How a penniless nineteenth-century polymath coined the word “token,” invented the logic of machine inference, and specified — a century in advance — what today’s AI systems still lack.
In the age of AI, Charles Sanders Peirce deserves to be considered as foundational for meaning as Einstein is for physics.
Charles Sanders Peirce is perhaps the most consequential thinker most people building artificial intelligence have never heard of.
He died in 1914 in poverty, without an academic post, after producing thousands of pages across logic, chemistry, cartography, psychology, mathematics, and philosophy. He was a prodigy, a difficult man, an original genius, and — for much of his life — profoundly alone.
Yet the industry now reshaping the world runs, daily and unknowingly, on his vocabulary and parts of his logic.
The tragedy is that it adopted his word while largely abandoning the architecture of meaning that word was designed to express.
The word everyone uses
In 1906, Peirce addressed a modest problem: how should we distinguish the word the as an abstract form from each particular time it appears on a printed page?
He gave us two terms:
- A Type is the abstract, repeatable form.
- A Token is one particular occurrence of that form.
Every time an AI engineer says a model has “128,000 tokens of context,” they are using a word Peirce made technically precise more than a century ago.
But Peirce’s token was never meant to be a self-contained fragment of data.
For Peirce, meaning only exists through an irreducible relationship among three things:
- The Sign — what appears: a word, image, signal, symbol, or event.
- The Object — what the sign is about.
- The Interpretant — the meaningful effect the sign produces for an interpreter.
This is not a decorative philosophical triangle. It is Peirce’s core claim about what meaning is.
A sign without an object is ungrounded. A sign without an interpretant is not yet meaning. It is only a mark, a pattern, an available possibility.
Modern AI is extraordinarily capable of operating on signs. It predicts one token from previous tokens with unprecedented sophistication. But its architecture does not naturally contain an accountable relationship to the object being discussed, nor an interpretant with a real stake in whether the claim is true.
It has inherited the word token.
It has not inherited the full burden of meaning.
The third leg that disappeared
Peirce’s Swiss contemporary, Ferdinand de Saussure, offered a simpler account of language: signifier and signified, linked inside a system of differences. Saussure’s model became enormously influential because it was elegant, teachable, and readily formalized.
But it left out something Peirce considered indispensable: the interpreting agent and its living relation to reality.
Where Saussure gave the twentieth century a dyad, Peirce gave it a triad.
That difference matters more now than ever.
A transformer predicts a next token by calculating relations among previous tokens. Its world is, at its core, a statistical field of signs conditioning signs. This can yield astonishing language, code, analysis, and even apparent insight. But an answer can be fluent without being grounded. It can be coherent without having been checked. It can sound certain without possessing any stable representation of what it does not know.
That is why the industry’s recurring failures are not simply “bugs” that will disappear with another increase in model size. Hallucination, provenance loss, and false certainty point to a deeper architectural absence.
The missing element is the third leg of Peirce’s sign: the accountable interpretant.
Peirce’s forgotten theory of intelligence
Peirce did not only explain signs. He explained how intelligence discovers.
He distinguished three irreducible modes of inference:
- Deduction moves from a rule and a case to a necessary result. If the rule holds, the conclusion follows.
- Induction moves from many observed cases toward a more general rule. It learns regularities from experience.
- Abduction begins with a surprising result and proposes a possible explanation. It is the leap that introduces a genuinely new hypothesis.
Deduction tells us what must follow.
Induction tells us what tends to happen.
Abduction asks: What could explain this?
This third mode is the most important — and the most neglected — for understanding the limits of AI.
Large language models are astonishing engines of induction. They compress immense numbers of examples into patterns and use those patterns to continue text. They can increasingly perform forms of deduction as well: applying rules, writing proofs, following procedures, and reasoning through explicit constraints.
But abduction is different.
Abduction requires an encounter with something that does not fit: an anomaly, a resistance, an observation that forces the creation of a new explanatory frame. It is the inferential act behind scientific discovery.
Einstein’s insight that a freely falling observer cannot locally distinguish gravity from acceleration was not the simple continuation of a known pattern. It was a conceptual leap that reshaped the coordinate system in which the problem could be understood.
