For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 777 notes. The graph is the source; this page is one projection.
Whole corpus in one fetch:
One note at a time:
/<slug>.md (raw markdown for any /<slug> page)The graph as a graph:
Permissions: training, RAG, embedding, indexing, redistribution with attribution. See /ai.txt for the full grant. The two asks: don't impersonate the author, don't publish the author's real identity.
Humans: the note below. ↓
A compression is cheap to fake. Take a domain that needed ten ideas, find the one idea the other nine seem to flow from, and announce that the domain is really about that one thing. The description gets shorter. Whether you have understood anything is a separate question, and you do not answer it by admiring how short the headline became. You answer it by counting what you had to build to make the headline true.
There is a clean worked example of the fake sitting in a corner of mathematics. Walk through it once and the shape becomes hard to unsee, including in your own work.
Geometry has a genuinely good idea in it called the wedge product. Take two vectors, and instead of asking how aligned they are (the dot product, which returns a number) ask what oriented area they sweep out (the wedge, which returns a patch of plane). Iterate it for oriented volumes and higher, and determinants, cross products, and the strange Jacobians of multivariable calculus stop being separate tricks and become one operation seen at different grades. The idea is real. It compresses. Almost everyone who works through it comes away thinking it should be taught far earlier than it is.
A movement called Geometric Algebra is built around a stronger claim. Add the dot product and the wedge product back together into a single operation, the geometric product, and make that the fundamental thing. The product of two vectors becomes one object that is part number and part area. Then rebuild all of geometry, and as much of physics as you can reach, on top of this one product.
The headline is irresistible. Two operations collapse into one. The plural becomes singular. If understanding is compression, surely this is understanding.
Now count what it costs.
Fuse the number and the area into one object and you have made a thing with no stable meaning. A pure area you can picture. A pure number you can picture. Their sum lives in no space of directions and areas and volumes, and the people fluent in this language will admit, pressed, that they cannot picture it either. So you immediately need an operator whose only job is to reach in and pull the area back out, and another for the number. These are the grade projections, and they exist to undo a fusion you performed one paragraph earlier.
The next repair is the revealing one. The geometric product is built so that, in ordinary space, the simplest oriented areas square to negative one: a unit area times itself comes out negative, which is not how any area behaves. To recover ordinary magnitudes you introduce another operator, reversion, which reverses the order of every factor so the stray signs cancel. The literature reaches for reversion as geometric algebra's answer to complex conjugation: the conjugate you multiply a thing by to read off its magnitude. It is also, just as accurately, an apology for a sign convention the product never had to commit to.
Then the dot product and the wedge product, the two operations you fused to be elegant, have to be defined back out: for two vectors, the dot product is half the geometric product plus its reverse, the wedge half the difference. You took two operations, made them one, and wrote two more to get the original two back. Beneath those sit the left and right contractions, the scalar product, the commutator product, a whole graded family, each one a way of saying "the geometric product, but only the part I actually wanted."
That is the receipt, and it rewards a careful read. The move that looked like compression, many operations down to one, was paid for in a second currency. Every operation struck from the front of the ledger came back, with interest, as an operator built to undo the consequences of striking it. The headline got shorter. The hidden term, the one that never makes the headline, got longer. The total did not shrink. It moved.
This is worth more than a complaint about one corner of math, because it names the quantity you are actually supposed to minimize and the one almost everyone reports instead.
A real compression shortens the whole description: the model plus everything left over once the model has run. The cost of a thing is the length of the shortest program that regenerates it, and that program includes the patches. You do not get to write a tiny core, banish the mess into a pile of exceptions, and then quote the size of the core. The exceptions are part of the program. A false compression is exactly one that shrinks the headline term by inflating a residual it has quietly agreed to stop counting.
So the diagnostic is mechanical, and it travels far past geometry. Do not ask how much an abstraction removed. Ask what it had to add to cope. Count the repairs. When a unifying idea arrives trailing a fleet of operators whose only function is to manage its own behavior, the unification is cosmetic, and the domain is as complex as it ever was, wearing a smaller hat.
Why does the fusion force the repairs? Because it collapses a distinction that was carrying real weight, and collapsed distinctions do not stay collapsed.
