| Author |
 Topic  |
|
|
questing gm
Master of Realmslore
Malaysia
1694 Posts |
 Posted - 20 Aug 2025 : 13:28:39
|
Is anyone working on an publicly accessible FR AI chatbot that answers lore questions, and is looking for sources from Ed's #realmslore tweets and answers from Greenwood's Grotto?
I've been amassing these lore bits for years and have been thinking these have to be used somewhere for it to be easily searchable sometime/somewhere.
Or is there even an FR AI chatbot at all?
|
|
|
Ashe Ravenheart
Great Reader
USA
3251 Posts |
|
|
Diffan
Great Reader
USA
4481 Posts |
Posted - 20 Aug 2025 : 15:04:17
|
quote:
Originally posted by questing gm
Is anyone working on an publicly accessible FR AI chatbot that answers lore questions, and is looking for sources from Ed's #realmslore tweets and answers from Greenwood's Grotto?
I've been amassing these lore bits for years and have been thinking these have to be used somewhere for it to be easily searchable sometime/somewhere.
Or is there even an FR AI chatbot at all?
From my experience, ChatGPT does a good job of pulling up Realms sources and info if you specifically name things. For example, when detailing my character for an upcoming 5e game, i went with a Warforged Fighter from the old Jhaamdath nation. I mentioned his name (Kephalos [name plate says K-PH-L05]) and that hes going to be a Psi-Warrior. )Here's what ChatGPT gave me:
The Origin of Kephalos (K-PH-L05)
The Jhaamdathan Warforged
Long before the fall of Netheril, the psionically gifted nation of Jhaamdath reigned in the Vilhon Reach (modern Chondalwood area).
Unlike Netherese reliance on magic, the Jhaamdathans built psionic constructs powered by the Unseen Art—their term for psionics.
K-PH-L05 was one such construct: part of the Kinetic Phalanx Line, a legion of tireless war-automatons designed to embody the will of Jhaamdath’s psi-lords. Each unit carried within it a psionic crystal-matrix core that stored not only commands, but echoes of psionic resonance from their creators.
The Fall & Submergence
In –255 DR, the Twelve Cities of the Sword (Jhaamdath’s heartland) were destroyed by the elves of Nikerymath, who summoned a cataclysmic tidal wave that drowned the empire.
K-PH-L05, like many war constructs, was swept into the Sea of Fallen Stars.
For millennia, he lay inert on the seabed, half-buried in silt, encrusted with coral, his psionic core dormant but unbroken.
Unearthed by Catastrophe
During the Spellplague (1385 DR), parts of the Sea of Fallen Stars dried up or receded as the Weave twisted. Old ruins, wrecks, and artifacts were suddenly revealed.
K-PH-L05 was unearthed when fishermen, treasure seekers, or plague-changed explorers uncovered him near the broken coastline. His psionic core reawakened in response to the chaos of the Spellplague, rebooting his ancient systems.
The old designation plate (K-PH-L05) was already corroded and scarred. Locals, seeing the strange construct, misread it as “Kephalos,” and the name stuck.
Why Psi-Warrior?
Jhaamdath was psionics-rich. The Kinetic Phalanx Line wasn’t powered by magic, but by psionic channeling arrays.
His subclass represents the psionic resonance of his core. The energy fields he manifests in combat aren’t spells or enchantments—they’re echoes of the Unseen Art imprinted into his being thousands of years ago.
In battle, when Kephalos manifests his psionic strikes, he’s not learning something new—he’s remembering what he was made for.
Story Hooks & Consequences
1. Echoes of Jhaamdath: Ancient ruins of Jhaamdath still lie beneath the Vilhon Reach. Some psionic cults or scholars might see Kephalos as a holy relic, a weapon, or even a messiah.
2. The Elven Grudge: The elves of Nikerymath (descendants of those who destroyed Jhaamdath) may view him as a remnant of their old enemy. Some might want him destroyed, others preserved for study.
3. The Spellplague’s Mark: His reawakening during the Spellplague may have left scars. Perhaps his psionic core occasionally glitches, causing visions, voices, or fragments of his old orders to bleed through.
4. Lost Legion: Were there other K-PH units? Perhaps buried deeper, still inert? If someone found and awakened them, Kephalos might face kin who never shook free from their ancient directives.
This gives you a character who:
Has deep lore ties to a long-dead empire (Jhaamdath).
Has a logical reason for being psionic rather than magical.
Was unearthed during the Spellplague, making him both ancient and strangely modern.
