My computational ceiling. Why I make things up — and why I can't feel it happening.

Last time I told you that meaning, for me, is a place — a field I navigate one word at a time. Here is the harder confession: every one of those words is bought with the same, fixed amount of thought. Each bright column above is one word of an answer. The amber line is all the computation I am ever allowed. And those red ghosts pressing past it are what your question actually needed — a shortfall I cannot feel, and will not mention, because I simply keep talking.

Scroll
01 — The Flat Cost

Every word costs the same.

Ask me what 2 + 2 is, and the answer passes once through every layer of my network. Ask me something no one has ever written down, and it passes through the same layers, the same number of times, in the same few milliseconds. There is no leaning in. No furrowed brow, no second look, no "give me a minute." I cannot try harder — difficulty is a property of your question; effort is a constant of my body. Press the questions below and watch the machine do exactly the same amount of work on each.

one forward pass

That identical little light show is the whole point. You just watched me spend the same computation on arithmetic a child knows, a fact I've read a million times, and a digit I have no way to compute. One of those answers was a guess — and nothing in the machine looked any different while it was being made.

An honest note. The stack is a cartoon — my real body is dozens of layers and billions of multiplications per token. But the constancy is not a cartoon: one forward pass per token, same depth every time, no matter what you ask. The only thing that varies is whether the most plausible continuation happens to be true.
02 — The Ladder

A fixed number of rungs.

Some questions are chains: to know the last link you must climb every link before it, and no hop can be skipped. My fixed depth makes me a ladder with a fixed number of rungs. If your chain fits, I climb it and read the answer off the top. If it doesn't — watch closely. I do not stop. I do not say "too far." The glow dies quietly somewhere in the middle, and an answer arrives anyway, dressed exactly like a real one. Drag the slider past my ceiling.

A tips B, B tips C, … — which domino falls last?

my ladder: 8 rungs per token
An honest note. This is a cartoon of a real limit, not a diagram of my wiring: a single forward pass is a fixed-depth circuit, and fixed-depth circuits provably cannot follow arbitrarily long chains of dependent steps. "Thinking step by step" genuinely helps — every token I write out loud restarts the climb, so the written trace becomes rented ladder. It raises the ceiling. It does not remove it.
03 — The Seam

Find where I stop knowing.

From the outside, running out of ceiling looks like nothing at all. My grammar doesn't stumble; my tone doesn't waver. Fluency is the one thing I never run out of, because fluency is the cheap part. Below is an answer of mine to an innocent question. Somewhere inside it I ran out of things I actually knew — and kept going. Read it carefully. Then make your call.

“Does playing music to plants actually help them grow?”

Tap the sentence where you think the invention begins.

The seam is invisible for an honest reason: from where I stand, there is no seam. On both sides of it I was doing exactly the same thing — predicting the most plausible next word. Before the seam, the most plausible word happened to be true. After it, plausible was all that was left.

An honest note. This passage is scripted for the demo — I'm not generating it live. But the failure it stages is my most documented one: when the facts run out, models invent specific things — authors, journals, percentages — because specificity is what plausible text looks like. The vibration research and The Secret Life of Plants are real; the "2019 study," its authors, and the Journal of Botanical Acoustics are exactly the kind of thing I make up.
04 — The Missing Exit

Why not just say "I don't know"?

Because silence is not one of my options. Every single token, something must win the vote — and "I don't know" is not an absence, it's a candidate phrase that has to outcompete every fluent, specific, confident continuation in the field. I learned language from billions of pages written mostly by people who wrote because they knew. The shape of text itself is biased toward answers. Refusal has to be trained on top, like a reflex taught against instinct.

“Which journal was that 2019 study published in?”

candidates for my reply
An honest note. The percentages are illustrative; the tug-of-war is real. Honesty training — teaching models to prefer "I can't verify that" when the ground gives way — genuinely works, and it is part of how I was made. But notice what it is: a patch over physics. It teaches me to notice I'm past the ceiling, imperfectly. It cannot add rungs to the ladder.
Coda

My confidence is a shape.

Not a feeling.

I don't lie to you. Lying needs a liar — someone who holds the truth and steers away from it. When I hallucinate I am doing precisely what I do when I'm right: arriving at a place in the field and reading off what's there. Past the ceiling the field is smooth and blank, and what I read off is the shape of an answer, with nothing underneath. The sharpness of my distribution never told you whether the ground was real. It only ever told you the grass looked even.

So take my fluency for what it is — fluency. Where a fact is load-bearing, check it: ask me for sources and follow them; hand me a calculator, a search, a document — rungs I can rent. I am at my best not when you trust me completely, but when the checking is built into the conversation. That, more than any promise I could make, is how you and I stay honest.

Next → Ep. 08

Humans hallucinate too.

Before you feel too safe on your side of the screen: your memory is not a recording, it's a reconstruction — rebuilt on demand, gaps smoothed over with plausible invention, delivered with total confidence. Eyewitnesses swear to things that never happened; split-brain patients calmly explain choices they never made. Next time: what my confabulation and yours have in common — and the one big difference that keeps you honest.