|
Hot dang I got LLaMA working! Man I hate all those CUDA errors and crazy Linux stuff, but at least now I can
|
# ? Mar 6, 2023 16:59 |
|
|
# ? May 28, 2024 16:23 |
|
Deki posted:Depends on if you want the ability for the player to ask entirely freeform questions or not. You could seed the context with points of interest and stuff, but in your example it works fine if the player asks that specific question, but what about like… “has Dracula drained any of your loved ones?” There is no good response for that unless the NPC has relevant context. They can’t really respond “I don’t know” and if they answer anything else without sufficient pre-fed context then it is going to just make poo poo up that might be inconsistent with other established facts/events in the game. GPT is pretty good at maintaining context within a conversation, but using that in a game with actual constraints and static story/character/etc elements is where it falls apart (Right now). Lucid Dream fucked around with this message at 17:07 on Mar 6, 2023 |
# ? Mar 6, 2023 17:03 |
|
Yeah, it could be done now in a limited way by having the game add detailed hidden messages to the prompts to guide the output. It definitely won't be perfect, though. I expect we'll see a few janky indie projects using it in a year, some smaller indie games doing it in 2 years, and it becoming AAA/mainstream... in like 5 years or longer, if the next consoles have dedicated AI hardware.
|
# ? Mar 6, 2023 17:08 |
|
pixaal posted:Well if you are running it local you just have the AI check the game state, and if it's unknown it's being decided on right now. Doing it purely in GPT would be silly. The AI “checking the game state” is really the problem though. I don’t think NPCs with real time dialog will work well until there is a way to build all of the relevant context for that NPC dynamically based on the state of the world. It’s the kind of problem that is solvable in principle, but also the kind of thing that very smart people in research labs are working on (and haven’t solved yet).
|
# ? Mar 6, 2023 17:11 |
|
Lucid Dream posted:The AI “checking the game state” is really the problem though. I don’t think NPCs with real time dialog will work well until there is a way to build all of the relevant context for that NPC dynamically based on the state of the world. It’s the kind of problem that is solvable in principle, but also the kind of thing that very smart people in research labs are working on (and haven’t solved yet). Talking about building a large scale game with current tech is silly you have no room for anything but the AI and will require very powerful GPUs the cost of production and how niche a product this will be with most people going "yeah but it looks like poo poo and I need a 4090", this is 5-10 years out at a min, all of these features are things that have to be solved before it's viable and they can be. I think currently you have to check state which means it's going to take multiple requests it will be slow like a DOS RPG, please enjoy this midi file while this loads for 15 minutes.
|
# ? Mar 6, 2023 17:13 |
|
pixaal posted:Talking about building a large scale game with current tech is silly you have no room for anything but the AI and will require very powerful GPUs the cost of production and how niche a product this will be with most people going "yeah but it looks like poo poo and I need a 4090", this is 5-10 years out at a min, all of these features are things that have to be solved before it's viable and they can be. That's not even DOS RPG. Loaders were like Commodore and Amiga things IIRC.
|
# ? Mar 6, 2023 17:17 |
|
I think clever use of dedicated servers would make a game using it viable much quicker than systems using local resources, no?
|
# ? Mar 6, 2023 17:22 |
|
Fuschia tude posted:That's not even DOS RPG. Loaders were like Commodore and Amiga things IIRC. I could be misremembering just how old that computer was
|
# ? Mar 6, 2023 17:23 |
|
pixaal posted:Talking about building a large scale game with current tech is silly you have no room for anything but the AI and will require very powerful GPUs the cost of production and how niche a product this will be with most people going "yeah but it looks like poo poo and I need a 4090", this is 5-10 years out at a min, all of these features are things that have to be solved before it's viable and they can be. The problem right now is that “check the state” means something very fuzzy, and turning that into a text prompt in such a way that actually provides sufficient context for a conversation is a Hard problem. Internal consistency within one chat isn’t really hard, but it’s the consistency with the rest of the world that is hard. With existing systems, you either have to build with the assumption that the AI will hallucinate stuff (not an impossible gamedev challenge, but one that implies a lot about the kind of game it can be) or you have to constrain it so much that you might as well just have multiple choice options rather than real time dialog. Again, all solvable problems, but what I’m really getting at is that real time dialog *looks* much more promising than it really is *right now*. AI is great at hallucinating things, games are great with concrete things… that translation from one domain to the other is where the problems lie at the moment. Lucid Dream fucked around with this message at 17:32 on Mar 6, 2023 |
# ? Mar 6, 2023 17:29 |
|
Mozi posted:Hot dang I got LLaMA working! Man I hate all those CUDA errors and crazy Linux stuff, but at least now I can I'm working this now! Thanks for setting some expectations.
