Indeed. I am absolutely sure their design process went: Mix a bunch of different stuff, teste it. Feed results of teste evaluation into computer, train ML model

Indeed. I am absolutely sure their design process went: Mix a bunch of different stuff, teste it. Feed results of teste evaluation into computer, train ML model to accurately pick out the mixture which was rated highest in evaluation for that sweet sweet ml pr-boost effect. If that's not the case then I am extremely curious for them to release an actual tech run through of how ML contributed to their process.