Theme: Measurement

-“Believable outputs are not believable because their contents are true of factual. They are believable because they emulate grammar”–

Trying to science that:

The contents of speech or written speech is perceived as non-predictive (false) because of not surviving tests of consistency (inconsistency of), or possibly-true because of survival of consistency of: predictions from identity, consistency, possibility, correspondence, rationality, reciprocity, or coherence against existing episodic memory.

The contents of speech or written speech are perceived as:
False, because the sequence of phonemes or words produces continuous recursive disambiguation insufficient for survival (as above)
Or;
Possibly true, because the sequence of phonemes or words produces continuous recursive disambiguation sufficient for survival (as above).

The brain predicts consistency. That’s all it does. From the cells at the back of our eyes, to the visual cortex, through disambiguation by the dorsal and lateral cortex, into the hippocampal system of episodic formation, to the auto-association of prior episodes, and the competitive sequences of those episodes, to the most concentrated effort of the frontal cortex to direct recursive searching, stacking the process as open pathways, and then directing attention to competing pathways, until we succeed or fail and start over at some point again.

So translating (sciencing):

-“They are believable because they emulate grammar”-

I can guess, is attempting to mean, that grammar( rules of continuous recursive disambiguation in a language and its paradigms, vocabulary and logic) …. what?

In the context of LLM’s that MIGHT mean that the marginal difference between sentence compositions will tend toward marginally indifferent framing. Those marginally indifferent framing that tend to converge together function as an adversarial system of competition for identity, consistency, coherence, etc?

Assuming the majority of marginally indifferent sentence compositions produce a probability distribution, then the marginally DIFFERENT sentence compositions contain associations that provide a domain of relations where we might investigate error (and tune the results).

This means the human composers of the source speech and text have determined a minimum disambiguation necessary for unambiguous communication at least within a given context.

This doesn’t tell us that the humans are lying, and often the majority are lying, or otherwise engaged in ignorance, error, bias, and deciet. Which is why the source data is limited.

So AFAIK, by the ‘gammar’ doesn’t tell us much.

-FIN-