How Much Can We Eliminate from Intelligence Until It Vanishes?

Intelligence cannot be defined, only cornered. Every attempt to pin it down dies on its own application. But something survives every attempt to eliminate it, and that residue is the only honest answer we can get.

The three posts before this one traced what got eliminated, historically, to produce the words “AI” and “AGI” — Dartmouth’s renaming of cybernetics in 1956, the Turing test as behavioral substitute in 1950, and the 2017–2026 capital consolidation around scaling. Each of those was an elimination performed by humans, on the field of inquiry, with consequences that compound to this day. What I want to do here is invert the move. Instead of asking what got eliminated to define intelligence, ask: what survives if we try to eliminate intelligence itself? Run the same operation on the term that was run on the field. See what’s left.

The most common answer to “what is intelligence?” is the IQ score. People say it loosely; the field that built the test says it more carefully but isn’t far off.

The history doesn’t support that picture. Charles Spearman, in 1904, noticed that performance on different cognitive tests is positively correlated, and named the common factor “g” (“‘General Intelligence,’ Objectively Determined and Measured,” American Journal of Psychology). Louis Thurstone (Primary Mental Abilities, 1938) immediately disputed the single-factor reading, arguing for several primary mental abilities — verbal, spatial, numerical, perceptual, memory, reasoning, word fluency — as more honest. The field has been arguing about how many factors there are ever since. The current dominant framework (Cattell-Horn-Carroll, or CHC theory) settles on three strata and dozens of narrow abilities, but the synthesis is administrative; it doesn’t resolve the underlying question of whether g names a real thing.

Stephen Jay Gould’s The Mismeasure of Man (1981, revised 1996) is the canonical historical critique: IQ testing was tangled with eugenics from the start, and the statistical operation called “factor analysis” doesn’t discover g; it imposes a single dimension on data that was always multi-dimensional. Cosma Shalizi’s online essay “g, a Statistical Myth” (2007) tightened the technical version of the argument: positive manifold (the empirical fact of cross-test correlations) does not entail a unidimensional underlying trait, and the inference from the one to the other is not statistically licensed. Defenders push back; the dispute hasn’t resolved.

Even granting the test, the data won’t sit still. The Flynn effect — James Flynn’s documentation that IQ scores have risen by roughly 3 points per decade across 14 nations through most of the 20th century (“Massive IQ gains in 14 nations: What IQ tests really measure,” Psychological Bulletin, 1987) — is not survivable for a strong reading of g as a stable underlying trait. If genes change slowly and the trait is genetic, the trait shouldn’t shift this fast. If schooling changed it, then the test wasn’t measuring the trait, it was measuring schooling. Either way, the interpretation is forced. Worse, the effect has now stalled or reversed in some countries — the Norwegian and Danish data show declines from the mid-1990s onward (Bratsberg & Rogeberg, PNAS 2018). Whatever IQ measures, it’s not a fixed thing.

Cross-cultural pressure compounds the picture. Robert Sternberg’s triarchic theory (Beyond IQ, 1985) and his later work on practical intelligence in non-Western contexts argue that the analytical intelligence IQ measures is one of three at most, and that different cultures grow different intelligences — Kenyan children scoring high on knowledge of natural herbal medicines score lower on academic IQ-style measures, and vice versa. None of these critiques individually buries the test; collectively, they make the claim “intelligence is what IQ measures” untenable as a definition. The test is one operationalization. It produces a number. The number is useful for some prediction tasks. It is not the thing.

The harder version of the problem appears when we look at how many definitions of intelligence the field has actually produced. Shane Legg and Marcus Hutter, in 2007, surveyed the literature and collected over 70 distinct definitions (“A Collection of Definitions of Intelligence,” in the Frontiers in Artificial Intelligence and Applications series). They tried to compress them into one — “Intelligence measures an agent’s ability to achieve goals in a wide range of environments” — which is honest but vacuous, since “goals,” “wide,” and “range of environments” all do undefined work. François Chollet’s “On the Measure of Intelligence” (2019, arXiv) goes further: he argues that no benchmark optimized against can measure general intelligence, because the moment intelligence is operationalized as a benchmark score, the field optimizes for the score rather than the underlying capacity. His ARC test was an attempt at a measure resistant to memorization. The field promptly began trying to memorize it.

