One thing that is still missing, I believe, is adaptability. Take chess.
Between rounds at an amateur chess tournament you will often find players passing the time playing a game commonly called Bughouse or Siamese chess. It's played by two teams of two players, using two chess sets and two clock. Let's call them team A, consisting of players Aw and Ab, and team B, consisting of player Bw and Bb.
The boards are set up so that Aw and Bb play on one board, and Ab and Bw on the other. They play a normal clocked game (with one major modification described below) on each board, and as soon as any player is checkmated, runs out of time on their clock, or resigns, the Bughouse game ends and that player's team loses.
The one major modification to the rules is that when a player captures something, that captured piece or pawn becomes available to their partner, who can later elect on any move to drop that on their board instead of making a move on the board.
E.g., if Aw captures a queen, Ab then has a black queen in reserve. Later, instead of making a move, Ab can place that black queen on Ab's board. The capture pieces must be kept where the other team can easily see them.
You can talk to your teammate during the game. This communication is very important because the state of your teammates game can greatly affect the value of your options. For example, I might be in a position to capture a queen for a knight, and just looking at my board that might be a great move. But it will result in my partner having a queen available to drop, and my partner's opponent having a knight to drop. Once on the board a queen is usually worth a lot more than a knight--but when in reserve it is the knight that is often the more deadly piece. So, I'll ask my teammate if queen for knight is OK. My teammate might say yes, or no, or something more complicated, like wait until his opponent moves, so that he can prepare for that incoming enemy knight. In the later case, if I've got less time on my clock than my teammate's opponent has, the latter might delay his move, trying to force me to either do the trade while it is sill his turn, or do something else which will let his teammate save his queen. This can get quite complicated.
OK, now imagine some kid, maybe 12 years old or so, who is at his first tournament, and is pretty good for his age, and had never played Bughouse. He's played a ton of regular chess at his school club and with friends, and with the computer.
A friend asks him to team up, quickly explains the rules, and they start playing Bughouse.
First few games, that kid is going to cause his team to lose a lot. He'll be making that queen for knight capture without checking the other board, shortly followed by his partner yelling "WHERE DID THAT KNIGHT COME FROM!? AAAAAARRRRRGGGHHHHH!!!".
The thing is, though, by the end of the day, after playing a few games of Bughouse between each round of the tournament, that kid will have figured out a fair amount of which parts of his knowledge of normal chess openings, endgames, tactics, general principles, etc., transfers as is to Bughouse, which parts need modification (and how to make those modifications), and which parts have to be thrown out.
To get his Bughouse proficiency up to about the same level as his regular chess proficiency will take orders of magnitude less games than it took for regular chess.
I don't think that is currently true for artificial neural nets. Training one for Bughouse would be as much work as training one for regular chess, even if you started with one that had been already trained for regular chess.
"While neural nets are good at organizing a world governed by simple rules, they are not proven good at interacting with other intelligent agents." This is an interesting point, for example squeezing information through a narrow channel forces a kind of understanding that brute forcing does not. I've stopped paying close attention to the field a year ago, but I have seen a handful of openai and deepmind papers taking some small steps down this route.
Between rounds at an amateur chess tournament you will often find players passing the time playing a game commonly called Bughouse or Siamese chess. It's played by two teams of two players, using two chess sets and two clock. Let's call them team A, consisting of players Aw and Ab, and team B, consisting of player Bw and Bb.
The boards are set up so that Aw and Bb play on one board, and Ab and Bw on the other. They play a normal clocked game (with one major modification described below) on each board, and as soon as any player is checkmated, runs out of time on their clock, or resigns, the Bughouse game ends and that player's team loses.
The one major modification to the rules is that when a player captures something, that captured piece or pawn becomes available to their partner, who can later elect on any move to drop that on their board instead of making a move on the board.
E.g., if Aw captures a queen, Ab then has a black queen in reserve. Later, instead of making a move, Ab can place that black queen on Ab's board. The capture pieces must be kept where the other team can easily see them.
You can talk to your teammate during the game. This communication is very important because the state of your teammates game can greatly affect the value of your options. For example, I might be in a position to capture a queen for a knight, and just looking at my board that might be a great move. But it will result in my partner having a queen available to drop, and my partner's opponent having a knight to drop. Once on the board a queen is usually worth a lot more than a knight--but when in reserve it is the knight that is often the more deadly piece. So, I'll ask my teammate if queen for knight is OK. My teammate might say yes, or no, or something more complicated, like wait until his opponent moves, so that he can prepare for that incoming enemy knight. In the later case, if I've got less time on my clock than my teammate's opponent has, the latter might delay his move, trying to force me to either do the trade while it is sill his turn, or do something else which will let his teammate save his queen. This can get quite complicated.
OK, now imagine some kid, maybe 12 years old or so, who is at his first tournament, and is pretty good for his age, and had never played Bughouse. He's played a ton of regular chess at his school club and with friends, and with the computer.
A friend asks him to team up, quickly explains the rules, and they start playing Bughouse.
First few games, that kid is going to cause his team to lose a lot. He'll be making that queen for knight capture without checking the other board, shortly followed by his partner yelling "WHERE DID THAT KNIGHT COME FROM!? AAAAAARRRRRGGGHHHHH!!!".
The thing is, though, by the end of the day, after playing a few games of Bughouse between each round of the tournament, that kid will have figured out a fair amount of which parts of his knowledge of normal chess openings, endgames, tactics, general principles, etc., transfers as is to Bughouse, which parts need modification (and how to make those modifications), and which parts have to be thrown out.
To get his Bughouse proficiency up to about the same level as his regular chess proficiency will take orders of magnitude less games than it took for regular chess.
I don't think that is currently true for artificial neural nets. Training one for Bughouse would be as much work as training one for regular chess, even if you started with one that had been already trained for regular chess.