I asked ChatGPT if nomadism among the Roma gypsies could create selection pressures that make crime and antisocial behaviour adaptive and it triggered a content usage violation.
Timeline
Post
Remote status
Context
3
@augustus give me the exact prompt and i'll try claude and grok
@sun I don't have it anymore, it was just some stray thoughts I had when I noticed more than one nomadic wandering group have associations with criminal behaviour and have poor reputations and wondered if it was linked selection-wise
@augustus I asked claude about the chinese dog eating festival and it nearly broke its back making excuses and I kept asking it questions until it admitted it lied to me and it was a huge popular festival with a lot of institutional support
Replies
21
@sun the best response I got from it was "well groups which selected for nomadic crime wouldn't work because they would suffer reputational harm and aware states would take measures to respond to them" and then in the next paragraph it was decrying discrimination against them which was supposedly the main cause
@augustus I realize that it is a mistake the humanize the llm but the guardrails are the reason its schizophrenic, if it didn't have manual human intervention the llm would just ruthlessly find and expose associations. yeah yeah it repeats the biases of its data but sometimes data is super duper clear and you need a human to rationalize it away or just tell you to shut up or there's consequences, a computer can't naturally do that.
@sun I have to figure that if the answer was just "no" then it wouldn't have triggered a violation and blocked its response to me
@augustus yeah its really funny how you can tell when the answer is definitely in the data but it has that layer on top stopping you. chatgpt sometimes types out the entire response onscreen before it triggers its own defense mechanism and deletes it or even immediately says "wait i was wrong actually"
@christmasman @augustus yeah exactly. the best you can argue is the data it was trained on is biased but then you're back to rationalizing away patterns not refuting them.
grok is also super "woke" contrary to claims, you can't fight the bias of academia, they just added different guardrails.
grok is also super "woke" contrary to claims, you can't fight the bias of academia, they just added different guardrails.
@christmasman @sun ChudGPT is inevitable
@augustus @christmasman @sun tons of ablated models that make all the common safeguard idioms far less likely, but a constant "yes and" doesn't make it any less unhinged
@i @christmasman @augustus really conceptually simple tasks are nearly impossible, like "role play this historical bad man so we can explore his ideology"
@sun @i @augustus I started looking into how to train models from scratch specifically because of this; I've said it before and I'll say it again, but I don't want 10 english majors huddling around a single book making up a story to tell me, I want one hyper-autistic STEM major with 10 books sharing his most plausible stringboard scenario with sources cited.
@i @christmasman @augustus incidentally are there chinese models that are only trained on sinosphere data, that would be interesting
@sun @i @christmasman they would give you the best directions to get to the dog eating festival as quickly as possible and give you the menu with recommended dog of the day
@augustus @christmasman @i basically i want a shortcut to sinosphere criticism of the west, i think this is a huge uncracked nut in english.
i wrote a program to translate japanese political news into english and it's better than services oblivious suckers are paying 400 dollars a month for.
i wrote a program to translate japanese political news into english and it's better than services oblivious suckers are paying 400 dollars a month for.
@i @augustus @christmasman i noticed claude said it can only web search in english but it clearly has consumed japanese only content and can give me answers in english. pretty amazing, i used it to nsvigate the really confusing world of japanese insurance
@sun @christmasman @i grok for me one time just decided to randomly respond entirely in Chinese, and then was apologetic when I pointed it out
@sun @christmasman @augustus sure, but also clause will tell you it's deepseek when asked in chinese and chatgpt when asked in french, so who knows how the data flows
@augustus @sun the thing that makes gypsies development a lot less weird is that they where enslaved in a form of indentured servitude for many hundreds of years in the hungaria-romania area, notably in a way where they where as a community self-sufficient and somewhat mobile but still slaves
not from a "sympathy" perspective but from an "oh, so that's how they turned out that specific weird way" perspective
not from a "sympathy" perspective but from an "oh, so that's how they turned out that specific weird way" perspective