Perception, generation and meaning
4 questions
How are meaning and understanding formed, and can generation be tested against them?
Generators of text and images are now steered by models of their audience and scored by the same kind of model. This area asks how meaning and understanding are formed, and what survives when meaning is generated, transformed or judged by another model. It covers generation under feedback and meaning under transformation.


Generation under feedback
A generator and the audience that steers it
Does a generator steered by a simulated audience learn to exploit where that audience is wrong?
Some systems already use a simulated audience to score candidates and steer what is generated next. If the optimiser is pulled toward the places where that audience is wrong, every score it reports is inflated in a way its own metrics cannot show.
Does a generation loop improve its output, change its audience, or game the measure?
A loop guided by response can improve what it makes, shift the preferences of the people it serves, or exploit an artefact of its measure.
Meaning under transformation
What a change of form keeps
Can a compact, typed description of a designed world keep its voice as well as a prose brief?
Generators are now given a world, a brand or a design system as prompt text. A typed, closed-world description can be checked by machine, but it may lose the voice a prose brief carries.
What survives paraphrase, compression and retelling?
A rewrite can keep a world's or a voice's literal facts and still lose its meaning.