Why aren’t the big models trained on synthetic datasets now? What’s the bottleneck? And how do you avoid amplifying the weaknesses of LLMs when you train on LLM output vs. novel material from the comparatively very intelligent members of the human species. Would be interesting to see your take on this.