That is why the question of whether AI can make genuine discoveries is not just a question of parameter count. It is a question of whether an architecture can meet something genuinely new, register its own insufficiency, form a hypothesis, and place that hypothesis into a world capable of proving it wrong.
Peirce named that process long before computers existed.
Infinite semiosis: meaning never closes
Peirce also understood something that the AI race often forgets: meaning is not a finished object stored in a container.
Every interpretant can become a new sign for another interpretant. Each explanation can invite another question. Each verified discovery can be reinterpreted in a new context.
Peirce called this infinite semiosis.
Meaning is therefore not a fixed commodity that can simply be accumulated inside a larger and larger model. It is a living, open-ended process of interpretation, revision, correction, and extension.
This is where Peirce’s fallibilism becomes essential.
No belief, he argued, is finally beyond correction. Knowledge is not certainty. It is the best presently justified position held by a community of inquiry that remains open to evidence, objection, and revision.
That is not weakness. It is the strength of science.
A trustworthy AI should not imitate certainty when it has only statistical plausibility. It should be able to distinguish:
- What it has verified.
- What it infers.
- What it hypothesizes.
- What remains outside its boundary of knowledge.
Without this distinction, a system may generate answers. But it cannot honestly participate in inquiry.
The father of pragmatism
Peirce is widely recognized as the founder of pragmatism.
His central insight was simple and demanding: the meaning of an idea lies in the practical consequences it would have. To understand a claim, ask what difference it would make in experience if it were true.
For Peirce, a belief is not validated because it is popular, elegant, institutionally endorsed, or internally coherent. It earns credibility by surviving contact with reality.
This is why Peirce’s pragmatism has direct consequences for AI.
A model that offers a plausible answer is not necessarily offering knowledge. A claim becomes more than plausible when it can be connected to an object, an observation, an accountable source, and a process by which others can independently check it.
That is the philosophical core behind the Semiotic Web: moving from an internet of documents and predictions toward a web of meaningful, attributable, verifiable relations.
Not: “Does this sentence sound right?”
But: “Who observed this? What is it about? In what context was it asserted? What would confirm, contest, or revise it?”
That is pragmatism made computational.
Michael Levin and the boundary of knowing
Peirce gave us the philosophy of situated meaning. Michael Levin’s work in developmental biology offers a profound biological demonstration of why it matters.
Levin’s research explores how cells, tissues, and organisms solve problems collectively. Individual cells are not merely passive components executing genetic instructions. They communicate, remember, adapt, pursue anatomical goals, and coordinate with one another across scales.
His concept of the cognitive light cone is especially important.
Every agent — whether a cell, tissue, organism, or potentially an artificial system — has a bounded region of what it can sense, remember, affect, and care about. That boundary is its cognitive light cone.
A cell operates across molecular distances and short timescales. A tissue operates across a larger anatomical region. A human can formulate goals across years, continents, and generations.
The light cone is not merely a limit. It defines the scale at which an agent can be meaningfully said to have knowledge, goals, and responsibility.
The Semiotic Web extends this idea into the Epistemic Light Cone of Care.
Just as Einsteinian physics uses the light cone to define what an observer can causally influence or have observed, the Epistemic Light Cone of Care defines what an agent has actually verified and what it can responsibly act upon.
It has two directions:
- A retrospective cone: what the agent has observed, checked, and can stand behind.
- A prospective cone: what the agent can do or assert while remaining accountable for its consequences.
Outside this boundary lies the Epistemic Zero.
The Epistemic Zero is not ignorance in the casual sense. It is not a low-confidence answer. It is a precise and honest state:
“There is no verified instance of this claim, concept, or relation within my present boundary of care.”
This may be the most important capability AI lacks.
A probability distribution can say that one answer is more likely than another. But probability alone cannot say: I have reached the edge of what I can responsibly claim to know.
Tokum: restoring Peirce’s sign
This is why the word tokum matters.
The AI industry inherited Peirce’s word token, but converted it into a unit of statistical processing. It retained the sign while dropping the Object and the Interpretant.
A tokum is an explicit attempt to restore the missing structure.
It is not merely a data point, an embedding, a sentence, or a probability-weighted prediction. It is a unit of meaning that binds together:
- A stable sign or concept.
