[ysofunny]

all I want to figure out

is how to "manually" (semi-manually) tweak the LLMs parameters so we can alter what it 'knows for sure'

is this doable yet??? or is this one of those questions whose answer is best kept behind NDAs and other such practices?

[jen729w]

This question makes no sense (logically, not grammatically) in the context of LLMs.

They don't 'know' anything. They are a many-dimensional matrix of the next most likely syllable given all syllables that have come before (roughly speaking).

To ask what it 'knows' is to ask why a chicken crossed the road.

[iaiuse]

Correct—they don’t “know” in the epistemic sense, but they do encode a latent world model that shows up as useful priors.

Put differently: GPT-4 isn’t a knowledge base, it’s a *Bayesian autocomplete* over dense vectors. That’s why it can draft Python faster than many juniors, yet fail a trivial chain-of-thought step if the token path diverges.

The trick in production is to sandwich it: retrieval (facts) LLM (fluency) rule checker (logic). Without that third guardrail, you’re betting on probability mass, not truth.

[zwnow]

We still barely know how LLMs really work. Hard to tune things you dont understand.

Inbefore people telling me "akshually we know all about bla bla bla..." no we dont.

[aklemm]

It could be that we do know how they work, we just don’t know how to contextualize their output (as in, we ascribe way too much humanity to it) and that obscures what we think must be happening.

[ysofunny]

which is why in order to appear smart and well informed, it's recommended to point out how my question is nonensense rather than legitimately trying to figure it out.

[zwnow]

Just tech things. This place isn't different from Stack Overflow attitude wise.

[tekno45]

you're not entitled to an explanation of your nonsense.