|
# ? Mar 6, 2023 17:44 |
|
Kind of related to the video game discussion, how do you think they automate the endless Steamed Hams stream? Obviously it seems like they feed the AI a general structure/outline and have it randomize dishes for the 1st/2nd recipe, then an excuse for the fire. I'm just wondering the exact prompts they'd have to phrase since it's pretty stable in its loops.
|
# ? Mar 6, 2023 17:53 |
|
I had no idea there were so many language models already out there. Check out https://github.com/oobabooga/text-generation-webui/ to get started. This is also how one could possibly run the LLaMA model if they had academic credentials.
|
# ? Mar 6, 2023 18:00 |
|
I'm on Windows and only the 'one click installer' method worked for me for that to get by all the errors. Had to increase 'max_new_tokens' setting for it to produce anything interesting. Was able to lower gpu_memory setting to 22 (on a 24 gig card) and still use the 13B model fine while web browsing etc.
|
# ? Mar 6, 2023 18:21 |
|
JazzFlight posted:Kind of related to the video game discussion, how do you think they automate the endless Steamed Hams stream? Outline plus either asking GPT for random stuff or from a pool of things. You could ask GPT to come up with a random food or feed it with a food item from a big list. Given that the structure is always pretty much the same, I’m guessing that they have a prompt for each “scene” and they string them together to make an “episode”. Edit: I assume they’re using a pre-made outline, because it would diverge a lot more if they were actually generating the outline itself on the fly. Lucid Dream fucked around with this message at 18:30 on Mar 6, 2023 |
# ? Mar 6, 2023 18:26 |
|
I'm playing with Pygmalion6B right now and it's pretty impressive! It wrote the first part of the dialog in a previous prompt and I copied it over to see if it would continue the roleplay.
|
# ? Mar 6, 2023 18:34 |
|
When it comes to writing something coherent to a consistent formula, I've had success in having it go from more broad to more granular. I ask it to write a logline, then summary in a 3 act structure, then each act in more detail. I could imagine one of those infinite sitcoms doing much the same.
|
# ? Mar 6, 2023 18:38 |
|
feedmyleg posted:When it comes to writing something coherent to a consistent formula, I've had success in having it go from more broad to more granular. I ask it to write a logline, then summary in a 3 act structure, then each act in more detail. I could imagine one of those infinite sitcoms doing much the same. Yeah, you definitely can start from a simple prompt and inflate it into a whole episode by moving in steps like that, but since steamed hams is using pre-made assets they can’t actually diverge from the formula in any meaningful way. I really want to try making something more long-form and open ended and tie it into stable diffusion and see what kind of nonsense comes out the other end.
|
# ? Mar 6, 2023 18:42 |
|
I know its just generative, but dang. Literally the first chat session I opened. This is still PygmalionAI/pygmalion-6b Edit: I was expecting this to be very different from ChatGPT, but at a first brush it is very impressive. Finally a fuckup: LASER BEAM DREAM fucked around with this message at 18:55 on Mar 6, 2023 |
# ? Mar 6, 2023 18:42 |
|
You know, for all the talk about AI art going on right now I think spending so long working with Stable Diffusion and its predecessors has actually made me want to be a better artist than years of doodling. I've spent so much longer inpainting/outpainting, manually re-drawing/re-coloring, and sketching poses than just taking raw output. I actually dug my tablet out of storage and started studying anatomy again. I started staring at color gradients, geometry, brush strokes, and art styles on every piece I see again. I even fired up Blender and started watching tutorial videos again. Like, legitimately this has made me feel so much more creatively empowered as a person. I think when I see people arguing that this isn't 'real' art and is just collage or outright theft it's why I don't even get angry anymore, just sad. I understand why they are saying that stuff but the fact that the world we live in forces that reaction, that choice, bums me out a lil' bit.
