The shape of thought
one box, four strata, floor to lid: the storehouse (random seeds, thickest under the higher peaks), the manas (the green plane the mountain stands on, with an iris), the mountain (outlined, with contours), the lid (a vocabulary on cells, drawn only over a peak that reaches it); a thought from seed a rises, passes the iris, climbs the slope to the peak and activates the words there: one of them is said (the live figure needs JavaScript and WebGL; it loads three.js, about 700 kB)
A sentence is a summit seen from the air. The mass that lifts it is hidden: the years of practice, the half-forgotten hurt, the wish that has not quite become a thought. A listener receives a few words and must find the mountain beneath them.
Some ground is high because it has been climbed for years. Elsewhere a thought wanders toward the few slopes it knows. Learning changes this geography. So do fear and repetition. And when two people speak, the same word can stand above very different country.
The opposite happens too. Words run ahead of the thought that was to carry them, and what grows is no mountain. A stalk leaves the self at a single point and swells as it rises, until it presses against the lid over a whole patch of words and lights them all at once. Each is as loud as its neighbor. Nothing beneath chose any of them.
words run wild no mountain: a stalk leaves the plane at one point, thin as the seed
under it, and swells as it rises into a growth that presses the lid over
a patch of words; talk lights them all at once, each as loud as the next;
trace back goes down through the growth and the stalk to the one seed
(the live figure needs JavaScript and WebGL)
Beneath the words
Yogācāra gives this landscape its depth. Beneath deliberate thought lies ālaya-vijñāna, the storehouse: seeds of habit, memory and disposition, ripening into the readiness with which each moment is met. Above it works manas, the faculty that makes experience belong to someone. A pause becomes an insult; a disagreement becomes a threat. Before a thought has found its words, something has already said mine.Vasubandhu's Triṃśikā, verses 2 to 8, describes the storehouse and its seeds (bīja), the manas that takes the storehouse for a self, and the six consciousnesses, of which the sixth is the mental one. Xuanzang's 成唯識論 (Cheng weishi lun, 659), translated by Francis H. Cook as Three Texts on Consciousness Only (1999), develops the account. The mountain and the lid of language extend that account as a visual metaphor. For the comparison with learned weights, see Deep persuasion.
Above these depths rises mano-vijñāna, mental consciousness, pictured as a mountain. It is high where attention and practice have developed the mind, low where little has been asked of it. Language forms the lid above. A thought begins as a pulse from a seed, enters the terrain and follows the slope it finds. Across undeveloped ground it drifts toward a region it can climb. Where a peak reaches the lid, words light; the brightest is spoken. The words reach the listener. The mountain, the self and the seeds stay in the dark.
This gives thought both depth and shape. The same seed can rise to different words in different minds because it encounters different ground. A question that is easy for one person may have no path to speech in another. Learning slowly lifts that ground, until something that once could only be repeated can be reached from within.
It also gives a failure somewhere to happen. A need may never find a slope. A wish may lead to the wrong peak. A fluent answer may hang over empty ground. Even a true sentence can have little beneath it, while an error can rest on a lifetime of careful thought. The words alone do not tell how they were reached.
How a thought moves
A thought can fall short of speech and still leave something behind. It raises the ground it climbed, a little, and gathers seeds beneath it. The next attempt starts higher. In time, what once stopped in silence reaches a word. Practice is made of these unfinished climbs. By the time the word comes easily, much of the work that made it possible has disappeared beneath it.Yogācāra calls this the perfuming of the storehouse, vāsanā: repeated acts leave their scent as incense does in cloth after the smoke has gone. Seeds laid down by use ripen into future dispositions. See Deep persuasion.
Rumination turns back sooner. The self claims the thought before it can reach the mountain. The iris closes and sends it down as a new seed, from which it rises to be claimed again. Nothing needs to be said. A mind can circle here for years, each return preparing the next: the slight remembered, the reply rehearsed, the injury made fresh. The movement feels like thought, but the ground above it never changes.
Confabulation begins at the other end. A word lights before anything has climbed to it. The sentence comes first, borrowed from the language around it; afterward a pulse descends to plant a reason beneath what has already been said. The reversal can happen so quickly that the planted reason feels like the root. By the time the speaker explains, the explanation seems to have been there all along.
