Post

I have just read this 🤦‍♂️: > Together with humans (who generally decide what counts as ‘reliable evidence’), LLMs currently form a powerful hybrid epistemic system that functionally resembles a type of neuro-symbolic architecture, typically operating under external supervision.

Isn't that just humans interacting with technology or something? What?

The thread (on email) started with the latest results obtained "with the help of an LLM". Then someone suggested that the LLM should be added as an author, or something like that. Then the excerpt above came.

Being able to analyze something with maths doesn't make it a formal system! A formal system is a set of axioms and rules that you can use to make logical deductions. LLMs aren't, because they don't really use "logic" most of the time, which is partly why they are so weird sometimes.

LLMs do use (derivative) probabilistic logics implemented on binary hardware which only ever is able to compute classical logics. This might seem weird as modern logics are formal systems lot of people dont know about and also human brains are incredibly bad at stochastics

Also as you cant implement actual probabilistic logic on normal hardware this leads to lot of tradeoffs and quirks which are logic breaking and therefore unpredictable. i.e. bugs.

Of course they use logic in the process, they're algorithms after all. But the end result is just a bunch of text, that, the way we interpret it, doesn't have to follow some specific set of axioms, or any formal logic at all.

LLMs aren't weird, they do exactly what they are designed to do, which is to predict the next character in a string

Plus Random variable to imitate life.

Right. This is exactly what I mean. They're not doing formal logic.