If you can privately build an LLM that:

1. Doesn't cause a disproportionate environmental impact in terms of energy and water consumption in relation to its output.
2. Isn't trained on plagiarized data.
3. Doesn't prioritize confidence over accuracy in its responses.
4. Doesn't hallucinate.
5. Is able to provide responses that couldn't be served up in equal capability by a search engine.
6. Can comprehend the scope of D&D just from the SRD.

Well you'll have a tool that not just would be novel to the D&D space, but probably the whole LLM space. This is assuming you're fully building your own model use your own training data.

But if your endeavor doesn't meet all of those criteria, well I personally wouldn't see any use for it, especially on points 1-3. If you're building off a framework from an existing LLM such as Chat-GPT, you'll likely have failed at one or more of the above points from the get-go.

"Since GPT already has been found to have pillaged from artists and writers, I am not sure that you can actually make an ethical AI based on the model. Thumbs down."

That's where you are wrong. GPT (Generative Pre-trained Transformer) is a concept not a proprietary holding. The data any given model is trained on the vast majority of which is in the public domain and by definition can't be stolen. Yes some bad actors have done some bad things, but that's not a legitimate reason to think that all such project are done with stolen data, when there is so much out there in the public domain. This include text, images, and sound all in the public domain. And WOTC already confirmed the project falls with in the Open Gaming License, I asked. So there's no Earthly reason for this project to steal data.

Quote from Duke_of_Vicenza >>

"Since GPT already has been found to have pillaged from artists and writers, I am not sure that you can actually make an ethical AI based on the model. Thumbs down."

That's where you are wrong. GPT (Generative Pre-trained Transformer) is a concept not a proprietary holding. The data any given model is trained on the vast majority of which is in the public domain and by definition can't be stolen. Yes some bad actors have done some bad things, but that's not a legitimate reason to think that all such project are done with stolen data, when there is so much out there in the public domain. This include text, images, and sound all in the public domain. And WOTC already confirmed the project falls with in the Open Gaming License, I asked. So there's no Earthly reason for this project to steal data.

(emphasis mine)

If the vast majority of the training data isn't stolen, the model is still based on stolen data.

Quote from scatterbraind >>

For what it's worth, Deepseek 3's recent release is kind of blowing the doors off the story we've been told about how energy-guzzling AI has to be.

That's if you assume everything we've been told about Deepseek is accurate. There's no real way to verify it, considering its source

How would it be useful? What would you have it do? These are the questions you would have to answer for people to find value in it. "Be an aid to DM's and Players" an aid that does what? And you need to consider the quality.
Many of the results of AI i have seen, have been atrocious in quality, or have felt off in ways that broke immersion.

The dislike for AI isn't just some neo-ludite impulse, or moral panic. It just has not proved itself to be of any worth to a lot of people. The fact it is constantly being pushed on us when we just want to be able to do our thing and be creative.

Quote from Davyd >>

If you can privately build an LLM that:

1. Doesn't cause a disproportionate environmental impact in terms of energy and water consumption in relation to its output.
2. Isn't trained on plagiarized data.
3. Doesn't prioritize confidence over accuracy in its responses.
4. Doesn't hallucinate.
5. Is able to provide responses that couldn't be served up in equal capability by a search engine.
6. Can comprehend the scope of D&D just from the SRD.

Well you'll have a tool that not just would be novel to the D&D space, but probably the whole LLM space. This is assuming you're fully building your own model use your own training data.

But if your endeavor doesn't meet all of those criteria, well I personally wouldn't see any use for it, especially on points 1-3. If you're building off a framework from an existing LLM such as Chat-GPT, you'll likely have failed at one or more of the above points from the get-go.

 Some of you standards are impossible to know or even achieve.

1 is impossible to know and probably impossible to achieve.

2 is easily achievable since there is plenty of data in the public domain

3 is achievable or at the very least can be mitigated

4 can be minimized but not completely eliminated

5 AI already does this and is oft time easier and faster to use than a search engine

6 Unrealistic and impossible with current technology. Comprehension can only come from consciousness and AI is not capable of that.

Quote from Duke_of_Vicenza >>

Quote from Davyd >>

If you can privately build an LLM that:

1. Doesn't cause a disproportionate environmental impact in terms of energy and water consumption in relation to its output.
2. Isn't trained on plagiarized data.
3. Doesn't prioritize confidence over accuracy in its responses.
4. Doesn't hallucinate.
5. Is able to provide responses that couldn't be served up in equal capability by a search engine.
6. Can comprehend the scope of D&D just from the SRD.

