I’m kind of surprised people are more concerned with the output quality for chatGPT, and not where they source their training set from, like for image models.

Language models are still in a stage where they aren’t really a product by themselves, they really need to be cajoled into becoming a good product, like looking up context via a traditional search and feeding it to the model, or guiding it towards solving problems. That’s more of a traditional software problem that leverages large language models.

Even the amount of engineering to go from text prediction model trained on a bunch of articles to something that infers you should put an answer after a question is a lot of work.