[Suggestion] Expand your mechanics. GPT Chat + Voice.

[Rc2F6]

I saw in one video how a guy bolted a GPT-3 to his VR game, bolted on a voice generator. He would ask questions through a microphone, the voice would be translated into text using a script, the text would be sent to GPT-3, GPT-3 would get the information the character needed to know, from that information GPT-3 would generate an answer based on the question and send the text, the text would be voiced by the voice generator and the NPC would respond. If such mechanics existed, the game would be the first game to use such mechanics. The video is called [OpenAI GPT-3 Powered NPCs: A Must-Watch Glimpse Of The Future (Modbox)].

P.S. X jester gave me X awards 😆.

[Rc2F6]

Автор сообщения: NotFuji

Автор сообщения: Rc2F6

The videos are not about that. And I'm not talking about the neural network influencing the game. You're wrong. The neural network is a chat bot. Just text. The game tracks events. Why are you trying to give neural networks the work that the game engine should do? You just didn't understand this technology. A neural network is a tool, just a chat bot. Which creates the text. There are keywords in the text that were written into the information that a certain NPC should know. Then the game engine already tracks keywords in the dialog and creates an event with the correct keyword. Everything is very simple. Why is it so difficult to understand that a neural network is just text, it's just a chat bot.

If it's just text then what's the point? Canned dialogue does the exact same thing, much cheaper, and much more efficiently.

And buddy, if I've learned anything from this conversation, it's that you really don't understand the capabilities of AI. I keep bringing up AI making inferences about the game environment, because every example you've given has required that it be able to do that.

I agree with you completely. According to your idea, you need to write an artificial intelligence that will be fully aware of the game, maybe even play for you. Do you even understand that this is an insanely big job. My idea is simplicity and elegance, your idea is to just say some words without knowing what they mean in programming. You probably don't even understand how a neural network works. Maybe first you will write your own neural network, the simplest one. And then it will be nice to talk to you. In the meantime, I have a feeling that I communicate with jocks and boxers who need root to put food there. No offense. No need to complicate the world. No need to complicate life. No need to complicate at all. A simple idea. We speak with a voice, the NPC responds. What's the big deal, I don't understand, and why. You need a head to eat there.

[DoremianCleff]

Автор сообщения: Rc2F6

I suggested an idea. I don't demand that it be everywhere and everywhere. If you treat ideas this way, the world will fall silent and just wait for the rich, who may come up with something new. I don't have that kind of money. But I'm not imposing. I just suggested an idea. In my opinion, you reacted too radically to a simple idea. I didn't hear anything from you except, "THIS IS A BUNCH OF ♥♥♥♥." I read your comment in the voice of Beavis and Butthead. You probably came up with the idea that I demand something from the developers, you elevated yourself to the rank of a hero to protect yourself from a fictional problem. And they themselves condemned me for an idea in which you don't understand anything at all. Standing ovation.

"Im just sayin'" isn't an argument or a suggestion or anything of substance.

It's either work in scale in practice or not. There isn't an "idea" here, it's just "HEY GUYS, WOULDN"T IT BE COOL IF..."

You can say that all you want - no war, no hunger, no disease or poverty. It doesn't change reality of things. And doesn't make you a smarter\better person if you stand on a soap box pissing about.

The what you called thread? [OFFER]? The hell are you offering my guy? What the transaction here?

[your fangirl ✧]

Just turn on screen reader and voila you can have your afwul Microsoft sam reading lines for you. Yw

[WILLHOU]

it would be cool if the player could fly, doesnt necessarily improve the game though

[Rc2F6]

Some geniuses created a mod for Skyrim and put CHAT GPT there, which revived the NPC. Now you can communicate with them.

The chatbot allows the NPC to navigate the environment, global events and time, as well as determine which items are in the player's hands by issuing appropriate replicas — and this is all voiced by a separate neural network.

The author of the mod also pays attention to the movement of the characters' lips so that the conversation is even more realistic.

https://youtu.be/Gz6mAX41fs0

[literally me]

Dumb idea IMO, I'm a dev doing LLM stuff for my company currently. LLMs are just too inconsistent and slow if you don't have tonnes of resources to throw at the problem. Let's start hosting

A) Hosted by a third party: Data security questions, insane costs for access, an additional point of failure, won't work offline.

B) Hosted by devs: Incredibly high up-front cost, requires them learning new skills, more maintenance, won't work offline.

