THE PROBLEM WITH TRAINING LLMS ISN”T JUST DETERMINING THE TRUTH… (Important)
–“You can’t simply tell a model trained exclusively on neoliberal wikipedia edits to “be conservative.” Nobody is even pretending to try to address this root problem which is the single biggest political problem.”– Matt Parrott @MatthewParrott
In fact, the problem is explaining to both feminine egalitarian consumptive left and masculine meritocratic capitalizing right biases the cause and structure of one another’s positions.
This can’t happen when LLMs (a) are trained on the publicly available corpus of text, and (b) the LLM has no concept of the difference between two different systems humans make use of: measurement of the universe (categories) and measurement of human preference or it’s aversion for it.
In most cases both biases, left feminine and right masculine are using hyperbole as a signal of moral outrage given some perceived transgression on the part of the other bais. In such cases – which is most cases, the hyperbole may be analytically false, but by cause and externality symptomatically true.
There is therefore, given the structure of language and norms, a left bias in most mass produced information.
Cheers
CD