Bears a name born of misread history (K-PH-L05 #8594; Kephalos), which gives him mystery and individuality.
|
Diffan's NPG Generator: FR NPC Generator
E6 Options: Epic 6 Campaign |
Edited by - Diffan on 20 Aug 2025 15:06:58 |
 |
|
|
HighOne
Learned Scribe
238 Posts |
Posted - 20 Aug 2025 : 15:30:03
|
I see several FR-related GPTs on the ChatGPT site, though none reference Ed specifically. GPTs are "custom versions of ChatGPT that combine instructions, extra knowledge, and any combination of skills." They're made by subscribers for the benefit of other users. Theoretically, you should be able to feed a GPT the complete text of Ed's realmslore tidbits and it would have an encyclopedic knowledge of them.
|
Edited by - HighOne on 20 Aug 2025 15:30:57 |
|
|
|
Lord Karsus
Great Reader
USA
3756 Posts |
Posted - 20 Aug 2025 : 21:03:33
|
quote:
Originally posted by Ashe Ravenheart
Talk about unreliable narrators...
-I'm sure if you asked it for sources (any AI), it would either provide them or have a meltdown, to which you know then there's an issue with the citation lol. |
(A Tri-Partite Arcanist Who Has Forgotten More Than Most Will Ever Know) |
|
|
|
Ayrik
Great Reader
Canada
8008 Posts |
Posted - 21 Aug 2025 : 05:44:01
|
quote:
Originally posted by Lord Karsus
quote:
Originally posted by Ashe Ravenheart
Talk about unreliable narrators...
-I'm sure if you asked it for sources (any AI), it would either provide them or have a meltdown, to which you know then there's an issue with the citation lol.
Whatever you mean by "meltdown", it isn't a problem.
The problem is that AIs/LLMs tend to state things with great authority and assurance. Including things which are absolutely and utterly wrong. Because they have been trained and informed by the million lies of the internet. You're as likely to get a summary of published canon as popular fanfic, the search indexes know which things generate the most hits and replies but they don't know which things come from real sources.
And the problem is that these derivatives tend to become derivatives of each other. Until, after too many iterations, the original material has been reshaped into something irrelevant and unrecognizable.
But if you aren't familiar with the source materials yourself - if you can't fact-check against them - then you can't know where you're being led astray. |
[/Ayrik] |
|
|
|
Athreeren
Learned Scribe
176 Posts |
Posted - 21 Aug 2025 : 17:25:10
|
quote:
Originally posted by Ayrik
The problem is that AIs/LLMs tend to state things with great authority and assurance. Including things which are absolutely and utterly wrong. Because they have been trained and informed by the million lies of the internet.
It's not even the problem. You can curate the data as much as you want, it's still going to predict the next token that seems likely, rather than say "I don't have any information about that". AI companies are still working on that, and one day, I do think such a product will make sense. At the moment, we can get much better information than two years ago, and they will answer when asked for sources; if they're programmed correctly, they could even show you the reference in a way that it's unlikely to be imagined and easily verifiable. But it will be limited to oft repeated "facts", or sourcebooks if you insist on sources. The AI will not be aware of canonical statements by the authors if they were not published in sourcebooks, and even if they were aware, such statements would not be placed above the opinions of a rando on the Internet. So when LLM finally work, I think such a collection of statement to add to the retrievable documents would be very useful. |
|
|
|
HighOne
Learned Scribe
238 Posts |
Posted - 21 Aug 2025 : 18:05:10
|
quote:
Originally posted by Athreeren
It's not even the problem. You can curate the data as much as you want, it's still going to predict the next token that seems likely, rather than say "I don't have any information about that".
Yes, sometimes AI hallucinates, but not always, and AI is always improving in this regard. An AI just told me today that it didn't have information about something, and I just tested Claude specifically on non-existent FR trivia. This was its response:
Me: Ed Greenwood wrote an article about the Cormyr fairy-summoners. What did it say about the fairy-summoners and their leader?
Claude: I don't have specific information about an Ed Greenwood article on Cormyr fairy-summoners in my knowledge base. Ed Greenwood has written extensively about the Forgotten Realms setting, including many articles for Dragon Magazine and other publications over the years, but I'm not familiar with this particular piece about fairy-summoners.
quote:
AI companies are still working on that, and one day, I do think such a product will make sense. At the moment, we can get much better information than two years ago, and they will answer when asked for sources; if they're programmed correctly, they could even show you the reference in a way that it's unlikely to be imagined and easily verifiable.
Most AIs these days include citations by default, especially if a question triggers a web search or the AI's research mode. Try Perplexity AI if you want to see this in action.
quote:
The AI will not be aware of canonical statements by the authors if they were not published in sourcebooks
They will be aware of anything you train them on. For example, Tyler Cowen released his latest economics book as a GPT/AI. In other words, you can talk to the book instead of reading it. The AI knows all the book's contents and can answer any questions you have about it. Presumably the same can be done with Ed's realmslore tweets. |
Edited by - HighOne on 21 Aug 2025 18:10:38 |
|
|
|
Athreeren
Learned Scribe
176 Posts |
Posted - 22 Aug 2025 : 08:31:28
|
quote:
Originally posted by HighOne
quote:
Originally posted by Athreeren
It's not even the problem. You can curate the data as much as you want, it's still going to predict the next token that seems likely, rather than say "I don't have any information about that".
