0:50 that is not the queen's gambit, queen's gambit is 1. d4 d5 2. c4, the c4 is the move that gambits, i.e., it offers up a pawn for intangibles. Come on, this is basic stuff.

fun fact: kings can flee over the edge

→ the world's two smartest AIs
→ two LLMs

"2 Ai's play normal chess with an infinite background"

I have been following this project since you first released the website, this is awesome

6:55  the "list of legal moves to choose from" would be infinite xD

Kings can jump over the edge

What I've found is that when playing novel games, the first iteration is usually close-minded like this, but they can write down lessons learned and get better over time.

When the world needed him most... he returned.

I enjoy this game so much and I’m glad you still care about it. hope to see more from it in the future and maybe even some more crazy infinite ordinal checkmate positions 🙏

This seems like an interesting AGI test, seeeing if the AIs they can actually look beyond their training data. How much time* did each AI take in total?

You can nudge them towards thinking in infinite coordinates by being explicit in the reasoning behind it:
"While normal chess notation uses an alphabetic character followed by an unsigned integer for brevity, infinite chess has more than 26 files and has no true starting point. Because of this, infinite chess notation uses two signed integers separated by a comma, with the bottom left rook (A1 in normal chess) being 1,1. An example of infinite chess notation that is impossible to represent with normal chess notation is -2439,125"

imo that is even more important than noting that kings can flee over the edge. In fact, it's likely key to stopping them from drawing on published opening lines which are a bunch of normal notations strung together.

6:23 They don't know. They just have the tokens for those words and keep repeating them.

No flurry of inventing pieces and moves that are impossible as soon as the state becomes novel, that's new.

2:32 I suspect the fact you didn't define what Infinite Chess was before inserting the "choose...legal move" led them to interpret this instruction as normal chess movement rules with some other meaning of "infinite". At least with older models, it was the case that you want to put the instruction at the end, after all needed background information.

Obviously, the actual rules should have gotten them to reconsider their earlier interpretation, but that's annoyingly difficult to get an LLM to do these days.

Now put a fly brain in infinite chess

Looks like AIs are limited to the classic chessboard by their theory they learned during machine learning. They might also be connected to a chess engine as a "skill" (api/tool they have access too) in the server's backend. Some chess engines such as stockfish one can support slightly larger boards than 8x8. If the AI does an illegal move, they might have fed their internal engine wrongly or misinterpreted its output.

Kings can flee over the edge

This is so fire 🔥

Off topic but the song in the background 1:45 was actually the same song as a SORA 2 video. I can now infer that SORA 2 had this song in its training data