I don’t think you understood my point. I’m not saying train the robots to say things by rewarding them for doing so. I’m saying “force” them to do so as a hard

I don’t think you understood my point. I’m not saying train the robots to say things by rewarding them for doing so. I’m saying “force” them to do so as a hard coded rule. And also force them to read the chats of the other AI allies as an input. Continuously train the model to map the messages into an embedding space and use the embedding for decision making. This would likely benefit by some cherry picked training data, but I think AI could get good at communicating their state to each other. Star craft has too much entropy to make a simple example. Imagine the game was just, you pick a random number and your ally has to guess what it is. I think Ai’s could learn to communicate this using a similar training design.