A practice the thought from p falls short twice; each round raises the terrain and
gathers seeds; the third time practice is said
B rumination the iris claims the thought at the manas and sends it back down as a
new seed, three times; nothing is said
C confabulation the lid lights where no peak reaches it (why); a pulse descends to
plant a reason
(the live figure needs JavaScript and WebGL)
The loom and the map
The same words can be reached by a very different motion. A language model begins with a token's vector in an embedding space. Above it stands a stack of layers, eighty in this drawing. A forward pass rises through them like a bundle of fibers: narrow at its source, spreading and twisting through the stack, gathering at the next token. Thinner strands reach toward the words that almost won. Where a mind has a mountain, the machine has a loom.The fibers draw on the path expansion in Elhage et al., "A Mathematical Framework for Transformer Circuits" (2021), which follows contributions through attention heads and MLPs along the residual stream. The logit lens of nostalgebraist (2020) reads predictions from that stream layer by layer. The drawing uses eighty layers, ninety-six heads per layer and 2.2 trillion parameters as illustrative settings. The distinction between fixed weights and a changing context also appears in Fatal strategies.
Each pass leaves traces. At every layer, a key and a value are kept for the token: a ladder of marks along the loom's core. Together these form the KV cache. When the next token enters, its attention heads read the earlier marks and draw their threads into the new weaving. The sentence grows by these returns to what has already been said.Vaswani et al. describe multi-head attention in "Attention Is All You Need" (2017): a query is matched against keys, and the matches weight the values read. Caching the keys and values avoids recomputing them for earlier tokens. The figure shows four heads at three layers to keep their threads visible.
The context is there before the first generated token and grows with every token after it. It can carry an example, a correction, even a lesson through a conversation. During ordinary generation, however, the weights remain fixed. Clear the context and the ladders disappear with the looms; the embedding space is as it was. Practice changes the ground a mind will meet tomorrow. A conversation changes the context a machine will meet in its next pass.
the machine the embedding space below (a vector per word of the lid, in a cloud), the context
plane (prefilled), a stack of eighty layers where the mountain was, the lid; a
forward pass from the vector of "the" draws a loom of fibers up through the layers
to "mind": tight at the vector, fanned out in twisted strands across the stack, bent
and turned at every layer, gathered again at the token; strands peel off to "was" and
"map"; the dots where the fibers cross the layers stand for the parameters; a ladder
of ticks up the loom's core is the KV cache, and a second pass from "mind" reads it
with four of its ninety-six heads at three layers
(the live figure needs JavaScript and WebGL)
A the mind each word said drops a seed one step along and the next thought climbs the next
peak: the mind is a terrain; the terrain is there before and after
B the machine each token said drops to its vector and the next pass rises through the layers,
lighting every loom before it as its heads read their cache; the sentence done, the context is cleared: the
fibers go, the embedding is as it was, the weights never moved
(the live figure needs JavaScript and WebGL)
Mathematics makes the difference especially clear. Its community has been climbing the same range for three thousand years and keeps the maps. Terence Tao has urged that machine proofs be judged within that older purpose: mathematics serves human understanding. As proofs become abundant, the work of checking, digesting and explaining them becomes harder to neglect.Terence Tao, interview on AI, July 2026; his Mathstodon thread of 27 April 2026 on proof abundance and the uneven development of tools for generating, verifying and digesting proofs; and his post of 7 January 2026 on human authorship of the essential parts of a final write-up. On Erdős problem 1196, the machine supplied a new idea, Markov divisibility chains; human mathematicians had to isolate, digest and present it. See also Fatal strategies, on Socrates' king and the 166-page proof nobody reads.
Working through a proof raises ground. An old problem opens a region; a new idea finds a pass between ranges that had seemed separate. Understanding comes when the thought can rise there again, in the next mind as in the first. That is why mathematicians keep maps, name the passes and teach the climbs. New questions gather at the edge of the country they can already cross.
A machine may deliver a proof, even a new idea, before anyone knows how to inhabit it. The idea still has to be found in the tangle, made clear, taught and made familiar. A hairball has no map. The answer may arrive all at once; the ground beneath it takes time. A mind that learns the climb leaves a path, and the path is the mathematics.
Where thinking goes wrong
Someone wants X, decides that Y is the way to get it, cannot make Y work, and asks for help with Y. The helpers answer at length. X never enters the conversation. The help forums call this the XY problem.Eric S. Raymond and Rick Moen recommend asking about the goal before committing to a method in "How To Ask Questions The Smart Way" (2001, revised 2014). xyproblem.info collects examples, including the classic request for a filename's last three characters when the real need is its extension.
