This week
- Bench
- 12 May 2027 Try the method on something fluent reconstructing, and failing
- Slides
- Open the deck
- Due
- Specimen log 14 May 2027
For ten weeks the argument has been that an error is evidence, and that reading it backwards gets you to a machine you were never given access to. Every step of that depended on the error being legible.
A general-purpose model’s characteristic failure is not a broken sentence. It is a fluent one that is wrong. There is no join to find, no entry to reconstruct, no register borrowed from committee minutes. The surface carries no information about what produced it.
So the method this course teaches runs out of material, and it is worth being precise about what that costs. The public was never given access to these systems. What it had instead was the error: a small, free, involuntary disclosure that arrived on every sign and every menu. Closing the seam did not make the translations honest. It made them unauditable by anyone standing outside.
Outline#
- the error that reads correctly, and why it implicates nothing
- what replaced the seam: evaluation sets, disclosure, and who holds them
- the argument this course has been making, stated plainly
- what you would need in order to keep doing this, and who has it
Teaching team
- Ines Barrantes — convenor
- Aurel Marchetti — tutor