LLMs never took off until they combined them with RL via RLHF. RLHF was discovered in their RL research on game playing. GPT3 was out for quite a while with much lower impact than the chatgpt release and finished training in like december 2019 I read somewhere, released mid 2020. There were later better checkpoints, but it still didn't have much impact except for code completion.
With just a raw language model instructions and chat didn't work to near the same degree.
Both elements are important and they were early in both. Illya's first email here talks about needing progress on language:
2016
Musk: Frankly, what surprises me is that the AI community is taking this long to figure out concepts. It doesn’t sound super hard. High-level linking of a large number of deep nets sounds like the right approach or at least a key part of the right approach.
Illya: It is not the case that once we solve “concepts,” we get AI. Other problems that will have to be solved include unsupervised learning, transfer learning, and lifetime learning. We’re also doing pretty badly with language right now.
With just a raw language model instructions and chat didn't work to near the same degree.
Both elements are important and they were early in both. Illya's first email here talks about needing progress on language:
2016
Musk: Frankly, what surprises me is that the AI community is taking this long to figure out concepts. It doesn’t sound super hard. High-level linking of a large number of deep nets sounds like the right approach or at least a key part of the right approach.
Illya: It is not the case that once we solve “concepts,” we get AI. Other problems that will have to be solved include unsupervised learning, transfer learning, and lifetime learning. We’re also doing pretty badly with language right now.