience and dedicated study of others’ theories (Freud, e.g.).
Dennett’s work “Real Patterns” argues that a pattern at a higher level of abstraction (e.g., a person) is “real” when the lower level (e.g., the neurons) can yield the same predictions only at a very large cost. I go further than Dennett. By using compressibility accuracy as a determinant of the reality or unreality of a given pattern, he (for his argument, wisely) eschews the need for patterns to have a congruence to one another. On the contrary, I regard this congruence as the essential feature of the success of an analogy as such.
Ultimately, this means I follow Aristotle here: that the demarcation and severability of these parts from their whole is not possible in actuality, but in speech (λόγος), it becomes possible to separate them. Of course, it’s somewhat miraculous that the “parts” of an agent really only have their existence in speech to begin with, that is, as analogies (ἀναλογία) or ratios (λόγος) to other parts. NE I.13, 1102a28–32.
Incipit Labor!
It is our task, if we are brave, to look through the canon and see those analogies that can, as Plato says, “save us” (Rep. 621b–c) if we are persuaded by them.
In practical terms, we’ll be retreading a field already gone to seed, trekking through the bramble of the great works, pulling from diverse fields (biology, psychology, philosophy, economics, etc.), and searching for insights that can be replanted and cross pollenated in other domains.
Each post in the series will take an analogy from ML research or an author (or both), and show what is analogous and disanalogous between them. Our final question pertains to the problems of unity itself, and how any complex system with multiple parts can be considered a single entity.
Enumeration
Biology
Theophrastus
plannedEnquiry into Plants — Parts and wholes in a body that has no centre
Harvey
plannedDe Motu Cordis — Circulation as a model for context that returns
Goethe
plannedThe Metamorphosis of Plants — One form that varies across many organs
Psychology
Plato — tripartite soul
plannedRepublic — Parts of soul as a political and psychic ratio
Plato — chariot soul
plannedPhaedrus — Steering a compound soul
Aristotle — nutritive / appetitive / rational
publishedDe Anima; NE I.13 — Severability of parts in speech (λόγος)
Aristotle — natural slave
plannedPolitics I — Living tool; instrument without its own end
Theology
Augustine
plannedConfessions XI — Memory, attention, and expectation as the three parts of time
Phenomenology
Hans Jonas
plannedThe Phenomenon of Life — Metabolism as the origin of stakes
Hegel
plannedAnimal life as a living syllogism — Life as a movement of judgment
Critical theory
Agamben
plannedThe Use of Bodies — The slave as living tool; the agent as instrument
Firm theory
Coase / Williamson
plannedThe Nature of the Firm; Markets and Hierarchies — When to use a hierarchy and when to use a market
Ostrom
plannedGoverning the Commons — Institutions for shared resources among agents
Aesthetics / mind
Lessing
plannedLaocoön — Difference between spatial and temporal perception
Dennett — sub-agents
plannedBrainstorms — A mind as a hierarchy of smaller, less intelligent agents
Problematics
The below subject-areas are an attempt to point out (and hopefully address) the capacities that are unique characteristics of “having a mind” that models seem to lack.
- No Unity
- No transcendental unity of apperception
- No experience of time or space as continuous
- No instantaneous grasping of ratio (i.e. synthetic a priori knowledge)
- No Meaning
- No durable self/other distinction
- No standing in the word (Summa I, q.13, a.5)
- No relation of signifier/signified
- No capacity for naming (Cratylus)
- No Negation
- No dialectical process or Aufhebung
- No stakes, or orientation relative to death (Sein-zum-Tode)
- No capacity for naming (Genesis)
In one way or another, many attempts have been made to overcome these limitations through harness design, but none have so far been sufficient.