Engine-invented variations/hypotheticals

Question

After watching a few alphaZero vs stockfish games, I was very intrigued. Clearly, engines can play at a different kind of level, but for the most part the openings seem pretty meta. Nothing extremely heretical jumped out at me in the first few moves. I dont know how much we humans can read into that; best case scenario the last few hundred years have modeled the opening permutations correctly if engines still use them. Still, engines crush humans on a regular basis.

Question

What would happen if we forced the engine to use a very absurd opening for the first few moves (like moving both rooks out) then have it resume against a human/GM?

Answer 1

I find watching the latest generation of self-taught AIs play is incredibly exciting - and of course they are relatively opaque as is much of AI these days.

Maybe openings turn out to be relatively tractable in that there are not so many fundamentally sensible options, given the transpositions which are possible. A player who begins with a few garbage moves however may still be able to utterly overwhelm a weaker adversary through superior midgame play. So the answer is it depends how much damage the AI inflicted on itself.

The midgame is where the real innovation of alpha zero seems to be. Particularly the received wisdom about pawn play has been turned on its head: pawn sacrifices can be powerful to win tempos and to clear space for line-moving pieces. Surely new concepts like this must already be a factor in top-level human analysis now.

Answer 2

Just my opinion.

Watch the GMs, epically Nakamura, play very unusual openings, odds games, and other stunts against the "mere" 2600s, and you'll find that they still win from a bad position. AlphaZero has shown that it can sacrifice several pawns against the highest rated chess engine. (Game Changer: AlphaZero's Groundbreaking Chess Strategies and the Promise of AI by Matthew Sadler and Natasha Regan is a good book giving many examples of exceptions to human evaluation the common positions.)

Answer 3

Nothing special happens.

Engines play whatever positions you give them. They have no understanding of the word 'absurd'. If you want them to play with a naked king against 3 queens, they'll do it. They'll lose, of course, but they'll do it.

If your question is actually "can a human GM beat an engine if the opening is absurd enough?" the answer is trivially yes. Top human GMs already beat engines at knight odds, so if your opening e.g. gives away a queen for no compensation, humans will win easily. The real question is how unbalanced an opening you need for the match to be balanced, which is addressed in your follow-up question.