Updating my priors - I've been assuming the bottleneck in LLM development is the training process itself, but this headline suggests that human annotation is a bigger barrier to progress than I thought. https://www.reddit.com/user/Neil-Sharma