Looking at the paper, as I understand it they're using Mineflayer https://github.com/PrismarineJS/mineflayer and passing parts of the state of the game as JSON to the LLM that are used for code generation to complete tasks.
> I still don't understand it and it blows my mind - how such properties emerge just from compressing the task of next word prediction.
The Mineflayer library is very popular, so all the relevant tasks are likely already extant in the training data.
> I still don't understand it and it blows my mind - how such properties emerge just from compressing the task of next word prediction.
The Mineflayer library is very popular, so all the relevant tasks are likely already extant in the training data.