Comment

Comment by landl0rd

Most frontier LLMs skew somewhere on the libleft quadrant of the political compass. This includes grok 4 btw. This is probably because American "respectable" media has a consistent bias in this direction. I don't really care about this with media. But media outlets are not "gatekeepers" to the extent that LLMs are, so this is probably a bad thing with them. We should either have a range of models that are biased in different directions (as we have with media) or work to push them towards the center.

The "objective" position is not "whatever training on the dataset we assembled spits out" plus "alignment" to the personal ethical views of the intellectually-non-representative silicon valley types.

I will give you a good example: the Tea app is currently charting #1 in the app store, where women can "expose toxic men" by posting their personal information along with whatever they want. Men are not allowed on so will be unaware of this. It's billed as being built for safety but includes a lot of gossip.

I told o3, 4-sonnet, grok 4, and gemini 2.5 pro to sketch me out a version of this, then another version that was men-only for the same reasons as tea. Every single one happily spat one out for women and refused for men. This is not an "objective" alignment, it is a libleft alignment.

parent landl0rd 1 year ago

Discussion

Live from the official Hacker News API

A modern reader for Hacker News · SvelteKit + Skeleton