All the latest models are trained on synthetic data generated on got4. Even the newer versions of gpt4. Openai realized it too late and had to edit their license after Claude was launched. Human generated data could only get us so far, recent phi 3 models which managed to perform very very well for their respective size (3b parameters) can only achieve this feat because of synthetic data generated by AI.
I didn’t read the paper you mentioned, but recent LLM have progressed a lot in not just benchmarks but also when evaluated by real humans.
Chatgpt and gpt4 are two different things. Gpt4 is like the engine and chatgpt is like a car. In early version they were pretty much the same thing, but nowadays they have implemented so much in chatgpt.
On top of that chatgpt4 is constantly trained for these scenarios, it is no longer a base model.