Can you beat a stochastic parrot? ParrotChess.com

Yes I can, apparently. Played as White. It played some variation of a Sicilian (haven’t played a human in 40 years so my opening knowledge is…sketchy). We each played solid moves, a few exchanges but it ended up with three pawn islands to my two. We exchanged all other pieces and rook. Rook and pawn endgame. I offered the remaining rook exchange that let me centralize my king and get connected Queen Rook and Queen Knight pawn against its isolated Queen Bishop pawn. Its king ran to the queenside while I blockaded its three-two pawn majority on the kingside. It pushed the Queen Bishop pawn (I think in English Descriptive Notation showing how old I am) out of desperation, I grabbed it and was close enough to the black Kings Bishop pawn be able to capture it before queening. The connected passed pawns trapped its King on the queenside and I could eat the black Kingside pawns at leisure.

But I couldn’t. When I reached the point of capturing the Kings Bishop pawn the program froze. An obvious win, but it didn’t let me make the next move, nor did it resign. Just froze. Weenie program.

I think I played optimal moves when we got to the rook and pawn stage. Magnus would have been proud. But it played way faster than I did. Trading the other rook I think was the crucial mistake because it gave me a free centralization of my King with no other pieces to confuse things, and at that point it was sort of a zugswangy forced mate in 15 or so. (Love saying that) But it took very precise counting and calculating at the end, a question of a single tempo. An alpha-beta engine would have seen it in advance and not exchanged the rook, thus forcing a draw with perfect play. My guess.

I took some sort of online rating thing on chess.com (?) a few years ago and it thought my rating was mid-1800s. Played probably a couple hundred 5 minute games or so with Stockfish giving Queen Knight odds at its highest rating level over the last few years and got to where I could hold my own, so I have played recently, just not humans. (At the beginning I couldn’t even win with Queen odds, humbling.)

See https://twitter.com/GrantSlatton/status/1703913578036904431.

The name references a well-known but non-technical paper "On the dangers of stochastic parrots: Can language models be too big?" with 2209 citations that included quotes such as "only have success in tasks that can be approached by manipulating linguistic form", "text generated by an LM is not grounded in communicative intent, any model of the world, or any model of the reader’s state of mind." It has not aged well.

Duplicate https://news.ycombinator.com/item?id=37603427