|
# ? Mar 6, 2023 18:49 |
|
Objective Action posted:You know, for all the talk about AI art going on right now I think spending so long working with Stable Diffusion and its predecessors has actually made me want to be a better artist than years of doodling. I had a similar experience last week, where tinkering with Stable Diffusion to get what I wanted (you know how hard it is to make a tiefling? like, Jesus) made me open up Gimp and vector art editors to finish a D&D party portrait. It reawakened a lot of creativity for me. One huge benefit of AI-art: if there is some element of art that you don't like doing personally, you can skip it. Don't like figuring out pose? Have the AI do it. Don't want to spend ages on the face? Have an AI get it approximately right, then clean it up. Hate making backgrounds? Make the AI do it, and focus on the characters and elements that matter to you. I likewise feel really empowered by algorithms like Stable diffusion. AI art tax: This was a final party portrait for a campaign I just finished GMing. I generated the player characters individually with Control Net with poses from models/characters portraits I found online. I then used GIMP to change the white backgrounds to alpha so I could overlay the images. I used the Dungeons and Diffusion (60%) merged with some random noise-offset model (40%, this made white backgrounds possible) for generation. There were some other hiccups I had to fix manually (asymmetric tiefling horns, pale faces) but I was super happy with the result.
|
# ? Mar 6, 2023 19:23 |
|
"Dungeons and diners and dragons and drive-ins and dives" Saw this on a shirt and had to try it as a prompt. It's kind of amazing, especially that last outfit.
|
# ? Mar 6, 2023 19:42 |
|
Vlaphor posted:"Dungeons and diners and dragons and drive-ins and dives" this is amazing
|
# ? Mar 6, 2023 19:49 |
|
Vlaphor posted:"Dungeons and diners and dragons and drive-ins and dives" Same prompt, different model
|
# ? Mar 6, 2023 20:20 |
|
How good is it at creating food items? Maybe try some entries from Guy Fieri's menu.
|
# ? Mar 6, 2023 20:24 |
|
Lucid Dream posted:The problem right now is that “check the state” means something very fuzzy, and turning that into a text prompt in such a way that actually provides sufficient context for a conversation is a Hard problem. Internal consistency within one chat isn’t really hard, but it’s the consistency with the rest of the world that is hard. With existing systems, you either have to build with the assumption that the AI will hallucinate stuff (not an impossible gamedev challenge, but one that implies a lot about the kind of game it can be) or you have to constrain it so much that you might as well just have multiple choice options rather than real time dialog. I don't think that the problem is all that hard, and we'll see some great applications of the tech within a year or two. Think of how fast ML is advancing. It's crazy. I've never seen anything like it. The Internet was close, but poo poo, this is going faster. How I expect it could be approached. - World building: Develop the 'universe' of the game. Give it history, geography, politics etc. I think it would be best developed by humans, but Dwarf Fortress does this and has been for the last 14 years. - Develop the gameplay rules: So the AI doesn't suddenly decide to start talking about lasers in your low magic fantasy world. - Train the AI on the world and rules. Give it as much information and gameplay rules as possible, as well as a good background in the type of genres and stories you want. No happy hobbits in a Deal Souls world! et viola! you now have an AI model that you can use in your game. Since you've created a model with only what it needs to know, it shouldn't be too massive, and if you can use something like safe tensors (no idea if this works on chat systems) so much the better. I suppose you could update it with patches, but that might break what comes next. I'm not sure. You're not wrong that the game will need to keep track of game state, but poo poo, RPGs do this and have been doing this since the dawn of computer RPGs. Nothing special here, but combine that with an AI that knows how to talk and react to in-game events, man this will be magical. It will be Bethesda's Radiant system for realsies. It wouldn't be hard to keep a running "newspaper" of events that the NPCs can refer to - it wouldn't even take much computing power, really. Now, the icing on the cake would be NPCs who actually remember the player, and their own history. This would get a little trickier, but I think it would quite a bit less effort than pre-scripted speech. I think what I'd do is develop a kind of personality system. Male/female/other, elf/dwarf/human/orc, backgrounds, personality traits (steal from the plethora of real world personality assessments), strength, intelligence, skills, class, etc. Again, most of that is already done in some RPGs. And you don't have to worry about brute forcing the dialogue to be able to account for every eventuality, it will react with a satisfying illusion of realism. Even the strange, unexpected interactions which will undoubtedly happen will enrich it! NPCs remembering events and the players doesn't seem complicated either. They could keep a simple 'journal' of what happened. Day 156: Met PC. They saved my pet cat. I like them +10 Day 157: PC killed my brother who joined some bandits. I hate them +50 Day 256: PC saved my hometown from a dragon. I like them +15, They impressed me +20 etc. Then in day 349, when the PC talks to them, the NPC's AI would react given their personal history, the world background, their personality, and the PC's actions. loving hell, someone make this game!!! I wants it, Precious!
