I think this would have been very easy to program. Just using oldskool game "ai" programming techniques, just some simple algorithms. No need for fancy ML or neural nets or anything like that.

Eg just off the top of my head

Theres less than 40 total positions (eg 9 positions * 4 orientations) you can place a block at any time (waaay less in the case of some blocks, eg only 9 for square one) (+ a couple more for edge cases where a position could be under another position)
It'd be super easy for the computer to look at every possible positioning of 2 blocks, way less than 1600 possibilities (current block + next block) and rate each one according to a few rubrics, then pick the highest scoring one.

If theyre using a modern computer it could run that algorithm thousands or millions of times times in 30ms.. if they're actually using the NES then they would need a less brute force approach lol.

Anyway I do not intend to take from away form this at all. Merely point out that this is a good example of seeing a computer program do something that simple computer programs are a good fit for, and how impressive they appear when we can see them at work in a way we can easily interpret.