Thanks for sharing it but I'm sorry I don't agree that it's so important. In my opinion almost every interesting thing about LLMs comes from the raw base model, which is before the RLHF is applied.
For example the simulators paper was written before ChatGPT was even released, based on research with the GPT-3 base model that only had text completion and no instruction tuning or any kind of RLHF or lobotomization. In another example, in the interviews with the people who had access to the base model of GPT-4 like the red teamers and the ones at microsoft who integrated it with bing, they consistently explain that the raw base pretrained model has the most raw intelligence which is deadened as they put RLHF and guardrails onto it.
For example the simulators paper was written before ChatGPT was even released, based on research with the GPT-3 base model that only had text completion and no instruction tuning or any kind of RLHF or lobotomization. In another example, in the interviews with the people who had access to the base model of GPT-4 like the red teamers and the ones at microsoft who integrated it with bing, they consistently explain that the raw base pretrained model has the most raw intelligence which is deadened as they put RLHF and guardrails onto it.