Yep, same way AlphaGo Zero learned to play. It evolved, generation by generation, by selection. People think an agent can't solve a problem unless it has full understanding, but that's not true, selection based on fitness can do it.
Cue fun time with etymology: understand translated to Latin and back yields support, which is significant, kinda, because support vectors are one model of machine intelligence, while DNA is basically a vector, in terms of transcription, or a matrix even, looking closer at the base molecules.
disclaimer: I have n o idea about support vectors, to be honest
Support Vector Machines (SVMs) are a classification technique that was in vogue before 2012, with great results for smallish problems. It's still useful today, but less talked about, I mean, it's just assumed as common knowledge.
i'm wondering if 15-25 generations (that's approximately how many generations of garden birds were there since the invention of the cigarettes with filters) are enough for evolution to actually show this trend here
Tangentially related - there was the russian experiment on fox domestication that had decent success in selecting for "tameability" with just 3 generations.
I think some birds are actually pretty smart and can learn by imitating other birds, so it's a cultural thing. All it takes is one bird to discover it.
Take a look at this video: "Crow uses a vending machine". It was an experiment where they trained crows to collect coins and put them into a vending machine for a few grains.
