I personally do not want the companies to release training data (at least for a while) because then it gives people leverage to neuter it. I don't want a sanitized LLM, and I don't have $60M lying around to train my own. Copyrighted material, sexual content, political opinions, throw it all in and release it please! Yes, reducing bias in the models is a noble goal, but introducing new bias and blindspots to do it is…
We really need to somehow separate a bias towards accuracy as distinct from some bias towards say, a sports team. Using everything would be like taking a bunch of students final exams and then claiming the most common answers are the correct ones. This isn't how expertise and accuracy works. Most things worth doing are not only genuinely hard and complicated but something that only a minority subset of accomplished p…
As we have seen from history, there is not often an absolute truth to questions, only clusters of truths. We want our LLM to be able to perform reasoning, mathematics, and science, but expecting absolute truths in anything outside of those fields is a bit much. Wikipedia often takes a good approach here and strikes this balance well. You can represent the "crazies" and show their reasoning, and then draw attention to how it is commonly refuted. You don't get this if you just omit the "crazies" to begin with.
Edit: What I'm saying is that you can include/represent these adjacent clusters of answers instead of asserting one cluster to be true and allow the correct answer to be represented organically. And guess what? Some people will still disagree for whatever reason they have to disagree. We should be taking all this AI alignment money and putting it into education imo.