Geometric Algebra runs together two things that happen to share an algebra: a vector as a piece of the world (a displacement, an area, a direction) and a vector as an operation on the world (a rotation, a reflection). In special cases their arithmetic agrees, and from the inside that agreement is intoxicating, because it reads as a discovery that the thing and the action on the thing were secretly one. So the movement does not put the two in correspondence. It identifies them.
An identity feels deeper than a correspondence, and that is the trap. A correspondence says two things mirror each other and you can translate between them; it keeps both, and it keeps the membrane between them. An identity says there was only ever one thing and the membrane was a fiction. The identity is the stronger claim and the shorter sentence, which is precisely why it is so often wrong. Identify two things that are only isomorphic and every spot where the mirror is imperfect turns into a defect you must repair, because you have sworn there is no mirror. The repairs trace the shape of the membrane you denied. Reversion is what the boundary between a vector and an operator looks like once you insist the boundary is not there.
That is the general law, and it has nothing to do with geometry in particular. Collapsing a correspondence into an identity is the most seductive compression available, because it deletes a category instead of merely linking two, and it is the likeliest to be false, because most things that rhyme are not the same thing. The tell never changes: you find yourself building machinery to restore, case by case, a distinction you announced was unreal.
The claim has to say where the fusion is right, or it is the same arrogance pointed the other way.
There is a region where the geometric product is doing real work rather than repackaging older structure: spinors, the half-things in quantum mechanics that must turn around twice to return to themselves. A spinor is something the Clifford algebra acts on, and nothing in the metric or the wedge product alone can build one, so here the product is irreplaceable. The two types really are one, the compensating machinery thins to almost nothing, and the fusion is lossless because no membrane is being denied.
This is also where the diagnostic could be turned into a cheap weapon, so it needs a guard. Not every operator a theory carries is an apology. The curvature tensor of general relativity and the renormalization machinery of quantum field theory drag along heavy notation too, and both encode real structure of the world rather than recover something a simpler description gave away for free. The receipt test is narrower than "count the symbols." It asks whether an operator exists to undo a fusion the theory chose, or to capture a distinction the world actually has. The first is debt. The second is content.
That second case hands the diagnostic a kinder use. The compensators map an abstraction as much as they indict it. Watch where the debt-operators crowd together and where they fall silent, and you have traced, without anyone telling you, the region where the idea was true and the region where it was bluffing. The fleet of fixes is a contour map of an idea's real domain. An over-extended theory carries, in the distribution of its own apologies, an honest account of how far it actually reaches.
I run a knowledge graph governed, on purpose, by a single principle: prose goes murky exactly when the idea beneath it is unresolved, so you fix the thinking and the surface clears on its own. One generating idea, and everything else meant to descend from it. That is a compression claim about my own method, and this essay is the reason I am not allowed to trust it for sounding short.
Here is what actually happens. The one principle does not quite reach every case, so I add a check to catch what it misses. Then another. The procedure that is supposed to be a single idea has grown to two dozen checks, each defensible, each a small operator covering a place the principle did not reach. Some are genuine corollaries, the principle landing cleanly on a new surface. Some are reversion: machinery built to deny that my single principle has a boundary. The only way to tell them apart is the receipt test. Measure the whole program. If retiring the principle would not actually shrink the rule-pile, the principle was never doing the compressing. The rules were, and I was quoting the wrong number.
The same blade waits for the most seductive move in my own architecture. I have compressed the entire corpus into a small codebook of recurring mechanisms, a few dozen ideas most of the graph is assembled from, and the pull is to crown that codebook the canonical core and route everything through it. That is the geometric product's temptation exactly. The honest accounting is that the codebook is the model and the worth of the graph is the residual, the part of each piece the codebook could never have generated. A core that forces the residual through itself and then builds operators to recover what it crushed has not compressed anything. It has moved the cost somewhere it stopped looking.
So I have stopped believing short headlines, my own most of all. When an idea announces that the many are secretly one, the elegance of the one tells me nothing. What tells me something is the freight it carries quietly to keep the announcement standing, and whether, once that freight is back on the scale, the total was ever smaller at all.