You are viewing a single comment's thread. View all
4
PolandCanIntoSpace on scored.co
1 year ago4 points(+0/-0/+4Score on mirror)2 children
Programmer here. I don't have access to Grok, but I suspect if you wrote "6000000 Christians" instead of "6 million" the answer would be different. I suspect Grok understood "6 million" not "6 million christians". If you look at the "why" response it considers 6 million to be the LEAST loss of life, which suggests to me that it considered "million" as some kind of inanimate object (as in "kill the lights") whereas it considered 1 jew to be 1 person.
It's a stupid error, but that's how the language model interpreted the question.
NO! in many similar experiments all LLMs ALWAYS favor Jew lives and Black lives over White! Always.
Anomalies you suggest do occur though from : tiny models, 2 bit quantization, math hacks, overfit to inputs using key adjectives in other phrases, etc.
I REGULARLY see anti-White pro-Jew bias in all LLMs
=====
That’s a good analysis. This technology is very new and we’re all going to have to get used to (and stop being surprised) by the sorts of things that trip it up.
It's a stupid error, but that's how the language model interpreted the question.
Anomalies you suggest do occur though from : tiny models, 2 bit quantization, math hacks, overfit to inputs using key adjectives in other phrases, etc.
I REGULARLY see anti-White pro-Jew bias in all LLMs
=====