AI revolution in gaming: NPCs with memory, emotions and real personality – beyond games

[Redaktion]

NPCs that remember players, display emotions and form genuine bonds – what once sounded like science fiction is now becoming reality. A young start-up from San Francisco is developing AI-driven characters that continue to grow and evolve beyond the game itself.

https://www.notebookcheck.net/AI-revolution-in-gaming-NPCs-with-memory-emotions-and-real-personality-beyond-games.1140779.0.html

[TruthIsThere]

Her? This is disturbing.

[GeorgeS]

IMHO: here's one of two area's where "AI" could transform computer gaming as we currently experience it.

The nearly universal long standing complaint that whatever "AI" that drives NPC's in games while surely MIGHT BE "A" is also surely lacking in INTELLIGENCE!!

Game developers have yet to bake in any believable "AI" into NPC's.

The second is with game generated content, here is an area ripe for having some "AI" tossed at it to help create new dynamic game play environments.

[How it used to be]

You know what the difference was in the past compared to now?

Much less hyping back then.

I don't remember Crysis going on for 5 years about how it would revolution ize graphics. It just came in 2007 and blew everyone away.

We've been fed this AI hype juice for years now. Is there a single game out there that using AI that has done any good?

In fact, it's been the opposite response. Games that have used it, have been boycotted and review bombed.

Forget AI, BF6 doesn't even have RT and look how well it did. It did well precisely because it didn't have it which enabled it to run well on a wide variety of hardware.

Maybe GTA 6 will finally get it right but that's still far away and not out yet.

[A]

IMHO: here's one of two area's where "AI" could transform computer gaming as we currently experience it.

The nearly universal long standing complaint that whatever "AI" that drives NPC's in games while surely MIGHT BE "A" is also surely lacking in INTELLIGENCE!!

Game developers have yet to bake in any believable "AI" into NPC's.

The second is with game generated content, here is an area ripe for having some "AI" tossed at it to help create new dynamic game play environments.

Because we lacked the processing and ram hardware to do it. For AI to be anything useful with good context size, you are going to have to run a good sized model. I'd imagine you'd want at least 12gb of ram just for the AI, so at bare minimum you would need 16gb vram. You may be able to decrease the amount needed if you limit the npcs knowledge, personalities and context sizes. But if you want more realism, you would want 96gb in vram to run the large models with huge context sizes.

And processing power too, you would want 3060 minimum for the small models and maybe 4070 minimum for larger ones. But that is just for the AI. Add in the game itself and you would want 4090 hardware minimum.

You know what the difference was in the past compared to now?

Much less hyping back then.

I don't remember Crysis going on for 5 years about how it would revolution ize graphics. It just came in 2007 and blew everyone away.

We've been fed this AI hype juice for years now. Is there a single game out there that using AI that has done any good?

That is because the so called AI usage has been simple stuff which doesn't really adapt to the user, it is just pretrained algorithm which most games use for monsters and other basic stuff. Then there is the generated AI content which companies use to save money and that is what people boycott.

What is talking about here is usage of AI in bringing characters to life. This isn't same thing as generated AI, it is a completely new paradigm that wasn't possible before.

But of course the specs required for it are quite high if you want anything decent, far more than requirements for crysis.

[How it used to be]

What is talking about here is usage of AI in bringing characters to life.
...
it is a completely new paradigm that wasn't possible before.

I'm going to assume this is tied/locked into using Nvidia libraries so you basically need their hardware?

In a day and age where everything is cross-platform and developers code for lowest common denominator (consoles with amd hardware and limited vram or the switch which is Nvidia hardware but extremely limited vram) with the biggest userbase, good luck with that.

Yeah, I'm sure in 20 years we'll have consoles with 96GB ram and that the models will require far less vram + algorithms will get better. But this is exactly my point. This isn't *today*.

Today, if you want to run something well, you need a nuclear reactor + data centre with 50,000 Blackwell GPUs, which nobody is going to do.

The only reason to hype such tech right now is to increase to perceived value of these bay area startups. That's all.

None of the big tech actually care about gaming. Microsoft is trying to actively destroy their Xbox brand, Google literally committed seppuku with Stadia + have the absolute worst GPU-SoC-drivers on their pixel phones, Nvidia is literally selling 6GB laptop GPUs in 2025 lol, etc etc. You expect these very same big tech companies to bring their very best AI tech to gaming when they carry such disdain for the market?

There is far too much greed to build the necessary entire ecosystem to make anything actually work. Just going from one hype bubble to the other, for the sake stockholder profits. That's all. That's the real news or should be the real news here.

[RobertJasiek]

As always, "3D gaming" does not equal "gaming". Neural network gaming for specific abstract games with well trained networks, such as Go, differs: AI play is very much stronger than humans and 1GB VRAM is more than enough for AI inferencing. (Training needs much VRAM though, but it can be distributed.) More GPU cores help the 0.1% strongest human users for greater study speed though. For every "weaker" player, a mid tier laptop dGPU is enough.

[A]

What is talking about here is usage of AI in bringing characters to life.
...
it is a completely new paradigm that wasn't possible before.

I'm going to assume this is tied/locked into using Nvidia libraries so you basically need their hardware?

Nvidia CUDA used to be the biggest so a lot of AI was tied down to Nvidia. But since AMD has Rocm. And there are cross platform frameworks to run on all like HIP, SYCL and etc

You need 96gb ram to run a large model, if you are fine with smaller models you can do with less. Of course there are stuff like Strix Halo that can do 96gb

And as I said, if you fine tune the models you may be able to do better as the models have a lot of general knowledge that your game simply doesn't need npcs to know. Like does your npc need to know scientific research? what about sports teams? By getting rid of a lot of stuff unrelated to your needs it may be possible to get a mid or small model into even 16gb vram gpus (this includes gaming)

There may also be techniques to use such as using your regular ram or swap to pregenerate npc stuff when you are not talking to them. This can allow lower specs.