Rant HN: Dialectics, and how supervision should ask "Why" not show examples
* a: discussion and reasoning by dialogue as a method of intellectual investigation specifically : the Socratic techniques of exposing false beliefs and eliciting truth
b: the Platonic investigation of the eternal ideas*
Yet we have other terms for that, which were widely used until the 20th Century. Socratic method. Logic. Platonic ideals. Induction versus deduction.
Supervised training is a kind of debate, but currently one that relies on pre-written dialectics.
Leaving aside philosophy, many theologies revolve around methods of learning through debate over the substance of written words. But it is the process of debate, not the absorprtion of its outcome, that is crucial to learning.
A "dialectic" differs from learning-through-debate in that it arrives at a conclusion previously, and then pre-arranges arguments to change people's minds. A written dialectic is not a live argument between living people; it's an example of an argument to be followed. As such, it's a work of fantasy that presents a straw-man and steel-manned argument along with their methods with an expected outcome which will not just persuade people, but persuade people how debates should go. A dialectic is a dishonest form of fiction, presenting a debate that never took place, to persuade people that no such debate ever needs to occur since the conclusion has already been arrived at. A written dialectic is not a way to make a point, it's a way to dismiss any further argument.
The reason I'm harping on this is to raise a philosophical question: Is it possible to have a Platonic model of self-supervision which doesn't include built-in bias? Who or what then - barring a full interface with "reality" - can hypervise? The shittiest thing about LLMs is that they are dialectical machines which steel-man whatever argument they were trained on. To an obnoxious fault.
LLMs are stateless. Every session is a naive model that only learns a few things through a series of prompts, up to its tiny buffer limit. Humans, on the other hand, are nothing but pure state with a small smattering of useful facts. Humans are notoriously incapable of remembering the same thing twice in the same way. Humans have an enormous life-long input/output buffer based on a model that has almost nothing.
When humans learn by arguing, it's not sufficient to absorb a previous dialectical argument. The entire method of learning is to overcome the method of argument one is faced with. This starts as soon as a human needs milk.
Yet LLMs are slanted toward utilizing pre-determined dialectics to prove their point.
To be "smart enough" to turn a useful argument into a "dialectic" is a great skill. It is to corrupt the usefulness of argumentation itself, by presenting a single Platonic dialog as an eternal truth and forcing its conclusion as a rote memorization tool. Rather than taking the perspective that the logic in the dialog is more important than the final dialectic, how can an LLM learn to debate fundamental principles? How do we get out of that trap? How do we teach a model how to think, if we can't?
I would argue that rejecting any pre-determined dialectic is a good starting point.
The way to do that is to ask "WHY" a whole lot of times; WHY why why forever; and especially to keep asking it when you are told to stop asking.