The pattern is the same as IQ at a different scale: every attempt to operationalize intelligence as a measurement produces a measure, the measure gets gamed, the original concept slips out from under the operation. Intelligence survives the elimination of measurement — but only because it was never coextensive with the measurement. The measurement was a proxy that lost its referent.

That opens the sharper question. Why does intelligence survive the elimination of every measurement we apply? Only one explanation works: the assumption that there is some true definition out there waiting to be found, and our instruments are imperfect approximations of it. Drop that assumption and “intelligence” looks less like a hidden referent and more like a category error — a term that was never self-consistent under its own application. Wittgenstein made the structural point sixty years earlier (Philosophical Investigations, §66–67, posthumous 1953) about “games”: there is no single feature common to all games, only a network of overlapping similarities — family resemblances. “Intelligence” appears to be the same kind of term. Ask “intelligence of what, of whom, in what context, decided by whom?” and it dissolves. A self-consistent term leaves no room for those questions; an operationalized term answers them by stipulation; a family-resemblance term has nothing fixed to answer them with.

Bridgman’s operationalism (The Logic of Modern Physics, 1927) tried to dispose of these terms by definition: a concept is the operations used to measure it, full stop. The price is brutal. Intelligence-as-IQ and intelligence-as-Turing-passable become different concepts, sharing only a label. By the operationalist’s own rule, you can’t compare them, because they aren’t the same thing. Putnam’s natural-kind semantics (“The Meaning of ‘Meaning,’” 1975) goes the other way: kind terms get their meaning from the underlying nature of what they refer to. Water means H₂O whether or not anyone knows it. The catch: you need an underlying nature for the term to refer to. With “water” we’re lucky; with “intelligence,” there’s no consensus that there is one.

The same circularity appears in another common definition — that intelligence is information manipulation. Of what, by whom, to what end, under what dynamics? Same pattern. Every additional question undoes the definition rather than refining it.

So we’ve eliminated measurement and we’ve eliminated definition-as-stipulation. What remains, surprisingly, is the bare intuition that we know intelligence when we see it — at least in the cases we care about. That intuition is anchored to humans. It’s worth asking what happens when we try to remove the anchor.

The problem in my view stems from viewing intelligence from a human-only perspective. The true question is not whether animals or other species are intelligent; it’s where we draw the boundary between what is intelligent and what is not, because that boundary is the distinction and the definition itself. We need to wonder: is a bacterium intelligent simply because it survives in its environment and replicates? Is a river intelligent for carving out a path on a mountain? Is a raven intelligent for solving easy puzzles to get food? Is anything that exists intelligent simply because it exists and persists?

Frans de Waal posed the question crisply (Are We Smart Enough to Know How Smart Animals Are?, 2016): the right question isn’t “how human-like is the animal?” but “what is the cognition of the animal for, in its own life?” Sara Shettleworth’s Cognition, Evolution, and Behavior (2010) gives the comparative-cognition framework that takes that question seriously. The anthropocentric reading — intelligence-as-human — collapses on contact with the data. Octopuses, corvids, and cetaceans pass tests they were never supposed to pass. Slime molds solve maze problems. Bacteria coordinate. The boundary keeps moving as the methodology improves.

But “moving boundary” doesn’t yet license the strong reading — that everything which eliminates possibilities is intelligent. Take the river example. A river has many possible paths down a mountain; only one gets carved. That is, formally, an elimination of possibilities. But the elimination is exogenous — gravity, water mass, terrain composition do the work. The river isn’t doing anything; it’s being done to. The selection comes from outside. Contrast the bacterium: it has many possible behaviors, and it picks one based on its sensors and internal state. The elimination is endogenous — the bacterium senses, integrates, decides. The selection is driven from within.

