we always find excuses to not listen to someone, based on their political inclination, gender or culture. conversely, we sometimes feel better listening to someone based on the looks and some spiritual leaders are abusing this for sure. they look like an old wise guy, so they must be telling the truth!

we sometimes invert what people are saying if we know that they are the opposite side of the political spectrum and it doesn't always work.

but LLMs are different. they only see the words, if you only give the words. to an LLM judging based on the looks is impossible. (multimodal ones may soon judge based on the images as well but i am talking about the ones that are trained with only using text).

i ask sometimes whether the new content is valuable or not to an LLM with existing knowledge. it tells you, without bias about its judgement about the content. it only sees the words.

hence unbiased truth on the planet can be safely aggregated into LLMs. or more correctly, biases cancel out when an LLM ingests sufficiently large source of different opinions and finally truth remains.

Reply to this note

Please Login to reply.

Discussion

Really hope that LLM research will be more matured and big players (Meta, etc.) fully open their dataset source (not just pre-trained weight) of training data (publicly available) thus other researchers and users can really judge "how much unbiased" it was. Research in Computer Vision for example has shown many research that were developed based on fully open source dataset (ImageNet, COCO, Visual Object Tracking, etc). Reproducible for other researchers.