Ask HN: How do companies make their AIs "woke"?

[litetime]

It's a common quip to say that LLMs are black boxes and that we have no idea how they "think". Of course that was always an exaggeration, but there is some truth to in that answers are generated from billions of miniscule connections between entities that are nearly impossible to reason about. How exactly then are they able to turn up and down the woke slider for their models? It seems impossible to tune them manually for a near infinite variety of use cases. But I could be wrong. Anyone have an intuitive explanation of how they get it done?

Edit: I wasn't trying to use the word "woke" to make any sort of political statement, although I see now how maybe it wasn't the best choice of words. The impetus for this post is the recent controversy around Google's Gemini being, uhhh, "overly tuned?".

[proc0]

For the most part it's probably reinforcement learning from human feedback (RLHF). This incorporates humans in the training loop and its done for alignment purposes (which is overall a good idea, but it does depend on who exactly the AI is aligning with). There may also be other areas where human bias can seep in, like the massaging of the training data, but more likely the biggest factor is the direct feedback training done by a select number of people.

https://aws.amazon.com/what-is/reinforcement-learning-from-h...

[nonrandomstring]

"woke" isn't a great qualifier. But you could re-frame the question as how do "AI" models encode and search within a political sentiment space? Here's some political spaces, usually in 4, 6, 9 and 12 dimensional variants [0..2] There are also personality axes, again with high or low dimensional nuance. Unlike "real" signals there's no component analysis to prove that these are orthogonal. If the training data carries-in any salient features you can get an NN to tell you where "woke" or "fascist" or whatever is within these for some task, then minimise or maximise it for some quality.

[0] https://www.thebehavioralscientist.com/glossary/political-co...

[1] https://politicaltests.github.io/12axes/

[2] https://9axes.github.io/

[3] https://en.wikipedia.org/wiki/Personality

[bell-cot]

I kinda suspect they're over-cooking the training data.

But with the ability of (so-called) AI's to hallucinate fictitious legal cases and such - pretty undesirable behaviors from either end of the political spectrum - I would not rule out "AI's are just too stupid to know better".

[EchoChamberMan]

Google tried to use prompts to force specific output. the companies use prompts, weights, and humans to try to keep the output sane.

[allears]

Please define 'woke.' As far as I know, it's simply an insult used by Republicans against Democrats.

[litetime]

Sorry I don't mean it as insult or to stir up any political arguments. I guess I could have said "personality", but I was motivated to understand this from the Google Gemini controversy.

[mtmail]

There are attempts to define it https://simple.wikipedia.org/wiki/Woke but I don't understand what the relevance to AI is.

[michaelmrose]

It started out having a specific positive meaning of being aware of social injustice and has in the hands of conservatives transitioned to meaning virtually anything I don't like that could even vaguely be attached however tangentially politically to anything left of Hitler or involving anyone except a white conservative man or his christian conservative wife and obedient white children.

If your minority waitress brings you burned toast it might be woke.