“Chatbot trained on conversations of real people acts like real people” Sounds like a very successful project.
Well, cynicism aside, the difficulty in creating language models etc. does not seem to lay in getting enough data anymore, but rather in getting the right data, together with identifying what is wrong.
Yup it’s a common sentiment since the GPT era. Trash in = trash out still applies and we should be focusing our efforts on collecting quality data. Unfortunately that’s not what funders are interested in. Grants generally just go to people who promise better metrics.
Don’t train models on user data without curation, folks!
Garbage in, garbage out. This is more a reflection on the devs and the training data they curated than anything else