Well you'll have a tool that not just would be novel to the D&D space, but probably the whole LLM space. This is assuming you're fully building your own model use your own training data.

But if your endeavor doesn't meet all of those criteria, well I personally wouldn't see any use for it, especially on points 1-3. If you're building off a framework from an existing LLM such as Chat-GPT, you'll likely have failed at one or more of the above points from the get-go.

 Some of you standards are impossible to know or even achieve.

1 is impossible to know and probably impossible to achieve.

2 is easily achievable since there is plenty of data in the public domain

3 is achievable or at the very least can be mitigated

4 can be minimized but not completely eliminated

5 AI already does this and is oft time easier and faster to use than a search engine

6 Unrealistic and impossible with current technology. Comprehension can only come from consciousness and AI is not capable of that.

If it's impossible for a new tool to achieve the standards set by an existing tool, then there's no reason for it to exist.

Quote from Duke_of_Vicenza >>

I have begun development and training on a GPT model , built from scratch. With the ultimate goal of it becoming and aid to players and Dungeon Masters. When ready it will be released under the Open Gaming License. as DnD-GPT. It will be fully open source. 

What do you all think? 

 Are you here to gain an understanding of what posters on this forum generally think, as your OP implied, or you are you just here to argue with anyone who has concerns about AI?

Quote from wagnarokkr >>

Quote from Duke_of_Vicenza >>

"Since GPT already has been found to have pillaged from artists and writers, I am not sure that you can actually make an ethical AI based on the model. Thumbs down."

That's where you are wrong. GPT (Generative Pre-trained Transformer) is a concept not a proprietary holding. The data any given model is trained on the vast majority of which is in the public domain and by definition can't be stolen. Yes some bad actors have done some bad things, but that's not a legitimate reason to think that all such project are done with stolen data, when there is so much out there in the public domain. This include text, images, and sound all in the public domain. And WOTC already confirmed the project falls with in the Open Gaming License, I asked. So there's no Earthly reason for this project to steal data.

(emphasis mine)

If the vast majority of the training data isn't stolen, the model is still based on stolen data.

 Nonsense. A model built from scratch including the training data taken from the public domain and data used under a legitimate License can not be stolen. 

Glad to see most of us are on the same page about AI. I don't want it hijacking our creativity or interfering with our own inventive and creative process.

Quote from wagnarokkr >>

Quote from scatterbraind >>

For what it's worth, Deepseek 3's recent release is kind of blowing the doors off the story we've been told about how energy-guzzling AI has to be. There is potential for massive efficiency improvements to be made, they are just not prioritized by the big players in the US. I do believe there's a possible future where AI results in overall reduction in energy consumption, if we can avoid destroying ourselves and others long enough to get there.

The idea that heavy use of generative AI could actually reduce energy consumption seems like pure fantasy to me, but if it actually happens I'll be pleasantly surprised.

Then we'll have a plagiarism machine that doesn't do anything useful, instead of a plagiarism machine that doesn't do anything useful while destroying the environment. Yay.

 I agree with Duke that AI is essentially a database that talks. When the data you put in that database is large swaths of the internet, you get the plagiarism machine you are familiar with. It can't reliably answer factual questions because it's full of conflicting information. "Garbage in, garbage out" has been a programming mantra for decades, and it applies here too.

What AI is good at is ingesting massive amounts of information and synthesizing that data. So for example, you input the entire energy infrastructure grid of the US and it outputs the areas for upgrades and fixes in order of priority. It could also suggest improvements or alternate routes that would minimize waste or loss from transfer. These suggestions would not be politically or economically biased. Of course, this requires a similarly unbiased public works machine that would then carry out the work, which may be even more idealistic than a productive use for AI.

It can also be applied to fields like material science to find better conducting materials or more efficient solar panels. My friend works in cancer research, and a huge part of their work is just churning through thousands of possible molecular configurations to find new treatments. AI is very good at this thing and has accelerated their work by quite a lot.

AI is ultimately just a tool that finds patterns. It's not inherently good or bad, but it can be used for good or bad. Most of the egregious waste and cultural damage is due to that tool being released into an overclocked capitalist engine that is far more concerned with profits and stock prices than concepts like ethics or safety.

I am currently a member of a lawsuit involving specific AI tools for plagiarism and copyright violation, and that I will not be using the tool regardless as long as that lawsuit continues.

 Ok, that's that.

Now, I'm not against the use of generative AI tools for the production of materials. I have no problem using it as a supplement and a small part of a process; I like being able to do a really rough sketch in my kindergarten level ability, plop it into a tool, and get an output that is significantly more pleasing to my eye.

