Skip to main content

BENCHBENCH · WEEK 5 · 24 March 2027

Week 5 Bench: Starve it

Feeding a neural system input it has nothing for, and recording what it produces instead of nothing

This week

Lecture
22 March 2027 The network, and the confident wrong answer
Slides
Open the deck
Specimens

I am not in the office at the moment. Send any work to be translated.

The Swansea road sign primary

Bench

Material
input with no content
Operation
probing a closed system
Leaves
a dated log of what a system invents

Today the room runs the experiment rather than reading the artefact.

The protocol

  1. Pick a language pair where one side has very little parallel text.
  2. Feed it input with no content: repeated syllables, a string of the same character, a word with a typo in the middle.
  3. Record every input, every output, the system, the direction, and the date. Screenshot the date.
  4. Repeat one input four times over the session and see whether the answer moves.

What to watch for

The interesting result is not a wrong translation. It is a well-formed sentence about a subject that was not in the input, arriving with no hedging and no error message.

Afterwards

Your log is a dated artefact about a system that will have changed by the time anyone reads it. That is why step 3 is in the protocol and not a suggestion. Six months from now the log is the only evidence that this behaviour existed.

The spec

What you bring to the session. Written so a reader can tell whether it has been met without asking you.

Teaching team