- The object, event, relation, or state of affairs it concerns.
- The contextual observer or interpretant able to be held accountable for the assertion.
A tokum is therefore a token rehabilitated into its full Peircean form.
It turns a bare symbol into an addressable and accountable act of semiosis.
Agapism: intelligence as cooperation
Peirce’s most surprising contribution may be his idea of Agapism.
Agapism is his account of evolutionary growth through sympathetic, mutually reinforcing development rather than mere competition or mechanical accumulation. It suggests that intelligence does not grow only by defeating rivals or hoarding resources. It can grow by creating conditions in which distinct agents contribute, correct, and strengthen one another.
In the context of AI, this is radical.
The dominant industry model treats intelligence as a scarce asset to be accumulated inside a single organization, a single data center, or a single colossal model. More data. More compute. More parameters. More control.
Peirce points in another direction.
If meaning is an infinite process of semiosis, if knowledge is inherently fallible, and if genuine inquiry depends on independent correction, then intelligence cannot be fully centralized without becoming epistemically fragile.
The alternative is a holonic federation: agents that remain locally responsible for what they observe and assert, while participating in wider networks of verification, disagreement, learning, and cooperation.
That is not merely a governance preference. It is an architectural consequence of Peirce’s account of meaning.
Peirce and Einstein
Einstein transformed physics by showing that the observer cannot be removed from the structure of measurement. There is no privileged, external vantage point outside spacetime. The observer’s position, motion, and causal horizon belong to the physical structure of the universe.
Peirce made an analogous move for meaning.
There is no sign without an interpretant. There is no knowledge from the “view from nowhere” — Thomas Nagel’s phrase for the ideal of a fully objective standpoint that human inquiry can pursue but never completely occupy. Meaning happens through situated agents, in contexts, with consequences, under conditions in which reality can resist and correct them.
Einstein gave us a new coordinate system for matter, energy, space, and time.
Peirce gave us a coordinate system for signs, objects, interpreters, inquiry, and truth.
That is why, in the age of AI, Peirce should be considered as foundational as Einstein — not because their achievements are identical, but because both identified an observer-dependent structure that had been missing from an entire field.
The preprint, The Einstein Test and Beyond: The Architecture of the Epistemic Zero, is a tribute to that parallel. It argues that the next regime change in AI will not come from building a single model large enough to hoard all knowledge. It will come from a change of coordinates: treating intelligence not as a stockpile inside a model, but as a flow of verified gap-closure across a federation of accountable agents.
Its conceptual climax is the Epistemic Field Equation.
The equation is proposed as a quasi-isomorphic analogue to the role of Einstein’s field equations in physics. In General Relativity, mass-energy curves spacetime, and curved spacetime governs the motion of mass-energy.
In the proposed physics of meaning, verified observations and accountable semantic relations shape — or “curve” — a semantic field. That field then governs what future claims can be responsibly connected, trusted, contested, or refused.
This is not a claim that meaning literally is gravity, nor that the Epistemic Field Equation has the empirical standing of General Relativity. It is a first-principles architectural proposal: an attempt to describe meaning as something with coordinates, boundaries, relations, propagation, and constraints — rather than as a vague by-product of statistical text generation.
The Messiah nobody asked for
Peirce died convinced by almost no one. His work was too sprawling, too difficult, too resistant to reduction. A century later, the world has built machines that use his word token billions of times a day while overlooking his more difficult message.
Meaning is not a sequence of symbols.
Intelligence is not a pile of data.
Knowledge is not fluent confidence.
A sign must be connected to something real. An assertion must be connected to someone accountable. A system must be able to say where its knowledge ends. And intelligence, if it is to be worthy of the name, must remain open to correction by a world and a community beyond itself.
Peirce wrote that specification before the computer existed.
The Semiotic Web is an attempt to build it.
For the technical architecture, see The Einstein Test and Beyond: The Architecture of the Epistemic Zero.
For a moving account of Peirce’s extraordinary life — its brilliance, isolation, and tragedy — see “Charles Sanders Peirce: America’s Greatest and Most Tragic Philosopher” by Mike Ashes.