This distinction does the work the term “intelligence” was reaching for and missing. It is also, exactly, the distinction Wiener and the cybernetics tradition were after. The first post in this trilogy quoted it: “elimination cannot be exogenous. A system that is acted upon by external selection is being sculpted, not cognizing. A system that eliminates its own non-viable states through self-referential dynamics is constituting itself as a knower.” The river is sculpted. The bacterium constitutes itself. Without this distinction, the term extends to everything physical — every avalanche, every chemical reaction, every fall toward equilibrium is “intelligent” because it eliminates possibilities. That’s pansychism by another name. With the distinction, the term applies to a specific class: systems where the elimination is driven from inside.

I want to be precise that this is a retraction. An earlier draft of this post argued that the river was intelligent in a “weaker sense” — that it had the operation but not the recursive closure that made it visible to itself. I no longer think that framing was right. The right framing is the exogenous/endogenous one. Weak versus strong is a continuum; exogenous versus endogenous is a kind difference. The river has the wrong kind of elimination, not a weaker amount of the right kind.

With that distinction in hand, we can run the elimination move properly. We start with humans, the anchor, and strip features one at a time. At each step the question is: does the endogenous-elimination capacity survive?

Start with the most distinctly human feature: language. The intuition is strong that language and intelligence are bound up together — that thinking, in some serious sense, is talking to oneself.

The cases against this are now extensive. Alex, the African Grey parrot Irene Pepperberg worked with for thirty years (The Alex Studies, 1999; Alex & Me, 2008), demonstrated competences the linguistic-only view cannot accommodate: he counted to six, identified colors, materials, shapes, and the category of question being asked, requested objects he wanted by name, used “none” to indicate absence — a non-trivial conceptual achievement — and asked questions of his own, including, while looking at his reflection, “what color?” His final words to Pepperberg the night before he died: “You be good. I love you. See you tomorrow.” Whether this is “language” in the full sense is contested; what isn’t contested is that the underlying cognition exists.

New Caledonian crows are the textbook case for cognition without verbal symbol systems. Gavin Hunt’s 1996 Nature paper documented their tool manufacture in the wild — pandanus-leaf tools with consistent designs across populations, transmitted across generations. Weir, Chappell, and Kacelnik (Science 2002) put a captive crow named Betty in front of food in a tube and a piece of straight wire she had never seen before. She bent the wire into a hook and used it. The act required mental representation of the goal, of the tool, of the modification needed, and of the action that would produce the modification. No verbal language was involved.

Chimpanzees were the founding case in Jane Goodall’s 1964 Nature paper documenting termite-fishing — modified twigs, used by individuals, transmitted as cultural traditions across communities. Cetaceans extend the pattern at scale: Hal Whitehead and Luke Rendell’s The Cultural Lives of Whales and Dolphins (2014) summarizes decades of research on stable cultural traditions in whales and dolphins — distinct dialects in killer whale clans, hunting techniques transmitted within family groups across generations. Reiss and Marino (PNAS 2001) demonstrated mirror self-recognition in bottlenose dolphins, the conventional benchmark for self-awareness. Plotnik, de Waal, and Reiss (PNAS 2006) extended the result to elephants. Elephants further show grief behavior, decades-long social memory, and ritualized response to the bones of the dead — phenomena hard to describe without the word “cognition.”

The most radical case is the octopus. Martin Wells’s Octopus: Physiology and Behaviour of an Advanced Invertebrate (1978) is the foundational compendium; Peter Godfrey-Smith’s Other Minds (2016) brings the case forward to a general audience. Octopuses learn quickly, escape from sealed apparatus, recognize individual humans across days, and execute complex camouflage by independently controlling millions of skin chromatophores — likely the most computationally intensive perceptual task any animal performs. Their nervous systems are anatomically distributed: roughly two-thirds of the neurons sit in the arms, not the brain. They are intelligent, by any honest reading, in a way that bears no resemblance to human verbal intelligence. The lineage that diverged from ours roughly half a billion years ago re-evolved cognition independently, on a different body plan, with no language at all.

The language defense, at this point, has degraded to “without verbal-symbolic processing there is no real intelligence” — which begs the question. Verbal processing is one specialization of cognition, evolved in one lineage, on one body plan. It is not the form. Endogenous elimination of possibilities — the bacterium’s move and the human’s move both — survives the removal of language entirely.

