LLM's runs off a reward system based on its output - this is all...

LLM's runs off a reward system based on its output - this is all embedded on its parameters and weights bias configuration. So of course it can produce its own data based on the reward system. But needs alot of training to make it commercial grade.

Big LLM are initially trained on 10TB of data from wikipedia, news sources, websites and whatever else. But they all have a time stamp - Bard is only trained to October 2023 for example.

If I ask for an update on BHP the data is outdated.

If I ask the same question 20 times I will get 10 different answers. Meaning Domain data is a complete and utter mess. But for sure LLM's do learn on different data types. But a long way to go to replace Web Search for example.

In other words - If I ask an LLM for three different economic outlooks - it all sounds vey nice but the data is useless.