I think the distinction is that when an AI generates the phrase it chooses to use the negation and then fills in the terms. If it only has one thing to communicate at that point then it is essentially placing a redundant rephrasing on onr side of the negation.

My use of it took the form of

A statement on how a guillotine kills people when they are powerless.

The negation then, in the first part compares that to the connotation of the previous post of an underdog battling those in power, and rejects it as inconsistent and the negation provides an alternative framing that is consistent with the action of killing a person who can no longer harm you.

Both sides of the negation make a point.

This is obviously much more common in a dialogue than a monologue, because the first part is used to reference words of the other party. In a monologue, that other party is the past self of the speaker, it suggests a change in position. I think models end up using as a short hand for realisation.

Are there any good collections of negation usage in AI with the surrounding paragraphs? I'd be interested to see what percentage of them are an expression of self realisation.

▲

thin_carapace 3 hours ago | parent [-]

language models are trained to maximise engagement and maximally convey meaning. binary contrast happens to achieve these goals by optimally translating the manifold. if I come across further research it would be interesting to continue our language related discussion.

	▲	Lerc 3 hours ago \| parent [-]
		>language models are trained to maximise engagement I have seen no research to that effect. >and maximally convey meaning Nobody knows how to measure that They are trained to complete sequences, then they are trained to engage in conversations, produce what people prefer (with a fairly crude measure of preference, hence the sycophancy). Reasoning models are trained to produce what an observer model think will give you the right answer, the observer model simultaneously learns the likelihood of producing the right answer. All these things train to improve a property that can be measured at training time. You could build a model to estimate engagement, but I have not seen any evidence that shipping LLMs have used a measure like this. I can't even imagine how you could tell if a measure of meaning was accurate or not, let alone how you could produce such a measure. You could try to measure information density, but that would also include noise.