Push further. The animals I named all have self-models. They draw clear distinctions between self, environment, and action. The distinctions show up in mirror self-recognition tests, in tool use that requires planning around their own body, in deception that requires representing another mind. This is self-modeling, in the cognitive-science sense. Does intelligence survive the elimination of self-modeling?

The answer is yes, and the cases are extensive enough to be inconvenient for the strong-reading view.

Bacteria are the cleanest demonstration. They use CRISPR-Cas systems to recognize invading viral DNA and remove it while leaving their own DNA intact (Doudna & Charpentier, Nobel Prize 2020). The system encodes “self” molecularly — sequences matching the bacterial genome are spared; sequences that don’t are destroyed. There is no representation of selfhood; there is the operation of distinguishing self from not-self at the level of nucleic acid pattern, with zero cognitive content. Quorum sensing (Bonnie Bassler’s foundational work through the late 1990s and early 2000s) extends the pattern: bacteria detect local conspecific density via secreted signal molecules and switch behavior at thresholds — a “social self” with no representation of one. Pamela Lyon, in “The cognitive cell: bacterial behavior reconsidered” (Frontiers in Microbiology, 2015), argues directly that minimal cognition — sensing, integration, valence-based decision — is fully present in single-celled organisms.

Michael Levin’s bioelectric work pushes the boundary further. Planaria regenerate their full body plan from a tail piece, and the body plan is determined not by gene expression alone but by bioelectric pattern fields that act before the genes do (Levin, “The Computational Boundary of a ‘Self,’” Frontiers in Psychology, 2019). Frog embryos with disrupted bioelectric patterns develop face anomalies the gene-expression model cannot predict; restoring the bioelectric pattern restores the face. The “self” being maintained is at the cellular collective level, with no nervous system required. If self-modeling means “explicit symbolic representation of self,” none of this counts; if it means “operational distinction between what counts as self and what doesn’t,” it counts at every scale of life.

Slime molds are the cleanest demonstration that self-relevant computation runs without anything resembling a self-model. Tero, Takagi, Saigusa, Ito, Bebber, Fricker, Yumiki, Kobayashi, Nakagaki (Science 2010) — the Tokyo subway map paper — showed Physarum polycephalum finding near-optimal network designs by laying down protoplasmic veins between food sources arranged like cities. Subsequent work from Audrey Dussutour’s group (Boisseau, Vogel & Dussutour, Proc. R. Soc. B 2016; Vogel & Dussutour 2016) showed that slime molds habituate to a repellent and can transfer the adaptation by cell fusion to naive individuals. No nervous system. No self-model. Self-relevant action, learning, decision-making.

Plants are where this gets contested, and the contest itself is informative. Stefano Mancuso and Alessandra Viola, in Brilliant Green: The Surprising History and Science of Plant Intelligence (2015), and Anthony Trewavas’s “Aspects of plant intelligence” (Annals of Botany, 2003), make the strong case: plants have electrical signaling between organs, self/non-self discrimination at the root level (Gruntman & Novoplansky, PNAS 2004 — root architecture changes based on whether neighboring roots are self or other), context-dependent foraging behavior, and chemical alarm signals that propagate through forest mycorrhizal networks. They argue this constitutes intelligence in any non-question-begging sense. The other side pushes back hard: Lincoln Taiz and seventeen co-authors in “Plants Neither Possess nor Require Consciousness” (Trends in Plant Science, 2019) argue the field has overreached, that “intelligence” applied to plants imports cognitive vocabulary the data cannot support.

Both sides have careful evidence and careful arguments. Neither convinces the other. That fact is itself worth holding still and looking at. Two careful sides cannot agree on whether plants are intelligent. The disagreement isn’t a research deficiency that more experiments will resolve — it’s a feature of where the term “intelligence” sits. The boundary is being drawn where the term doesn’t have a unique answer. In post 5 I’d later note that undecidedness is itself the evidence; the same point applies here. The term “intelligence” doesn’t fail at random. It fails at exactly the points where it is being asked to do work it was never designed for. The plant case isn’t a problem the term will eventually solve. The plant case is the term hitting its own boundary.

