I wanna see the ML acceleration hardware in current GPUs used for something other than graphics. Imagine a decently-sized model with inputs like the player's last known position, time since last seen, last weapon used, enemy's health and ammo reserves, presence/absence of nearby buddies, etc. Outputs would be a target move position, player's estimated position, aggression factor (used to blend between "take cover" vs "guns blazing" tactics)...
You'd need to initially train it by specifying suitable behaviour for many random scenarios, but could then refine it further by capturing player data and phoning it home. Whenever the AI succeeds in killing the player, use that data for the next round of training.
In fact, what if you did training locally too, not just inference? Now I'm imagining a roguelite where the only meta-progression is enemy AI learning from you each time, forcing you to switch up your tactics between runs to surprise them.
As for real games that currently exist though, I'm going to give it to TLOU2. It's hard to say how smart its AI actually is, because it leans on a lot of simple-but-expensive tricks like named callouts. But it successfully creates the illusion of smart and menacing enemies, and that's the important thing for an immersive atmospheric game like that.