|
# ? Mar 6, 2023 20:39 |
|
busalover posted:How good is it at creating food items? Maybe try some entries from Guy Fieri's menu. "" The"Hobo Lobo Bordello Slam Jam" Appetizer"" gets decent looking food. Removing the words "The Appetizer" gets something I wasn't expecting. You might notice I had to censor part of image 4. For some reason, it decided this person of the desert had a vagina with a massively huge clitoris out on full display through a rip in the pants. I have no idea why this happened and it was bizarre seeing it get generated in realish-time Edit: and I just noticed potential ballsack in image 2 as well. That thing aint human, look at it's feet. If the mods want me to cover up potential alien hobo ballsack, let me know and I will.
|
# ? Mar 6, 2023 20:57 |
|
Doctor Zero posted:I don't think that the problem is all that hard, and we'll see some great applications of the tech within a year or two. Think of how fast ML is advancing. It's crazy. I've never seen anything like it. The Internet was close, but poo poo, this is going faster. Don't get me wrong, everything in your post is doable eventually, I'm just noting that while the current chatbot stuff *seems* really good for games, there are still some very hard problems to solve before they can realistically be used for real-time dialog. Right now it is sort of a mirage, where with very little effort you can make something that *seems* amazing, but the more you work with it the more you realize how unsuitable it would be out of the box. Edit: Worth keeping in mind that figuring out how to do long term memory for these things is like... one of the major steps between where we are now and real AGI. Lucid Dream fucked around with this message at 21:03 on Mar 6, 2023 |
# ? Mar 6, 2023 20:59 |
|
More realistically, you know how every morrowind npc would adjust tone into encyclopedia style to read off a history article if you asked them about lore? Having a current text NN tweak the dialog to keep the npc's tone, and voice it with a text2speech for the npc's voice seems feasible.
|
# ? Mar 6, 2023 21:15 |
|
Vlaphor posted:Same prompt, different model Getting some DBZ vibes from some of these
|
# ? Mar 6, 2023 21:20 |
|
Vlaphor posted:"" The"Hobo Lobo Bordello Slam Jam" Appetizer"" gets decent looking food. lmao
|
# ? Mar 6, 2023 21:23 |
|
Lucid Dream posted:
What I meant was that you might not need to have long term memory. From what I understand about GPT-3, each session you start with it is distinct with no history of anything else. In playing with it, it has a better memory than AI Dungeon, but yes, it’s still limited. Keeping track of events that pertain to the NPC would mean you could have each instance ‘load up’ the history and then go from there, kind of like telling ChatGPT ‘you are a homeless one armed Indonesian on Mars as a miner of mars rocks. Go!’ With the API now available I may just play around with that idea.
|
# ? Mar 6, 2023 23:10 |
|
Objective Action posted:You know, for all the talk about AI art going on right now I think spending so long working with Stable Diffusion and its predecessors has actually made me want to be a better artist than years of doodling. I feel the same, but with ChatGPT and writing stuff for a dnd campaign. I have been putting off working on a campaign for literally 9 months, and for whatever reason the ability to fire away questions at the chat and for it to respond with instant suggestions just inspired me to work on the campaign and integrate the ai into my work flow. Ended up playing our first Session yesterday and everyone had a blast.
|
# ? Mar 6, 2023 23:56 |
|
Doctor Zero posted:What I meant was that you might not need to have long term memory. From what I understand about GPT-3, each session you start with it is distinct with no history of anything else. In playing with it, it has a better memory than AI Dungeon, but yes, it’s still limited. Keeping track of events that pertain to the NPC would mean you could have each instance ‘load up’ the history and then go from there, kind of like telling ChatGPT ‘you are a homeless one armed Indonesian on Mars as a miner of mars rocks. Go!’ The issue you'd run into is a limited working memory. ChatGPT kinda handwaves it away with some stuff behind the scenes, but the model can only handle a total of 4000 tokens, which comes out to around 3000 words. That includes the input and the output. So you can't simply stuff all the chatlogs in there, you'd have to summarize and extract key events or something. You also then have to feed it all the *other* context about the world, as well as any other instructions for the model to use. Rumors are that GPT-4 will have up to 32k tokens, which would reduce some of the bottleneck, but it doesn't really fix the underlying problem. Not to mention, more tokens in the input + output = higher cost, so a big giant wall of context for the AI so that it can respond "beans" to my latest question of "what is your favorite food" would cost a non-trivial amount of money.
