The Challenges of Crowdsourcing Human Instruction Data

Crowdsourcing is one of the biggest missing resources if we want to have real true GPT for even chat GPT three quality models. I don't know whether it's possible to crowdsource that, or whether at some point you just have to spend a lot of money to get high quality data. As we've learned lately, it's quality and not quantity that really matters here. And getting good human instruction data is a really hard thing to do.