C) On user's PC: Mediocre results at best, heavily increased file space usage, heavily increased performance requirements. Even with quantisation, a model giving good results will still take several seconds to process and quite a few GB of space. Also hardware requirements will be all over the place as Nvidia users might need CUDA toolkit installed, a compatible GPU with the right amount of RAM and god knows what else.

LLMs just aren't ready to be inserted into a game like this yet. They're too heavy and unreliable to be implemented into a full game although progress is being made rapidly (LLaMa for example quantises very well into a nice small, decent performing model).

Also consider how annoying it would be for an LLM to hallucinate details about a case that are just completely false.

[Rc2F6]

Автор сообщения: literally me

Dumb idea IMO, I'm a dev doing LLM stuff for my company currently. LLMs are just too inconsistent and slow if you don't have tonnes of resources to throw at the problem. Let's start hosting

A) Hosted by a third party: Data security questions, insane costs for access, an additional point of failure, won't work offline.

B) Hosted by devs: Incredibly high up-front cost, requires them learning new skills, more maintenance, won't work offline.

C) On user's PC: Mediocre results at best, heavily increased file space usage, heavily increased performance requirements. Even with quantisation, a model giving good results will still take several seconds to process and quite a few GB of space. Also hardware requirements will be all over the place as Nvidia users might need CUDA toolkit installed, a compatible GPU with the right amount of RAM and god knows what else.

LLMs just aren't ready to be inserted into a game like this yet. They're too heavy and unreliable to be implemented into a full game although progress is being made rapidly (LLaMa for example quantises very well into a nice small, decent performing model).

Also consider how annoying it would be for an LLM to hallucinate details about a case that are just completely false.

You just described GPT-4 on 2TB. For this game, it's stupid. I also suggested the idea for your voice to be translated into text, processed by GPT2, which weighs 2GB. The text was returned and converted to voice. Do you think that schoolchildren in Modbox can do this on their knees. And the developer company can't? You probably don't know how to prioritize. This game is not a super realistic simulator, it is a simple investigation generator game, the dialogues there are insanely primitive. You think that GPT2 will not cope. That GPT2 needs a lot of resources and money. It's strange where the students have such money and resources. But since you are a cool developer, I will not argue with you. You're like a caveman, you don't want to look for a solution, you just don't care about progress. I would not work with you for your narrow-mindedness, even despite your experience in programming, if there is one, of course. You are a loss-making person by yourself.

[literally me]

Автор сообщения: Rc2F6

Автор сообщения: literally me

Dumb idea IMO, I'm a dev doing LLM stuff for my company currently. LLMs are just too inconsistent and slow if you don't have tonnes of resources to throw at the problem. Let's start hosting

A) Hosted by a third party: Data security questions, insane costs for access, an additional point of failure, won't work offline.

B) Hosted by devs: Incredibly high up-front cost, requires them learning new skills, more maintenance, won't work offline.

C) On user's PC: Mediocre results at best, heavily increased file space usage, heavily increased performance requirements. Even with quantisation, a model giving good results will still take several seconds to process and quite a few GB of space. Also hardware requirements will be all over the place as Nvidia users might need CUDA toolkit installed, a compatible GPU with the right amount of RAM and god knows what else.

LLMs just aren't ready to be inserted into a game like this yet. They're too heavy and unreliable to be implemented into a full game although progress is being made rapidly (LLaMa for example quantises very well into a nice small, decent performing model).

Also consider how annoying it would be for an LLM to hallucinate details about a case that are just completely false.

You just described GPT-4 on 2TB. For this game, it's stupid. I also suggested the idea for your voice to be translated into text, processed by GPT2, which weighs 2GB. The text was returned and converted to voice. Do you think that schoolchildren in Modbox can do this on their knees. And the developer company can't? You probably don't know how to prioritize. This game is not a super realistic simulator, it is a simple investigation generator game, the dialogues there are insanely primitive. You think that GPT2 will not cope. That GPT2 needs a lot of resources and money. It's strange where the students have such money and resources. But since you are a cool developer, I will not argue with you. You're like a caveman, you don't want to look for a solution, you just don't care about progress. I would not work with you for your narrow-mindedness, even despite your experience in programming, if there is one, of course. You are a loss-making person by yourself.

GPT-2 sucks ass we tried it at work and the results were awful. Also there's a difference between loading up python and using Hugging Face to do some simple inference, or dropping a model into oobabooga's text generation webui and actually integrating it within software.

Shadows of Doubt is built on unity, which is C#. Most ML tools are on python sadly, especially LLM stuff. However C# does have TorchSharp which could be used to load a PyTorch model but the process would be incredibly manual and require the devs to learn a huge new skillset.