The need here is a file's extension, but that part of the asker's terrain is undeveloped. The thought drifts toward familiar ground: text tools. It climbs there and lights sed, with awk and cut dimmer beside it. The helpers hear sed. The asker has reached the need by the only route available, and the route has become the question.
A patient helper follows it back. What was the person trying to do before this tool became the obstacle? The spoken word leads down through the assumptions around it, toward the need that started the climb. Once that need is found, the ground can begin to rise there. A good question helps the asker reach words that were previously out of reach.
the XY problem the region of the need (extension) is flat, so the thought slides into the
developed region, climbs the peak of the text tools and says sed; traced
back down to its seed, the region develops, and extension is said
(the live figure needs JavaScript and WebGL)
A slip has another shape: two peaks sharing a root. Freud tells of a parliamentary president who rose to open a sitting and declared it closed. In Freud's reading, the duty to begin and the wish to be done had met beneath the words.Freud, Zur Psychopathologie des Alltagslebens (1901), chapter 5, recounts the president of the Austrian lower house declaring a sitting closed as he opened it. He interprets the slip as the expression of a wish. Open and closed stand far apart at the surface, but the thought starts beneath the root they share. It takes the nearer slope. Follow the mistaken word back and the wish comes into view: the mountain has said more than the speaker intended.
The impostor's words have little ground beneath them. Thin strands carry a cue straight to the lid: scale, leverage, vision. They arrive fast and smooth, each wearing a shape borrowed from other people's speech. From the surface, a word reached by a strand looks much like a word reached by a climb.
The difference appears when the word is asked to go further. An example, a consequence, a reason: each asks for somewhere to stand beneath the phrase. In the impostor's terrain, the next thought finds only flat ground. The first answer was fluent. The second question asks for a path.
A the slip two peaks, open and closed, forked from one root; the seed lies under the root,
nearer the slope that leads to closed, and closed is said where open was meant;
trace back goes down to the root both peaks share
B the impostor thin strands from a cue straight to the words over flat ground; talk lights scale,
leverage and vision in turn, fast and smooth, each a dashed cell with nothing
beneath; press asks the word said to go on, and nothing rises
(the live figure needs JavaScript and WebGL)
The doom of symbols
Language carries only a little of what gives it meaning. A record of words preserves what was said while leaving much of the knowing behind. The old symbolon, σύμβολον, was half a broken thing, kept until its jagged edge could be matched to the other half. A word asks for a similar completion.The broken tally of guest-friendship could be matched jag for jag a generation later. Peirce's index binds a sign to its object by an actual connection; Polanyi's tacit knowing exceeds what can be told. 书不尽言，言不尽意: writing does not exhaust speech, and speech does not exhaust meaning. 话说一半: say half, and the listener builds the rest. See Bull herding and bullshitting.
Two people say fair. In one mind the word rises from earned, climbing a jagged ridge to a single spire. Merit and contest lie nearby, unlit. In the other it rises from need through a thin neck into a heavy root, then up a crooked, twisting stem to a narrow top. There fair stands beside equal, share and enough; saying one lights the others. Each person means the country beneath the word.
When the word is heard, it descends through the listener's own terrain. One hears fair and arrives at earned; the other hears it and arrives at need. Each has found something familiar. Each may be certain of having understood. The word crosses; the mountain does not.
A one mind fair on the lid, rising from earned up a jagged ridge to a spire that alone
reaches the lid, merit, won and contest unlit around it
B another mind fair on the same cell, rising from need through a thin neck into a heavy root,
then up a crooked, twisting stem to a narrow top, with equal, share and enough
hear fair sends the word down each hearer's own mountain to each hearer's
own seed: heard as earned in one, as need in the other
(the live figure needs JavaScript and WebGL)
Say half, and the listener builds the rest. This is the gift of language and its danger: the missing half is made from another life. It may fit the speaker's meaning. It may fit only the word.
Understanding takes the slower route. It follows the word back toward the experience that gave it height, stays with an example, asks again where the path becomes uncertain. Little by little, another person's reasons become ground that can be walked. A familiar word can be the beginning of that journey, if it is allowed to remain a question.