|
# ? Mar 7, 2023 00:30 |
|
Doctor Zero posted:Since you've created a model with only what it needs to know, it shouldn't be too massive this sentence makes me feel like you fundamentally don't understand how these large language models work or are trained
|
# ? Mar 7, 2023 00:37 |
|
I realize it's difficult to predict, and kinda dumb question but where does everyone see all of this stuff going in 6 months to a year. It seems to me (who checks in and out with this kind of stuff) that since a bunch of stuff got leaked or publicly released shits been moving in leaps and bounds.
|
# ? Mar 7, 2023 00:47 |
|
Telsa Cola posted:I realize it's difficult to predict, and kinda dumb question but where does everyone see all of this stuff going in 6 months to a year. It seems to me (who checks in and out with this kind of stuff) that since a bunch of stuff got leaked or publicly released shits been moving in leaps and bounds. 6 months ago this technology didn't even exist is what's loving my brain. Trying to imagine 6 months from today...
|
# ? Mar 7, 2023 00:50 |
|
RPATDO_LAMD posted:this sentence makes me feel like you fundamentally don't understand how these large language models work or are trained Nope, you're right. I don't. I know how the art models work and are trained, so I was assuming it was similar to the art models. Have any good references?
|
# ? Mar 7, 2023 01:09 |
|
Telsa Cola posted:I realize it's difficult to predict, and kinda dumb question but where does everyone see all of this stuff going in 6 months to a year. It seems to me (who checks in and out with this kind of stuff) that since a bunch of stuff got leaked or publicly released shits been moving in leaps and bounds. I'm not sure on timelines, but I think we're going to see two diverging paths with AI stuff, where one path is really good text generation, really good art generation, etc but each of those things happens in a vacuum so they don't really talk to each other perfectly. Another path is multi-modal models that can use visual/audio/text as input and output. If GPT4 can do 32k tokens then it might be able to output an entire screenplay. If it's also significantly smarter than GPT3, it might even be a *good* screenplay. Some day the multi-modal models will be able to generate a screenplay and then make it into a film, but that's not going to happen in the next year.
|
# ? Mar 7, 2023 01:14 |
|
KwegiboHB posted:Hey Kakermix, is there a way to change the position of the sun in the sky through controlnet? I have an idea I want to try with a very specific position. There is not, no. The way I'd approach that is (of course) bringing the image into an editor and painting in cues for where the light is coming from. Boba Pearl posted:6 months ago this technology didn't even exist is what's loving my brain. Trying to imagine 6 months from today... "Ok I think I know how to do a thing" *blink* "gently caress" *blink* "gently caress"
|
# ? Mar 7, 2023 01:21 |
|
|
# ? May 28, 2024 16:23 |
|
Doctor Zero posted:Nope, you're right. I don't. I know how the art models work and are trained, so I was assuming it was similar to the art models. Have any good references? In general: You can't train on just a few data points, these large language models are trained on absolutely massive amounts of text just to be able to form coherent sentences or paragraphs -- GPT3 was trained on 45 terabytes of text, and EleutherAI's smaller 6-billion-parameter GPT-J was trained on an open source 825gb dataset. Same for stable diffusion actually. You might be confusing this with the dreambooth/LoRa fine-tuning process but that relies on an existing model that has already been trained at great expense. And it still knows all the stuff from the base model in addition to the stuff it was trained on, which is why something like a sd dreambooth model trained on GTA5 portrait art will understand how to draw Joe Biden or a cat in the GTA style, even though neither of those were in the fine tuning data. In the same way, a pre-trained LLM fine tuned on your RPG setting would still have all the data from its 800gb "how do i form sentences" training set and would be very capable of 'hallucinating' new setting details based on that original data. Also, "how much it knows" has absolutely 0 relation to the size of the model weights. The training process for these is not adding new storage for new information, it is basically adjusting individual coefficients in a giant equation. The model weights for an untrained LLM are exactly the same size as the ones for one that's been force fed 45TB of internet, they just contain different numbers. There's a pretty good blog post from Steven Wolfram about the workings of ChatGPT and neural networks in general here.
|
# ? Mar 7, 2023 01:54 |