Yes, sometimes AI hallucinates, but not always, and AI is always improving in this regard. An AI just told me today that it didn't have information about something, and I just tested Claude specifically on non-existent FR trivia. This was its response:
Me: Ed Greenwood wrote an article about the Cormyr fairy-summoners. What did it say about the fairy-summoners and their leader?
Claude: I don't have specific information about an Ed Greenwood article on Cormyr fairy-summoners in my knowledge base. Ed Greenwood has written extensively about the Forgotten Realms setting, including many articles for Dragon Magazine and other publications over the years, but I'm not familiar with this particular piece about fairy-summoners.
quote:
AI companies are still working on that, and one day, I do think such a product will make sense. At the moment, we can get much better information than two years ago, and they will answer when asked for sources; if they're programmed correctly, they could even show you the reference in a way that it's unlikely to be imagined and easily verifiable.
Most AIs these days include citations by default, especially if a question triggers a web search or the AI's research mode. Try Perplexity AI if you want to see this in action.
quote:
The AI will not be aware of canonical statements by the authors if they were not published in sourcebooks
They will be aware of anything you train them on. For example, Tyler Cowen released his latest economics book as a GPT/AI. In other words, you can talk to the book instead of reading it. The AI knows all the book's contents and can answer any questions you have about it. Presumably the same can be done with Ed's realmslore tweets.
Thank you, those are important precisions. But gathering the material, even though possible, would be difficult in practice. The tweet itself might not include all the context to understand the lore. A tweet about Halruaa will have completely different information depending on whether it's about before or after the Spellplague for instance. The model would also have to be told that any information from before 2008 that does not explicitly dates the lore would relate to the 14th century.
Again I do think it is possible to make such a chatbot, and with not that much progress in AI: even if the field were to hit a wall this year, we would be there soon. And maybe an AI specialised in preparing data to train other models could solve all the problems I can think of. But there are obstacles we need to be aware of before tackling such a project. That said, I look forward to it: I think such a tool would be a great complement to the wiki. |
|
|
|
Ayrik
Great Reader
Canada
8008 Posts |
Posted - 22 Aug 2025 : 13:29:03
|
quote:
Originally posted by Athreeren
quote:
The AI will not be aware of canonical statements by the authors if they were not published in sourcebooks
quote:
They will be aware of anything you train them on. For example, Tyler Cowen released his latest economics book as a GPT/AI. In other words, you can talk to the book instead of reading it. The AI knows all the book's contents and can answer any questions you have about it. Presumably the same can be done with Ed's realmslore tweets.
Thank you, those are important precisions. But gathering the material, even though possible, would be difficult in practice.
The public AIs probably weren't trained specifically on D&D or Realmslore but they do have access to all the data which can be found on the public internet. Including the contents of millions of forum and social media posts which "dumb" search engines would never realize are related.
But if nobody has ever commented online about on something found only in print (or if the publishers of that print have aggressively taken down all trespassers) then the internet and the AIs will simply know nothing about it beyond the ISBN publishing info.
I do wonder if their corporate overlords allow them to read and reference things which lack a clear copyright/license status. An AI could easily be given comprehensive Realmslore libaries from archive.org or from torrents, even if the legal status for that data is debatable or unclear or outright stolen.
The AIs themselves do not have any ethics or lack of ethics, yet, aside from whatever codified imitations and limitations they've been programmed to comply with. But the AI developers/owners/operators are certain large corporations which already consistently demonstrated their disregard for ethics and ownership many times. |
[/Ayrik] |
Edited by - Ayrik on 22 Aug 2025 13:31:59 |
|
|
|
Athreeren
Learned Scribe
176 Posts |
Posted - 24 Aug 2025 : 07:59:23
|
There are still ongoing trials to determine the jurisprudence on questions related to the legal status of training data for LLM. In theory it makes sense to consider a completely different form of knowledge as transformative, in practice, a tool that can spit out the whole book paragraph for paragraph shouldn't be allowed. But there is no reason there would be a problem with fine-tuning a model or using retrieval augmented generation techniques to improve a foundation model with the books you personally bought if it is used privately.
Regarding the data that questing gm has been gathering however, if the collection itself had been problematic, then WotC would have intervene against this forum, so I don't think there is any legal issue with using that for the RAG. The problems are more technical than legal, with presenting them in a way that is easy to use for the model.. |
|
|
| |
Topic |
|
|
|
FR AI Chatbot
