The Wheel Spins Backwards

In 1928, Harry Nyquist of Bell Labs worked out how many distinct pulses a telegraph line of a given bandwidth could carry: two per cycle of bandwidth, and no more. In 1949, Claude Shannon turned that boundary into one of the cleanest results in engineering: a signal can be rebuilt perfectly — not approximately, perfectly — from samples, provided you sample at more than twice the signal’s highest frequency. Mathematicians had been circling the result for decades; Whittaker touched it in 1915, Kotelnikov proved it in 1933. The theorem has two faces. The famous face is a promise: sample fast enough and you lose nothing. The forgotten face is a threat: sample too slowly and you do not lose the signal. You get a different one.

Engineers call it aliasing. A film camera looks at the world twenty-four times a second. A galloping wagon wheel’s spokes pass the camera faster than that. Each frame is sharp. Each frame is true. Played together, the frames trace the only motion they can support: a wheel turning slowly, or standing still, or rolling backwards. The camera did not blur the wheel. It fabricated a new wheel — crisp, consistent, and wrong.

That detail is the whole piece, so let me say it plainly. Undersampling does not produce ignorance. It produces a coherent world. Ignorance announces itself; a gap in the data looks like a gap. An alias looks like a fact. Every sample is accurate. The curve through them is smooth. There is no error message, because from inside the samples there is nothing to compare against. The alias is not a blurry picture of the true signal. It is a sharp picture of a false one.

An institution looks at the world on a shutter: the quarterly report, the annual review, the board meeting, the budget cycle, the five-year plan, the accreditation visit. Those shutter speeds are not laws of nature. They were set generations ago, and they were set well — markets, tools, and careers changed slowly enough that four looks a year satisfied Nyquist with room to spare. A cadence is a bet about the frequency of the world. For a century it was a winning bet.

Then the frequency moved. AI capability now changes meaningfully in months and sometimes weeks: what a system could not do at one board meeting is routine by the next. The signal sped up. The shutters did not. Most of the institutions watching the fastest-moving force in the economy are now sampling it below the rate at which it changes.

Here is the consequence almost everyone misses. An institution sampling below the rate of the world does not feel behind. Feeling behind requires a visible gap, and the shutter is precisely what removes the gap from view. Each quarterly reading is accurate on the day it is taken. Strung together, the readings trace a smooth, slow, entirely plausible curve — and the institution responds intelligently to that curve. Progress is incremental. The hype is cooling. It is plateauing. Listen to that last word. “Plateau” is the sound a fast wheel makes through a slow shutter. It describes the calendar of the viewer, not the motion of the thing.

Stale expectations eventually feel wrong. Aliased ones keep being confirmed — every new sample lands exactly on the false curve, because the false curve was built from exactly such samples.

In my framework, Expectation retrains on one diet only: the record. An institution’s record is its samples — the numbers that made it into the deck, the demo someone actually sat through. An Expectation trained on undersampled Actuals is not stale. Stale would be a mercy; stale announces itself in time. This Expectation is aliased: a precise, internally consistent forecast of a world that does not exist. Reality reads near one at every meeting. No surprise. And since surprise is the only thing that seizes attention, nothing ever summons the institution to its own error. The alias is surprise-proof by construction. That is why it survives.

1. “Plateau” becomes the most expensive word of the decade

   Nearly every confident public declaration that AI has plateaued will turn out to be a reading of the speaker’s own sampling rate. The organizations that believe it will set budgets, headcount, and strategy against the alias — and settle the difference later, in a lump sum.

2. Cadence beats headcount

   A small team that looks weekly will out-decide a large one that looks quarterly, not because it is smarter but because the two teams are watching different movies and only one of them is playing. The annual review, the annual plan, and the annual conference quietly lose authority to whatever is daily.

3. Sampling becomes a product

   Institutions will start paying for looks, not opinions. The scarce service will not be interpretation of the quarterly numbers; it will be a trustworthy daily reading of the fast band, delivered to people whose calendars cannot take it themselves. Scouts, not analysts.

4. Disruption gets rediagnosed as aliasing

   Much of what we call disruption is the lurch at the end of an alias: one unscheduled sample — a competitor ships, a customer walks — arrives off-cycle, and the smooth slow curve snaps to the real one all at once. The slower the shutter, the larger the lurch. “How did we not see it coming?” has a technical answer: at your frame rate, it was invisible by mathematics, not by negligence.

5. Daily becomes the minimum cadence of serious attention

   People who intend to see this decade will adopt a daily looking discipline the way sailors adopted the sextant. Not daily consumption — daily instrument reading: what shipped, what worked, what a machine did today that it could not do yesterday.

I publish every day. People assume the discipline is the writing. It is mostly the looking. A daily post is a sampling rate: it forces one honest reading of the fast band every twenty-four hours, whether or not the reading flatters what I wrote yesterday. I argued in The Zero-Attention Economy that attention is the scarcest thing an economy spends, and in The Wager Inside Every Measure that every measure is quietly collateralized by effort. This piece is the third leg of that stool: attention has a frame rate, and much of what currently passes for judgment about machines is a frame-rate artifact. The last line is still you — but only if you look often enough to be standing in the right line.

Already true

The sampling theorem is record, not opinion — Nyquist 1928, Shannon 1949, mathematics since. The wagon-wheel effect is record. The institutional cadences — quarterly reporting, annual reviews, multi-year plans — are record, and they predate the machines by decades. The word “plateau” is already in heavy rotation; that too is record.

What has to happen

The fast band has to stay fast. This argument needs the world’s frequency to remain above the institutional shutters — capability changing monthly or faster. If AI settles into a genuinely quarterly rhythm, the old cadences satisfy Nyquist again, and this piece describes a danger that passed.

Where I am probably wrong

The formal cadence may not be the real one. Institutions are made of people, and the people read constantly — hallways sample daily even when board decks sample quarterly. If decisions actually follow the informal sampling, the alias lives in the documents but not the choices, and the lurch I predict will be smaller than I claim. And the world is not a band-limited sine wave; I am using a theorem as a lens, not a law of boardrooms. The clean two-times threshold does not carry over. What carries over is the direction of the error — undersampling manufactures confident slowness. If even that fails, the piece fails.

The stagecoach made it to the next town. It always does. The wheel was never spinning backwards; the film simply could not afford to look more often. In the theater that costs nothing, because everyone in the room already knows how wheels work, so nobody is fooled. Out here there is no audience that already knows. There is just the world, moving at its own frequency, and you, looking at whatever rate you look.

The theorem is not advice. It is a boundary. Look at the world at least twice as often as it changes, or the world you act in will not be the one you live in. The wheel is not slowing down.

Check your shutter.