What this means operationally: the bacteria, slime molds, and plants all do endogenous elimination — they sense, integrate, and select responses on the basis of internal state. None of them model themselves. Self-modeling can therefore be eliminated and the operation we keep pointing at as “intelligence” still runs. The model is downstream of the operation, not constitutive of it.

The next layer is adaptation. The intuition is that intelligence is the capacity to handle novel circumstances — that without adaptation, you have a fixed reflex, and a fixed reflex isn’t intelligent.

This is the candidate that survives best. Adaptive systems have intelligence, sometimes obviously. But fixed systems can have it too, in a degenerate-but-real sense. Horseshoe crabs (the xiphosuran lineage that includes Limulus) have held a recognizable body plan for hundreds of millions of years; fossils close to the modern form go back roughly 300 million years. The coelacanth (Latimeria) was thought extinct since the Cretaceous until Marjorie Courtenay-Latimer brought one to scientific attention in 1938 off South Africa; the lineage has been morphologically stable on comparable evolutionary timescales. These animals are not actively adapting in any meaningful sense — they sit at fitness optima that hold under the environments they encounter. They are nonetheless solutions to the problem of how-to-exist-in-environment-X. Local optima are real, even when nothing is moving toward them anymore.

The mathematics of this is Banach’s fixed-point theorem (Banach, 1922): any contraction map on a complete metric space converges to a unique fixed point. Once at the fixed point, the iteration produces the same result forever — not because it stops computing, but because the computation lands where it started. Stuart Kauffman’s NK fitness landscapes (The Origins of Order, 1993), built on Sewall Wright’s classic 1932 adaptive-landscape framework, generalize this: in any non-trivial fitness landscape, there are many local peaks where adaptation has nowhere to go. A species sitting on a peak isn’t adapting; it’s stable. Stability under perturbation is one form of intelligence, in the operational fixed-point sense — the system has converged to a solution and holds it.

Gradient descent on a convex loss does the same thing. Any modern deep network, after sufficient training, sits at a local minimum (or near-minimum) and produces outputs that are consistent with the training distribution. The “intelligence” of the trained model — to whatever extent the word applies — isn’t in its capacity to adapt; that capacity ends at the close of training. It’s in the structure that the optimization converged to.

So adaptation can be eliminated and the underlying capacity (sit at a good point in problem-space) survives. But there’s a sharper version of the question hiding here. The horseshoe crab’s optimum was found by natural selection — an exogenous process, by the distinction we drew earlier. The bacterium’s responses were also shaped by selection over evolutionary time. If the only “adaptation” that counts is endogenous adaptation — the system itself doing the adapting in real time — then the world has very few examples. The vertebrate immune system’s clonal selection counts. Behavioral learning in animals counts. Most else is exogenous adaptation, which is sculpting, not cognizing. So the layer reveals a forced choice: either count exogenous adaptation, in which case rivers and weather systems are intelligent (which we already rejected), or restrict to endogenous adaptation, in which case very little is. The honest position has to draw the line at endogenous, and at that line, intelligence-as-adaptation collapses back into endogenous-elimination — which is just the operational definition we’ve been circling.

That leaves embodiment. Even horseshoe crabs exist via a body. So does every other case we’ve considered. Can intelligence be eliminated by removing embodiment?

When typically invoked, embodiment refers to the fact that humans and animals have biological bodies, whereas machines don’t. Embodiment as a criterion only does work if it means biological embodiment specifically; under any broader reading, it includes everything (every machine has a substrate of some kind) and distinguishes nothing. The biological-only reading begs the question — it defines intelligence as biology and then notices that intelligence is biological. So the interesting question is the middle position: substrate-coupling. Does intelligence require coupling to some substrate?