[Rc2F6]

Автор сообщения: literally me

Автор сообщения: Rc2F6

You just described GPT-4 on 2TB. For this game, it's stupid. I also suggested the idea for your voice to be translated into text, processed by GPT2, which weighs 2GB. The text was returned and converted to voice. Do you think that schoolchildren in Modbox can do this on their knees. And the developer company can't? You probably don't know how to prioritize. This game is not a super realistic simulator, it is a simple investigation generator game, the dialogues there are insanely primitive. You think that GPT2 will not cope. That GPT2 needs a lot of resources and money. It's strange where the students have such money and resources. But since you are a cool developer, I will not argue with you. You're like a caveman, you don't want to look for a solution, you just don't care about progress. I would not work with you for your narrow-mindedness, even despite your experience in programming, if there is one, of course. You are a loss-making person by yourself.

GPT-2 sucks ass we tried it at work and the results were awful. Also there's a difference between loading up python and using Hugging Face to do some simple inference, or dropping a model into oobabooga's text generation webui and actually integrating it within software.

Shadows of Doubt is built on unity, which is C#. Most ML tools are on python sadly, especially LLM stuff. However C# does have TorchSharp which could be used to load a PyTorch model but the process would be incredibly manual and require the devs to learn a huge new skillset.

Already on the merits. Now I like you better as a critic of this proposal. GPT-2 as an example, I mean you don't need a super smart neural network. There are a bunch of GPT models. Pick what's most suitable, with an intelligence that could produce at least a five-word sentence. That would be enough. I understand, there are always pitfalls like python. But agree, when there weren't any in programming. Especially in C++. I've written a lot of sheets of code. And I can tell you none of the assignments were so perfect and smooth. It's the same here. It's just another puzzle that needs to be solved correctly. Without a lot of money and a lot of work and a lot of quantum computer power. For a project like this, even more so.

[literally me]

Автор сообщения: Rc2F6

Автор сообщения: literally me

GPT-2 sucks ass we tried it at work and the results were awful. Also there's a difference between loading up python and using Hugging Face to do some simple inference, or dropping a model into oobabooga's text generation webui and actually integrating it within software.

Shadows of Doubt is built on unity, which is C#. Most ML tools are on python sadly, especially LLM stuff. However C# does have TorchSharp which could be used to load a PyTorch model but the process would be incredibly manual and require the devs to learn a huge new skillset.

Already on the merits. Now I like you better as a critic of this proposal. GPT-2 as an example, I mean you don't need a super smart neural network. There are a bunch of GPT models. Pick what's most suitable, with an intelligence that could produce at least a five-word sentence. That would be enough. I understand, there are always pitfalls like python. But agree, when there weren't any in programming. Especially in C++. I've written a lot of sheets of code. And I can tell you none of the assignments were so perfect and smooth. It's the same here. It's just another puzzle that needs to be solved correctly. Without a lot of money and a lot of work and a lot of quantum computer power. For a project like this, even more so.

There are a bunch of GPT models and the small ones suck ass and the big ones are unsuitable. The fact that when LLMs come into question, you only think of GPT is very telling of your experience with them. You've seen the headlines and played around with other people's tools built on it because it has hype but have never spent a significant amount of time actually trying to integrate it into a real project yourself.

If it's as easy as you think it is to deploy an LLM within a game without being dependent on cloud processing then go do it yourself and sell it on steam since you think it's such a golden idea.

It's a very new field with a high cost of entry in terms of processing, it needs to mature. This time next year, or even a couple of months from now this idea might be more feasible for an indie dev team but as of now it seems like just trying to jump on a hype train.

[Zothen]

Imagine a bunch of squirrels discussing rocket science....

[literally me]

Автор сообщения: Zothen

Imagine a bunch of squirrels discussing rocket science....

Working with LLMs is literally part of my job description. Admittedly I'm not experienced on the actual development of one from scratch, but most people who work with them aren't and that area isn't relevant here.

It's not my whole job but it's a large chunk of it. Creating a chatbot for use by our customers and ourselves internally (can't say much more than that). I handle data, hardware, some of the fine tuning/training work, the front and back end.

[DrDeadRed]

This just seems like a really pointless tacked on gimmick idea. Not to mention the work required to actually integrate something like that into the game.

Like what reason is there for it to be added to this game specifically? It justs seem tbh you've seen a video about it and then gone 'thats so cool I want a game with that' and suggested it here effectively randomly.

Would be like me suggesting they add VR support and motion controls, becuase it would 'so cool' to have that in the game, when there realistically is no reason this game would benefit from them more than any other game would.

