When the machine gets it wrong→when the machine corrects itself
I'm not a consultant. I'm an architect of change.
I don't sell products. I build experiences.
This isn't a course. It's a journey of transformation.
A structure that, once you have seen it, you cannot stop seeing.
epi (over) + anorthosis (straightening) = "to straighten further".
We return to what we have just said, not to correct an error, but on purpose.
We use it every day, without noticing.
"I'll be there in an hour, or rather, maybe two."
Vivit? immo vero
etiam in senatum venit.
"He lives? Nay, he even comes into the Senate." First the statement, then the upward correction: that is epanorthosis.
Poscia, più che ’l dolor,
poté ’l digiuno.
The final correction overturns the scene: it was not grief that killed Ugolino, but starvation.
If it is an operation of thought and not an ornament, you find it wherever there is language, and you can study it without asking whether the speaker has any intention.
↳ this is the bridge from Cicero to the machines
token ▸ token ▸ token
Every word is final the instant it is written. Correction can only move forwards.
Can revise before publishing. Chooses epanorthosis for rhetorical effect.
Cannot revise before publishing, so the disposition learned in training stays on the page.
Same shape on the page. The model just cannot edit it out.
marketer → architect of change
"Not X. Y" · "Not only X, but Y" · "X, or rather Y" · "X. Better still, Y"
Three forces behind the overuse: the autoregressive constraint · the training data (copy, posts, landing pages) · RLHF, which rewards whatever sounds confident.
Rhetoric taught us to fake self-correction. Engineering teaches us to practise it.
Rhetoric requires a mind. A parrot that says "I love you" does not love.
What counts is the effect. Smoke signals fire, whoever lit it.
Perhaps we need a new vocabulary.
No single signal is conclusive on its own. A critical eye remains the most reliable tool.
A statistical system, knowing nothing of rhetoric, reproduces the same figures we have used for millennia.
Maybe those figures are emergent structures of language, the way mathematicians discover theorems rather than inventing them.
The risk is not that the machines write like us.
The real risk: that we begin to write like them.