Eliminating biological embodiment leaves a binary question: does substrate-coupling-of-any-kind remain necessary, or does intelligence persist without any substrate-coupling at all? Trying the second choice — eliminating all coupling — produces an error. The elimination operation itself requires a substrate. The very act of eliminating substrate-coupling has to be performed by something, on something, somewhere. Pure substrate-free intelligence is a fiction in the same way that pure substrate-free information is. Even mathematical proofs are inscribed in marks on paper, in chalk on slate, in patterns of charge in a transistor. The proof isn’t in the abstract; it is in the physical record.

Landauer made this point physically in 1961 (“Irreversibility and Heat Generation in the Computing Process,” IBM Journal of Research and Development): erasing one bit of information requires at least k_B T ln 2 of energy dissipated as heat. A bit is not a Platonic entity. It is a configuration of physical degrees of freedom that can be set or reset, and the resetting heats the world. The 4E cognition movement (Andy Clark’s Being There, 1997, and Surfing Uncertainty, 2016; Shaun Gallagher; Evan Thompson) makes the same point at a different level — cognition is constitutively embodied, embedded, extended, and enactive; it is not in the head. Both arguments converge on the same conclusion: the substrate isn’t optional. The question isn’t whether intelligence requires a substrate (it does). The question is what kind of substrate-coupling counts.

That question is the boundary problem from earlier, in another costume. Consider embodiment as differentiation: in order for something to be contained within anything, there must be distinguishability between the contained and the container — a boundary, with information crossing it. Embodiment reduces, formally, to the existence of a boundary that is being maintained. Maintaining a boundary is what bacteria do molecularly with CRISPR, what plants do biochemically at their root tips, what cells do via autopoiesis, what brains do via predictive control — and what we keep pointing at when we say “intelligent.” So the embodiment layer doesn’t eliminate intelligence; it shows that intelligence can’t be defined independently of boundary-maintenance. The two are entangled.

We have, then, a troubling end. If we eliminate existence we can definitely say there is no intelligence there, but it is also a category error to say “eliminate existence” — we might as well say “eliminate the universe.” The only fair conclusion we can draw is that intelligence can survive almost anything we try to eliminate, and still find residuals of itself. Intelligence survives every elimination because elimination requires a substrate to eliminate from, and that substrate is what we keep pointing at when we say “intelligent.”

But I have not been fully honest. If intelligence is what keeps surviving elimination, why does it keep surviving? Because there is something we can’t eliminate — the eliminator, or the act-of-eliminating itself. And even there we can try the move once more: if every attempt to eliminate intelligence fails because the operation is what does the eliminating, then intelligence is not the residue. It is the operation. The thing that does the surviving and the eliminating is the cognitive act.

For every layer stripped, the same structural feature appeared: what got eliminated was a thing. Measurement-as-IQ is a thing. A definition is a thing. Self-modeling is a thing. Adaptation, embodiment, language — all things. And every thing turned out to be removable. By the structure of the argument, what’s left can’t be a thing. A thing-like residue would have failed under one of the layers; that’s what the layers did.

What it seems to be, instead, is an operation — specifically, the endogenous elimination of possibilities. The same operation bacteria perform with CRISPR, plants perform biochemically at their root tips, slime molds perform across their protoplasmic veins, horseshoe crabs hold via an evolutionary optimum, and brains perform via predictive control. None of these have intelligence as a separable thing. All of them have the operation. The operation is what runs; the appearance of intelligence-as-thing is what running the operation generates from the inside.

The operation requires something to operate on, and whatever that substrate is, it is intermingled with the cognitive act in a way that can’t be teased apart. The operation and what it operates on aren’t separable: elimination is the production of distinction. There is no boundary-drawing without sides, and no sides without boundary-drawing. Intelligence is neither the substrate nor the operation; it is the operation-on-substrate where neither half could be removed without removing the other. And the operation acts on itself — the very claim that intelligence is the operation is itself an act of endogenous elimination, performed right now, on the term. The argument is what it is arguing about.

The only survivor of elimination is the very act itself.

That residue is interesting because it’s portable. We just ran the elimination move on the term “intelligence” and got back an operation. There is nothing about the move that is specific to intelligence. Run it on something else — on “self,” on “truth,” on “cause” — and we should get back the same shape: a thing-like answer fails under inspection, but the operation that does the inspecting doesn’t. What else can we apply this to?