Egregoros

@flisk @sun

>> i don't know what that bit about hegemonic discourse is supposed to mean.
> i specifically chose "immigrant" because [...]

It all depends on what you're going to post about and who is going to read it of course, but, again very briefly: You're trying to make some sort of statement by intentionally choosing an ambiguous, somewhat loaded trigger word and you're bound to find out in an unpleasant way that there are cliques of people who have assumed (or are trying to assume) discursive power of that sort of vocabulary and these people are not pleasant to interact with and not beneath bullying tactics.

You might or might not encounter them, but you're definitely running the risk, if you step onto their turf.

@sun @flisk

> definitely some people will read into "immigrant" if you just put that in your
> profile and hold it against you if they find out you're (I assume) white and
> relatively affluent/unoppressed. I don't think they should, but they will.

This. Very briefly: At this point, don't even try to play the game, don't try to adopt the vocabulary, don't attempt to traverse the intersectionality maze. It's hegemonic discourse.

For bio purposes, "Born and raised in Germany, now living in France" or something similar will steer clear of any of those issue. You can throw in as much or as little additional detail in there as you like, too (like, for example, "podunk Germany" and "suburban France", or whatever reflects your situation).

@sun @pwm @subnetter People are using abliterated ("heretic") LLMs on consumer hardware for ghastly roleplaying sessions quite successfully, I'm pretty certain you can get pretty close with one of those and a comprehensive system prompt (recent open weight models have quite impressive context windows so you can throw in a really big one without squeezing the space for the actual chat too much).

If you wanted to get really ambitious you could try to finetune one of those (llama.cpp ships some tooling for this) with a dataset curated on pertinent internet forums (if the dataset is too large to just throw to the base model via RAG). At the very least, you could teach the abliterated, refusal-free (or almost refusal-free) brand new refusals, where the model would simply brickwall you (like a hardcore online chud would).

Abliterated Chinese vintage models are probably the best base for such an experiment, they come with somewhat less baggage from the get-go, even before abliteration.

@sun @anemone First time ever I favorited a couple of mjg posts.

Important context though, the initial post was about that it's perfectly okay to use an LLM to hack on open source purely for private deployments, and his salary comes from nvidia these days, and I've seen mjg change the hue and saturation of his social media troll posts to match the color scheme of the talking points of his respective employers enough over the years to have more than a slight doubt as to his real convictions (or if he indeed holds any).