Science

Including their own output means including their "hallucinations". But they also use other LLM AIs.

The weights are the amount of impact that a given artificial neuron (out of millions or hundreds of millions) has on the next neurons in the connection sequence from input to output. Training is an iterative process of homing in on those weights. It takes millions of runs and a lot of time and a lot of energy (enough that big players are looking to locate beside nuclear power plants).