Being able to take a costume design and give it a bit of life, or a patched together composite image and make it more uniform is all very, very ideal.

I wouldn't use it to create anything I am already competent at -- so there's nothing in my 650 page lore book where the tables, text, concepts, and ideas are generated, but there are a few images that are and derive from sketches i did with that aforementioned lack of any real ability. 

Using them with D&D, though, requires awareness of a specific purpose and goal.

1 - GAI tools operate on a most common denominator basis. As a result, anyone using them to generate things that are outside those common denominator aspects will have to put more effort into them in order to achieve a result, and the more effort needed, the less likely it is to be used for that purpose. Bluntly, the most simple example of this is that if you want a Black Woman Lesbian Warrior with a sword, it is going to require a whole lot more effort to generate one of those.

2 - D&D is a series of complex systems that require an incredible number of potential options and tons of freedom of choice -- and with 60% of all games happening outside any standard official setting, the number of potential options is going to be far larger than one can effectively predict or model for.

3 - To use it to improve character generation, you will have to find a way to enable and disable options without invoking their use in a prompt, or you'll need a prompts space of probably around 30,000 characters.

4 - You might be thinking of doing so for Adventures or Campaigns -- well, that's cool, but you'll have to be able to develop web, linear, tree, and other styles of approach in order to approximate it, and the system is almost always going to generate defined result scenarios. That is, it will railroad the hell out of people because it will always default to the least common denominator. That's not even paying attention tot he lack of broader term awareness -- you can ask it to generate a full campaign and it will do outlines of adventures, and then you can ask it for those adventures singly and it will break them down -- but it will note link them together or be able to hold onto a commonality or basis among all of them, even with a prompt to do so, because that's a cognitive function outside the scope of its ability set.

5 - It will not be able to generate genuinely usable original monsters, Classes, sub classes, Species, or similar, because it does not understand archetypes. That specific tool.

So, that's my thoughts.

Quote from AEDorsay >>

I am currently a member of a lawsuit involving specific AI tools for plagiarism and copyright violation, and that I will not be using the tool regardless as long as that lawsuit continues.

All the public domain models are based on scraped data. Most of them were derived from collegiate research efforts that developed the sets to establish the baselines -- and they licensed those models to the big companies (who then violated the terms of those agreements), and those are the foundations of what became the "black box" that none of the current versions of them use any longer (because of lawsuits and the theft/plagiarism thing). If you can  identify all the training material sources to be able to provide compensation for all the creators, or can prove that all of the data has a license (and it is important to note that if you use anything not in Creative Commons, that includes any and all D&D materials) that explicitly permits use in LL/LI/LA models (because otherwise you are already competing with the IP owners), you might be able to effectively use this. But that's a tall order. It is also important to note that the product of generative systems is not copyrightable at this time, even if significantly altered by an artist using digital tools (because digital tools, including photoshop, are caught up in this). The biggest concern is that you want to train it on data about D&D -- and you almost certainly have not acquired permission from Hasbro to do this. If you scrape DDB, you violate the terms of your agreement, and holy heckfire the legal risk there is not one I even want to consider, lol. 

 Ok, that's that.

Now, I'm not against the use of generative AI tools for the production of materials. I have no problem using it as a supplement and a small part of a process; I like being able to do a really rough sketch in my kindergarten level ability, plop it into a tool, and get an output that is significantly more pleasing to my eye.

Being able to take a costume design and give it a bit of life, or a patched together composite image and make it more uniform is all very, very ideal.

I wouldn't use it to create anything I am already competent at -- so there's nothing in my 650 page lore book where the tables, text, concepts, and ideas are generated, but there are a few images that are and derive from sketches i did with that aforementioned lack of any real ability. 

Using them with D&D, though, requires awareness of a specific purpose and goal.

1 - GAI tools operate on a most common denominator basis. As a result, anyone using them to generate things that are outside those common denominator aspects will have to put more effort into them in order to achieve a result, and the more effort needed, the less likely it is to be used for that purpose. Bluntly, the most simple example of this is that if you want a Black Woman Lesbian Warrior with a sword, it is going to require a whole lot more effort to generate one of those.

2 - D&D is a series of complex systems that require an incredible number of potential options and tons of freedom of choice -- and with 60% of all games happening outside any standard official setting, the number of potential options is going to be far larger than one can effectively predict or model for.

3 - To use it to improve character generation, you will have to find a way to enable and disable options without invoking their use in a prompt, or you'll need a prompts space of probably around 30,000 characters.

