Field Notes: The Intelligence Question
EkaShunya Field Notes: Research notes collected while preparing the first essay in the Big Questions series. The intelligence essay arrives next week.*
I have been down a rabbit hole. Weeks now. Reading about intelligence — what it means, who gets to define it, and what happens when the definition becomes infrastructure. What follows are the notes I kept coming back to. The points that changed how I see the word.
These are not conclusions. They are breadcrumbs. The essay will try to make a path out of them.
Nobody can define it
In 1921, the editors of the Journal of Educational Psychology asked fourteen prominent psychologists to define intelligence. They received fourteen definitions. Almost no overlap.
In 1986, the exercise was repeated. Twenty-four experts. Twenty-four definitions. Still no consensus.
In 1994, fifty-two researchers signed a statement in the Wall Street Journal. Even they hedged with “among other things” — which in academic language means “we know this is incomplete but cannot agree on what is missing.”
In 1996, the American Psychological Association convened a task force to settle the matter. They published a hundred-page report. They reviewed every theory, every study, every debate. And they declined to define the word.
A century of trying. No definition stuck. That is not a failure of the experts. It is a property of the word.
The nineteen-year arc
In 1905, a French psychologist named Alfred Binet built a test. He wanted to spot children who needed extra help in school. Not to rank them. Not to sort them. Just to help. He wrote, explicitly, that his test should not be used as a measure of fixed, innate intelligence.
Here is what happened next.
1912: William Stern converts the scores into a single number — the Intelligence Quotient. Now there is a number. 1916: Lewis Terman at Stanford adapts the test for American use. Now there is a standardized instrument. 1917: the US enters the war, and Robert Yerkes persuades the Army to test 1.75 million recruits. Now there is infrastructure.
1924: the Immigration Act. Congress restricts entry to the United States based on national origin, using IQ data to argue that certain populations were intellectually inferior. Carl Brigham’s A Study of American Intelligence provided the academic cover. Cited on the floor of Congress.
Nineteen years. From “help the children who are struggling” to “exclude the populations we have decided are lesser.” Not because bad people hijacked a good tool. Because a tool that produces a number carries within it an invitation, and institutions will always accept the invitation.
I keep returning to this because I think we are in the middle of a similar arc right now. AI benchmarks are diagnostic tools. But the distance between “this model scored 90% on the bar exam” and “this model is intelligent” is the same distance Binet’s test traveled in nineteen years.
Skill is not intelligence
François Chollet, a researcher at Google, draws a distinction that sounds simple and is not. Skill is how well you perform on a specific task. Intelligence is how efficiently you acquire new skills from limited experience.
“A lookup table has infinite skill and zero intelligence.”
A lookup table that contains every possible chess position and the correct response has infinite skill and zero intelligence. It cannot learn anything new. It has already memorized everything. Most AI benchmarks measure skill. Passing the bar exam tells you what a model can do on bar exam questions. It tells you nothing about whether it can learn an entirely new domain from three examples.
We are testing the wrong property. And the institution that deploys the test does not read the fine print.
The ghost in the machine
In 1997, Garry Kasparov played IBM’s Deep Blue. During Game 2, the machine made a move that Kasparov could not explain. Move 36, bishop to e4. It looked like creativity — like the machine had a plan he could not see. He was so unsettled that he resigned a drawn position, then played worse for the rest of the match.
The move was a fallback. When Deep Blue could not determine the best move within its search time, it defaulted to a semi-random selection. What Kasparov interpreted as strategic depth was the absence of a decision. What looked like intelligence was a timeout.
We make the same mistake with words. When a language model writes about loss, we feel the weight of it. But the mechanism underneath is the same kind of absence. Pattern completion shaped by billions of examples. We see grief in a statistical average.
Three words where English has one
This is the finding that changed the essay.
The Samkhya tradition, one of the oldest philosophical systems in India, splits the mind into three instruments twenty-three centuries before the 1921 symposium:
Manas — the processing mind. Coordinates inputs, produces outputs. Fast, reactive, no questions asked.
