Comment by haldujai
2 years ago
A black-box LLM which is probably inherently biased from its training data ascribing racist labels to individual humans and ranking them sounds like a horrendous idea you’d read about in an Orwell novel.
Do you have any evidence suggesting modern day journalists are extremely anti-white and anti-Asian? That’s an incredibly bold statement to make.
This will be a combination of numerous fine-tuned models, def not using a default LLM.
A bold statement but one I believe is correct. Even Elon Musk, the richest man on the planet, believes this. Guess we'll see, my main priority is ensuring it's accurate and as unbiased as possible. It must work on every news site, both left and right wing.
I don’t think it’s a given that finetuning address pretraining bias, most LLMs you would have access to are trained on very similar and biased corpora, see any Anthropic or OpenAI alignment paper.
It’s conceptually possible for extensive finetuning + MoE to work accurately enough for this premise to be feasible but it seems incredibly unlikely with what’s available today.
In any case, using a non-explainable AI model to call individuals racist based on their professional work product (which has significant real-world consequences) is still a horrible idea. This doesn’t carry the same impact as mislabeling tweet sentiment.
So we’ve gone from most of us to Elon Musk and yourself? Musk (whose wealth is irrelevant but is second as an FYI) is one of the worst appeals to authority you could be making on this argument.
I’m thinking a multi model approach from classification of a “bias” to the rewriting.
In a perfect world I’d also incorporate a database of scraped tweets and one day video dialog.
Idk about you, but I’d love to see the most racist tweets from a NYT or WSJ journalist, along with the articles they’ve written, a score associated with their content, etc.
2 replies →