4 - You might be thinking of doing so for Adventures or Campaigns -- well, that's cool, but you'll have to be able to develop web, linear, tree, and other styles of approach in order to approximate it, and the system is almost always going to generate defined result scenarios. That is, it will railroad the hell out of people because it will always default to the least common denominator. That's not even paying attention tot he lack of broader term awareness -- you can ask it to generate a full campaign and it will do outlines of adventures, and then you can ask it for those adventures singly and it will break them down -- but it will note link them together or be able to hold onto a commonality or basis among all of them, even with a prompt to do so, because that's a cognitive function outside the scope of its ability set.

5 - It will not be able to generate genuinely usable original monsters, Classes, sub classes, Species, or similar, because it does not understand archetypes. That specific tool.

So, that's my thoughts.

 

 

 Thank you for your thoughts and candor. But do not confuse the Model which is just software code, and the training data, they are two separate things. Big companies with armies of lawyers  are not the same as one individual who isn't seeking to profit . And any data the someone makes a claim against is easily resolved as they come to light. So copyright issues are  unlikely. As for the Intellectual Property of WOTC, that is covered under their Open Gaming License and I most certainly have contacted them about it.  I don't need to scrape the DnD Beyond Website, nor do I have a desire to.  I am well aware of the limitations of AI and don't expect this model to be a digital all knowing oracle for D&D. Anyone that expects that are kidding themselves. And I would expect any potential user if any to understand this.

Bad idea even if you didn't stick to the free D&D stuff, but with the very limited free content there really wouldn't be enough to "trian" "ai". Heck if you pirated everything for D&D there isn't enough to do that, it is just a really big spreadsheet at best. As a pet project have at it; as a business model it doesn't even come close to viable.

For the record what we are being sold as ai is not even close to ai, it is however a wonderful tool for piracy and plagiarism and has no place in my life i wish it was 100% illegal.

Quote from Duke_of_Vicenza >>

I have begun development and training on a GPT model , built from scratch. With the ultimate goal of it becoming and aid to players and Dungeon Masters. When ready it will be released under the Open Gaming License. as DnD-GPT. It will be fully open source. 

What do you all think? 

Absolutely not.  Have a lovely day.

Quote from Erik_Soong >>

Since GPT already has been found to have pillaged from artists and writers, I am not sure that you can actually make an ethical AI based on the model. Thumbs down.

Not the attitude on this I would've expected from a Soong!

Quote from Duke_of_Vicenza >>

An alcoholic knows that they are an alcoholic, it's not the alcohol's fault that the alcoholic doesn't care what it is doing to them.  

I am going to touch on the thread in a second, but this needs to be addressed first.  As an expert on the subject of addiction, I simply cannot allow this comment to stand.  This is an incredibly reductive, harmful, and flagrant stereotype of alcoholism that fails to acknowledge two utterly obvious realities of addiction.  First, no, many addicts do not know they are an addict--that is why the First Step is admitting you have a problem.  Second, many addicts absolutely do care about what the substance is doing to them - they simply cannot stop due to a mix of physiological and psychological reasons beyond their control.

Now, why do I bring this up?  In a single sentence, you managed to perpetuate two extremely harmful stereotypes about addiction.  Stereotypes that perpetuate stigmas; stereotypes that keep people from seeking help... Stereotypes that kill people.  Given how casually you made this statement, I seriously doubt it even crossed your mind that you were furthering longstanding harms.

Which brings us to AI.  AI is only as good as the data set it is trained on--and there are countless, countless studies showing that AI tends to take on the implicit biases of the information it is being trained with.  You clearly have implicit biases - and you demonstrated an extremely harmful one on this thread without realizing it.

I do not really care what you do with AI - your project does not interest me in the slightest.  But D&D is already a game full of harmful tropes - a side effect of being created by a self-proclaimed bigot.  The last thing this game needs is a digital tool created by someone who so cavalierly makes harmful statements.  That indicates to me that you (a) will very likely taint the AI algorithm with your own biases by not holding up yourself to self-evaluation and (b) will not recognize problematic elements of others' creations when training the AI on them.  That all leads to the situation we see too often with AI - the black box of AI producing bigotry.

Do what you will, but I do sincerely hope you hold up your AI modelling to a higher standard than you did your posts here.

Quote from wagnarokkr >>

Quote from Erik_Soong >>

Since GPT already has been found to have pillaged from artists and writers, I am not sure that you can actually make an ethical AI based on the model. Thumbs down.

Not the attitude on this I would've expected from a Soong!

 Whenever someone gets the reference, it always makes me smile. +1