In an ideal world they would add any and all features to make the game perfect, but this is an early access title being dev'ed by a small team. Would be best they stick to actually adding content that would improve the gameplay a lot rather than just make the AI covos mildly more interactive which doesnt really matter if the gameplay is the exact same.

[Rc2F6]

Автор сообщения: DrDeadRed

This just seems like a really pointless tacked on gimmick idea. Not to mention the work required to actually integrate something like that into the game.

Like what reason is there for it to be added to this game specifically? It justs seem tbh you've seen a video about it and then gone 'thats so cool I want a game with that' and suggested it here effectively randomly.

Would be like me suggesting they add VR support and motion controls, becuase it would 'so cool' to have that in the game, when there realistically is no reason this game would benefit from them more than any other game would.

In an ideal world they would add any and all features to make the game perfect, but this is an early access title being dev'ed by a small team. Would be best they stick to actually adding content that would improve the gameplay a lot rather than just make the AI covos mildly more interactive which doesnt really matter if the gameplay is the exact same.

If you think like you, then why all these new games, with new ideas, play Tetris at all.

[literally me]

Автор сообщения: Rc2F6

Автор сообщения: DrDeadRed

This just seems like a really pointless tacked on gimmick idea. Not to mention the work required to actually integrate something like that into the game.

Like what reason is there for it to be added to this game specifically? It justs seem tbh you've seen a video about it and then gone 'thats so cool I want a game with that' and suggested it here effectively randomly.

Would be like me suggesting they add VR support and motion controls, becuase it would 'so cool' to have that in the game, when there realistically is no reason this game would benefit from them more than any other game would.

In an ideal world they would add any and all features to make the game perfect, but this is an early access title being dev'ed by a small team. Would be best they stick to actually adding content that would improve the gameplay a lot rather than just make the AI covos mildly more interactive which doesnt really matter if the gameplay is the exact same.

If you think like you, then why all these new games, with new ideas, play Tetris at all.

⠀⠀⠘⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡜⠀⠀⠀
⠀⠀⠀⠑⡀⠀⠀⠀⠀⠀ ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡔⠁⠀⠀⠀
⠀⠀⠀⠀⠈⠢⢄⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⣀⠴⠊⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⢸⠀⠀⠀⢀⣀⣀⣀⣀⣀⡀⠤⠄⠒⠈⠀⠀⠀⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⠘⣀⠄⠊⠁⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⠀⠀⢀⡠⠔⠒⠒⠒⠒⠒⠢⠤⣀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⡰⠉⠁⠀⠀⠀⠀⠀⠀⠀⠀⠀⠈⠑⢄⡀⠀⠀⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⡸⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡀⠀⠀⠀⠀⠙⠄⠀⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⢀⠁⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠃⠀⢠⠂⠀⠀⠘⡄⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⢸⠀⠀⠀⠀⠀⠀⠀⠀⠈⢤⡀⢂⠀⢨⠀⢀⡠⠈⢣⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⢀⢀⡖⠒⠶⠤⠭⢽⣟⣗⠲⠖⠺⣖⣴⣆⡤⠤⠤⠼⡄⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠘⡈⠃⠀⠀⠀⠘⣺⡟⢻⠻⡆⠀⡏⠀⡸⣿⢿⢞⠄⡇⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⢣⡀⠤⡀⡀⡔⠉⣏⡿⠛⠓⠊⠁⠀⢎⠛⡗⡗⢳⡏⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⢱⠀⠨⡇⠃⠀⢻⠁⡔⢡⠒⢀⠀⠀⡅⢹⣿⢨⠇⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⢸⠀⠠⢼⠀⠀⡎⡜⠒⢀⠭⡖⡤⢭⣱⢸⢙⠆⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⡸⠀⠀⠸⢁⡀⠿⠈⠂⣿⣿⣿⣿⣿⡏⡍⡏⠀⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⢀⠇⠀⠀⠀⠀⠸⢢⣫⢀⠘⣿⣿⡿⠏⣼⡏⠀⠀⠀⠀⠀⠀⠀
⠀⠀⠀⠀⣀⣠⠊⠀⣀⠎⠁⠀⠀⠀⠙⠳⢴⡦⡴⢶⣞⣁⣀⣀⡀⠀⠀⠀⠀⠀
⠀⠐⠒⠉⠀⠀⠀⠀⠀⠀⠀⠀⠀⠠⠀⢀⠤⠀⠀⠀⠀⠀⠀⠀⠈⠉⠀⠀