From random samples, the model successfully rediscovers the classic Big Mac®, both in correct ingredients and weights, although the Big Mac® was never part of the initial training data (Fig. 2a). Across ten independent randomizations, rediscovering the Big Mac® requires on average 7.3 million samples which demonstrates that exact replication of recipes is a low-probability event under the learned distribution (Fig. 2e).

my siblings in christ that's not recreating the big mac, that's keep hitting your head against the wall untill you start to bleed lmao. i don't have anything specific to say against the paper (because i can't be bothered to look deeper) but this really specific and horrible to test benchmark is way too funny to not dunk on.