Language models have shown remarkable progress in generating human-like text. One of the most advanced language models is ChatGPT, based on deep learning techniques and pre-trained on vast amounts of text data. This technology has been used to generate text for various applications, including chatbots and language translation tools. One potential application of language models is to use them in dialogue trees for non-player characters (NPCs) in triple-A games, which could lead to more engaging and immersive gameplay.
If a major triple-A game (as defined by industry standards) is released by the end of 2023 that features dialogue trees for NPCs that have been confirmed to incorporate language models or similar natural language processing technologies, the market will resolve as "Yes."
If no major triple-A game is released by the end of 2023 that features dialogue trees for NPCs incorporating language models or similar natural language processing technologies, or if the use of such technologies in triple-A games by the end of 2023 is only in a limited or experimental capacity, the market will resolve as "No."
Related questions
🏅 Top traders
| # | Name | Total profit |
|---|---|---|
| 1 | Ṁ1,086 | |
| 2 | Ṁ483 | |
| 3 | Ṁ360 | |
| 4 | Ṁ246 | |
| 5 | Ṁ210 |
Unwittingly made a similar market on a longer timescale - I agree with issues around compute requirements at least
https://manifold.markets/Tomoffer/will-there-be-a-wellreviewed-video?r=VG9tb2ZmZXI
There's a bunch of duplicate comments below which I can't delete :)
.
.
.
No chance
AAA Game timelines are 2-3 years on the short end
A change on this scale would upend the entire narrative and structure of a game, forcing designers to build a game around this "feature"
Y'all are underestimating the amount of compute this would require. The economics of games simply don't allow companies to incur charges that scale at the rate this would require.
Do mods count?
I'm kind of kidding here as I think this would be "limited in capacity" but I think mod implementation of this would be likely...
@Kormann which AAA game developer do you think is most likely to fulfill the YES criteria this year?
@LukeHanks I don't know much about game developers . My confidence stems from my knowledge about the technology.
The idea that the are not willing to replace their current static dialogue trees with ml at all doesn't make sense to me. The idea that they are not capable of implementing and training a new model before 2023 even if they are not willing to employ eg. Llama.cpp for legal reasons seems completely unlikely too
@Kormann One, AAA games usually take several years to make. Any game on that scale that's planned for release this year is already scrambling to fix bugs and implement polish in preparation for release. They don't have time to design, implement, and test some huge new experimental gimmick feature at this point.
Two, LLMs of reasonable quality are extremely expensive to run. They'd either have to run them locally on high-powered graphics cards (which would take resources away from rendering pretty graphics, and AAA developers have historically always put graphics ahead of any other feature), or they'd have to rely on a cloud service (in which case the whole game becomes unplayable if the network has problems, not to mention the ongoing cost of operating or licensing such a service).
Three... Why? The whole point of an NPC is to stand around and give you information about where to go in the game. If NPCs start hallucinating nonsense about the game world, players won't have any reliable idea about where to go, and the game will become unplayable. Not to mention the risk they'll start saying horribly racist things if you prompt them the wrong way.
(I do play as a professional game developer in real life, fwiw.)
@NLeseul Also, arent LLM weights way too big to fit reasonably in a single server's memory, let alone a gaming PC's memory ?
But i have to disagree with your last point. Today i saw a post about BibleGPT, QuranGPT and GitaGPT. This says that given enough information about the world/character/book, the LM can act as a interface/flawed interface to the world. Asking the right question to figure out something itself can become a good RPG element. But i imagine it would be hard to test it for bugs and all.
@NLeseul I agree. I don't play games, but to me this sounds like something that would be a big part of the game. It's not something you just slap on a 90% done game, but something you start with and build around it. Like you make a Sherlock Holmes game where characters are trained with data from the books. And of course somehow deal with hallucinations, jailbreaking,...
@HarishGanesan It's probably not impossible; people have mentioned LLaMa as being usable on consumer hardware on here. I don't know how high-end a machine you'd need to have, though; developers would be pretty unwilling to publish something that only 5% of their target market can actually run. (Plus you'd probably still have to completely turn off GPU rendering during conversations.)
With GPU requirements for rendering, you can always give people options to turn off advanced rendering features or lower the framerate to make the game playable, if suboptimal, on lower-end hardware. I'm not sure if there's really an equivalent of that for LLM processing; an LLM of a certain size is always going to have pretty fixed requirements. At most, you might be able to turn off LLM processing completely and use pregenerated text on lower-end machines.
(I think a cloud service is probably a more likely path if anyone chose to actually pursue this. No worries about hardware compatibility, and you can deploy emergency updates for exploits much more quickly.)
I did poke at BibleGPT a little bit just now. It does at least support its advice with actually-existing Bible citations, which is more than ChatGPT can currently do.
Llama models can run on under a GB of space. This would of course produce no performance like chatgpt but still easily 10 times better then just reading of a static script.
Graphics would not need to pause as long as the GPU has enough space to load the model on top of the other stuff.
A little model can be trained without ever seeing a single racist sentence.
Cloud hosting:
If a game costs 50 bucks would it be crazy to take 5 more so you could rent a small model from openai which would run for one Standart playthrough. And if the connection goes down you fall back on the Standart method
@Kormann no, a model that fits into the free space of a 3060 will not give more enjoyable results than a mildly competent human writer creating a few options in advance
@CodeandSolder heh, I honestly don't know why I had bought yes. Probably was intending to trade hype.
That said the larger llama models do well on the 24gb GPUs and decent on 16gb of vRam with the 4bit quantisation.
It's gutting me right now that I don't have a 48gb GPU for my own local model. I need to find a job that will give me one so I can make my hologram gf.
This market should almost be; "will there be a triple A game that requires a minimum of 24gb vRAM this year?"
(However, I think we are all being a bit silly not appreciating the indy market and having such models running on a network that the player connects too. I don't think it will be triple A though)
If you are interested I can point you in the direction of systems already utilising systems like this in a dungeons and dragons esque way.
@Fivelidz that could be possible, though it would require some clever resource sharing algorithms but the other big argument against it is making it work the required 99.9% of the time would nearly certainly take longer than the time from LLMs becoming good to market close minus testing and bugfixing
More shoe-horning it into Skyrim: https://www.youtube.com/embed/Gz6mAX41fs0?autoplay=1
Needs a bit of work but you can see where it's going