Hacker News Reader

Comment by ttul | Hacker News Reader

Rumour has it that the NPCs will engage in realistic dialogues with the player that are generated by an LLM. By the time this game releases, running a reasonable LLM locally would not be out of the question. That alone would greatly increase the realism of the game.

View on HN

Wouldn't be better from business perspective to sell subscription for the game that will enable realistic AI behaviour by offloading LLM functions to the cloud? It will provide technical excuse for subscription, since it's not realistic to expect that players will have high-end systems that can run LLMs when also running that open world game with stunning graphics.

There are 2 problems to that.

First seeing how much it costs OpenAI to run their infrastructure, that rockstar will have a one time payment for the game and not a subscription from the players, and that you will have maybe 50-100 query per hour per player, it would not be feasible economically.

Second even if it was, that mean that the whole game depends on an internet connection and a backend that will need to be here forever. 10 years later they cut the backend you can’t play the game anymore.

GTA 5 sold 190 million copies and ChatGPT has 180 million users, I think Rockstar would run into problems procuring the number of cloud GPUs that could actually handle something like that.

Rockstar can rent from Azure just like OpenAI I guess. Buy capacity that matches the actual paid demand.

This would be a bit disappointing to me. They spent a lot of time authoring interesting and crazy utterances that were often absurd and funny in their previous games.

You still need author utterances to create an AI personality, but an LLM isn't restricted solely to reproducing them verbatim.

(Don't think LLM charaxters are ready for realtime use on that kind of game soon enough for the game being discussed, but I don't think it would have to compete with creative character line writing as much as being a force multiplier for it when it is ready.)

I think LLM is ready for non-critical-path NPC chatter that you "overhear" in the game.

Putting it in the critical path will take a lot of development/testing.

https://github.com/opyate/godot-llm-experiment

They've likely spent the last 4 years recording voice artists for everything. I doubt they're going to embed a local LLM into the game in the last year that has unpredictable results and replace all of that. On top of actually having the overhead needed to process it on current gen consoles.

I can see it happening in a future game in some way maybe.

Rumors that are based on nothing?

You're also ignoring that dialog has to be spoken, and while AI can do somewhat decent text to speech for things like audiobooks, it can't do anything like the kind of vocal "character personalities" in GTA. Voice actors aren't being replaced yet.

For GTA VII? Sure, almost definitely. But VI? We're not there yet.

Have you used the speech capabilities of ChatGPT? We are there.

Natural, neutral conversational tone is worlds away from having personality.

Video game characters need personality full of emotion.

No, we are not there. In a few years, we'll probably have something like it though.

As it is, even in its restricted form (it's forcibly neutral and avoids tone changes), we are there. We are talking about NPCs here too.

We are. https://www.youtube.com/watch?v=aihq6jhdW-Q

Hmm, I dunno. Is it practical to do that while the GPU is also busy with, y'know, rendering the game?

The modern NVIDIA cards have dedicated silicon for deep learning workloads, used for features like DLSS or RTX Voice. So you could theoretically tap into that for such a feature without meaningfully lowering the game's framerate.

AFAIK Apple also includes similar accelerator hardware in their modern chips.

For people who have integrated AMD or Intel GPUs attached to their main CPU, it's also possible to tap into that thing for compute workloads even if your game isn't running onto it.

You're thinking about it on a PC. Rockstar isn't known for releasing for PC first, or well even stably for PC in the first year. So unless the PS5 and XBoX have this setup, which they don't really, I don't think they'll lean into this.

> The modern NVIDIA cards have dedicated silicon for deep learning workloads, used for features like DLSS or RTX Voice. So you could theoretically tap into that for such a feature without meaningfully lowering the game's framerate.

But...what if you're already using that for DLSS? Can you do that at the same time?

GPUs render games too?

I don’t know, they’d really have to bank on it being viable to run an LLM and speech generator at the same time without it impacting the game’s performance. You’d have those two components, plus rendering the game itself all running on the GPU simultaneously. They might be trying it but I feel like it’s going to be up in the air whether it’s working by release.

I doubt they will enable mic or other free-form speech input from the player, so the LLM will only have to handle player's gestures and state of the outside world. I think smaller, light LLMs fine-tuned on their data would do good locally without having to hog the whole GPU.

Reply on news.ycombinator.com