I predicted something similar. Basically: LLMs generating "fluff" speech of unimportant NPCs, so they don't repeat the same thing over and over. Then the underlying tech (neural networks) replacing scripts to decide opponent actions; the main appeal will be that opponents will adapt to your playstyle, as if you were "training" the model without realising it.

For example, let's say you're playing a racing game. Your first runs will be pretty much the same as now. But later runs, the AI "catches" on your strategy, and reacts to it, like:

• if you're often bumping enemies out of the track, the AI avoids getting close to you, even if it needs to slow down a bit.
• if you lead the run from the start, and fuel consumption be damned, the AI might want to stay right behind you, as if "riding" on your pressure drag. Until the right moment where it speeds up to run past you.
• if you avoid helping the opponents with pressure drag, and keep the boosters/nitros for the last lap, the AI might do the same.
• etc.

It would be challenging for game designers, too; they'd need to find the right metrics for the AI. For example if you simply "reward"/"punish" the AI based on wins and losses, the AI will use strategies like ganging up against the player - sure, this is effective, but it is not fun at all.

BTW here's a cool video of someone "teaching" AI to play Monopoly. I believe the process will be similar-ish, except models would be already a bit pre-"trained" a bit when you buy the game, and further "training" is simply from the player playing it.

CPU/IGP RAM inference is quite viable; the software/training infrastructure just isn't there yet. And honestly this is a good case for a swarm or APIs.

And yeah, the slop will be horrible until they're used in a more integrated/modest manner.

...Or it might just be crap forever? I dunno, there are a lot of dud ideas in gaming software world.

To replace the existing AI/bots in games? I admit, I have no idea how current game AI works, but LLMs might make them way smarter than before.