4 ms·
I feel like this is just what happens when you average out a large amount of the internet and train a system to pick the next word, and is not an issue inherent
by voidUpdate 2y ago
I feel like this is just what happens when you average out a large amount of the internet and train a system to pick the next word, and is not an issue inherent to LLMs. If you trained one exclusively on horrible, sexist data, you'd get a system that chooses horrible sexist words. If it was perfectly equal, it would pick gender norms essentially at random. Seeing as reddit was a source for a sizable proportion of GPT training data, and there are some parts of reddit that are less than savoury, it's understandable that in some situations, it can choose less than savoury results
- drexlspivey 2y agoMake no mistake, these biases have been explicitly put there in the RLHF stage (see the Gemini fiasco). They don’t come from the training data.
- t-writescode 2y agoThere is a very good chance ( I imagine ) that these biases are put in place explicitly because unrestrained internet scraping and paper reading produces *wildly* sexist and awful material. See: what happened to Microsoft's old chatbot when let loose on Twitter.
- rhdunn 2y agoIt looks more like the refinement of the model contained views prominent in Silicon Valley that skewed the model to be anti-male/pro-female -- such as the examples relating to abuse. All abuse is bad regardless of gender, so the model should have treated them the equally.