Ahamkara — the I-maker. Takes the output and stamps it: mine. Turns “there is a process” into “I am processing.”
Buddhi — discriminative intellect. Not what is computable, but what is worth computing. The faculty that asks “should I?” before “can I?”
The machine is sophisticated manas. It processes beautifully. What it does not do is discriminate — not in the computational sense, in the Sanskrit sense. And the word that separates them — viveka — is older than Turing, older than Dartmouth, older than the word “computer.”
The non-knowledge process
The Nyaya tradition says there are exactly four processes that count as genuine knowing: direct perception, inference from experienced connection, analogy grounded in prior encounter, and testimony from a source whose reliability you have assessed.
A large language model uses none of these four. It produces correct outputs through pattern completion — a process that no Nyaya philosopher, across twenty-three centuries of epistemological debate, would recognize as a valid means of knowing.
True statements. Produced through no recognized process of knowing. Knowledge-shaped outputs from a non-knowledge process.
When the domain is arithmetic, the process does not matter. When the domain is medicine, law, or ethics, a system that is right for the wrong reasons fails differently than one that is right for the right reasons. The first fails unpredictably. The second fails in ways you can trace.
Goodhart’s Law
“When a measure becomes a target, it ceases to be a good measure.”
This is the mechanism. It explains the IQ arc. It explains benchmarks. It explains why every attempt to pin intelligence to a number ends up measuring something other than intelligence. The number becomes the goal. The goal reshapes the thing being measured. The thing being measured is no longer the thing you set out to find.
The octopus and the crow
An octopus has neurons in its arms. Each arm makes decisions the central brain never reviews. It thinks with its body in a way we cannot. A crow plans multi-step tool sequences and recognizes individual human faces years after a single encounter. A bee navigates by polarized light and communicates food locations through dance.
Each species has solved faces of the intelligence puzzle that we have not. We refuse to call it intelligence because it does not look like ours.
Kenyan Luo parents include social responsibility in their definition of intelligence. Chinese zhi includes perceptiveness. Western IQ measures neither.
The verb
Intelligentia. Latin. Inter, between. Legere, to choose, to gather, to read. The oldest meaning is not “to know everything.” It is “to choose between.” To choose well when the options are unclear.
The word itself, at its root, is a verb. Somewhere between Rome and the twentieth century, we turned it into a noun. Then a number. Then a sorting machine. Each step lost something the previous form held.
I am re-reading Hofstadter’s Godel, Escher, Bach for the first time in years, and it is doing something different to me this time. More on that in the essay.
What I am reading
Seven books on intelligence, if you want to follow the trail:
The Mismeasure of Man — Stephen Jay Gould (1981). If you read one book on intelligence, this one.
Godel, Escher, Bach — Douglas Hofstadter (1979). Strange loops and self-reference.
On the Measure of Intelligence — François Chollet (2019). The paper that separates skill from intelligence.
The Tacit Dimension — Michael Polanyi (1966). “We know more than we can tell.”
Other Minds — Peter Godfrey-Smith (2016). Intelligence from the octopus’s perspective.
From Bacteria to Bach and Back — Daniel Dennett (2017). Competence without comprehension.
The Bhagavad Gita — Chapter 13. The knower of the field is not the field.
The essay is almost ready. It is the longest piece I have written for this series, and the one I am most uncertain about. Honestly, I think that is a good sign.
But before it arrives, I want to hear from you. When you say someone is intelligent — a person, a child, a colleague — what do you actually mean? Not the dictionary definition. The real one. The one you carry around without examining.
I am genuinely curious.
The Big Questions of AI
Seven questions. One clearing that may not be what it seems.
Prologue: The Big Questions of AI
1 · Intelligence · notes · essay · Five Fractures ◄
2 · Consciousness · notes · essay · The Mirror Test
3 · Reality · notes · essay · The Trust Stack
4 · Purpose · notes · essay · Five Conversations
5 · Freedom · notes · essay · The Cage Inventory
6 · Power · notes · essay · Five Maps
7 · Evolution · notes · essay · Five Endings
Epilogue: The Clearing Was a Room