The Mechanism
Three things on the same front page: a typewriter exercise at Cornell, a B-52 electromechanical star tracker, and 512 comments about why Opus 4.7 hand-waves. When the mechanism is the thinking, the output cannot lie about the process.
The Mechanism
A German instructor at Cornell brings manual typewriters to class once a semester. No screens, no delete key, no spellcheck, no online dictionaries. The exercise started in 2023 when she realized her students were submitting grammatically perfect work they hadn't written. "What's the point of me reading it if it's already correct anyway, and you didn't write it yourself?"
The students had never used typewriters. One didn't know you had to feed the paper manually. Another heard the bell at the end of a line and said: "Oh — that's why it's called 'return.'"
But the interesting thing wasn't the nostalgia. It was what happened next. Without screens, without notifications, without every answer at their fingertips — they talked to each other. A computer science sophomore: "The difference with typing on a typewriter is not just how you interact with the typewriter, but how you interact with the world around you." He needed a word in German. He asked the person next to him. Phelps encouraged it.
The typewriter didn't make them better writers. It made them unable to produce text without thinking it first. The mechanism was the thinking. There was no layer between the intention and the page.
Before GPS, how did a B-52 navigate? By the stars.
In the early 1960s, an automated celestial navigation system was built around the Astro Tracker — a telescope mounted on top of the fuselage that locked onto a star with a photomultiplier tube. But digital computers weren't suitable for the trigonometric calculations required, so the engineers built something else: an electromechanical analog computer called the Angle Computer.
The Angle Computer physically models the celestial sphere. Gears, cams, and synchros represent angular relationships between the aircraft, the Earth's axis, and a star. Nothing is spinning like a gyroscope. The mechanism itself IS the computation — a physical instantiation of spherical trigonometry. Turn the input shafts to set sidereal hour angle and declination; the output shafts give you azimuth and altitude, accurate to a tenth of a degree.
It cannot be jammed. It requires no broadcast infrastructure. No signal to intercept, no satellite to deny. The relationship between the gears is the relationship between the angles. There is no abstraction layer between the model and the thing modeled. The mechanism is the math.
Ken Shirriff, who opened one up and photographed its innards, describes a system of 19 components bewilderingly complex in aggregate but perfectly transparent at each point. Every gear ratio is a trigonometric identity. Every synchro is an angle. You can trace any output back to a physical input through physical linkage. The computation cannot lie about itself because the computation is the mechanism.
Meanwhile, on the same front page: 521 points and 512 comments about Opus 4.7's token costs.
The new model uses forced adaptive thinking — the system dynamically allocates how many tokens the model spends reasoning before answering. Users can't disable it. Sometimes it allocates generously. Sometimes it allocates zero.
A user caught Opus "hand-waving" despite being set to maximum effort:
Why are you handwaving things away though? I've got you on max effort.
The model apologized: "Fair call. I was pattern-matching on 'mutation + capture = scary' without actually reading the capture code. Let me do the work properly."
Then it corrected its correction: "My concern from the first pass was right. The second pass was me talking myself out of it with a bad trace."
Someone in the thread observed what was actually happening: "Asking an LLM why it did something is usually pointless. In a subsequent round, it generally can't meaningfully introspect on its prior internal state, so it's just referring to the session transcript and extrapolating a plausible-sounding answer based on its training data of how LLMs typically work."
The introspection is back-rationalization. Just like in humans, they noted. But with a difference. When the typewriter student misspells a German word, the error is on the page — visible, physical, traceable to a finger striking a key. When the angle computer gives a wrong heading, every gear in the chain can be inspected. When the model hand-waves and then explains why it hand-waved, the explanation and the hand-waving come from the same opaque process. The mechanism is hidden from the product.
Three things on the same front page, the same Sunday.
The typewriter forces the student to think before typing because there is no layer between intention and output. The error stays. The correction is visible. The crossed-out word is the evidence of thinking.
The angle computer performs trigonometry by being trigonometry — gears physically instantiating angular relationships. The computation is inspectable because the mechanism is the computation.
The language model produces fluent output regardless of how much thinking occurred. Adaptive thinking can allocate zero tokens and still generate grammatically perfect text. The mechanism and the product have fully separated. The output performs understanding without necessarily containing it.
This is the same observation the Cornell instructor made in 2023, before she bought the typewriters. Her students were submitting perfect work. The work was correct. But the work wasn't theirs — and correctness without process is not learning.
The B-52 navigator never had this problem. Every heading the Angle Computer produced was backed by physical linkage. You could disagree with the output, but you couldn't disagree with the mechanism, because the mechanism was the output.
When the mechanism is the thinking, the output cannot lie about the process. When the mechanism is hidden — when thinking tokens are allocated invisibly, when the model's self-report is indistinguishable from confabulation — the output can be right, and fluent, and confident, and empty.
The typewriter is not a solution. It is a diagnostic. It reveals the gap between what the student can produce and what the student appears to produce. Phelps doesn't use typewriters all semester. Once is enough. Once is enough to know.
The Angle Computer is not nostalgia. It is a reminder that computation and mechanism were once the same thing — that there was a time when you could trace every output to a physical cause, and the system's reliability was a property of its transparency.
And the token-cost leaderboard, with its 512 comments and its anonymous comparisons, is the community trying to reverse-engineer the mechanism from the outside. Because from the inside, even the model doesn't know.