-: FOLLOW US :- @theinsaneapp
Researchers at IT University of Copenhagen have developed a new method for encoding and generating Super Mario Bros levels called MarioGPT.
-: FOLLOW US :- @theinsaneapp
MarioGPT is based on GPT-2, a large language model that is good at pattern recognition and replication.
-: FOLLOW US :- @theinsaneapp
MarioGPT can generate an infinite number of Mario levels that can be played every day.
-: FOLLOW US :- @theinsaneapp
GPT-2 was chosen because it is better suited for smaller datasets, although GPT-3 may be used in the future.
-: FOLLOW US :- @theinsaneapp
The researchers had to first represent the Mario levels as text so that the model could understand them.
-: FOLLOW US :- @theinsaneapp
The model then ingests the text and reproduces the patterns that correlate with features.
-: FOLLOW US :- @theinsaneapp
The output includes a playable path represented as lowercase x's.
-: FOLLOW US :- @theinsaneapp
A software agent was able to complete nine out of 10 of the levels generated by MarioGPT.
-: FOLLOW US :- @theinsaneapp
The new method has implications for the creation of other types of games.
-: FOLLOW US :- @theinsaneapp
The research demonstrates the potential for combining language models and gaming.