How well can a dedicated RPG GenAI perform?

[FrogReaver]

For what it’s worth, I’d mostly ignore @Art Waring

From past interactions id summarize him as pretty much on the end ai any way possible bandwagon.

The actual policy for open AI’s API data is found here:
https://platform.openai.com/docs/concepts

Note that it explicitly says they don’t train on API data.

 

[Gorgon Zee]

So, just to be clear, my job handles sensitive personal information that is highly regulated, and involves using it in data science models including LLMs. I speak regularly to data protection lawyers and although I am not one, I advise them on technical matters and meet with them weekly. I spoke at an invited panel last week on the topic of Ethical AI together with other professionals, lawyers and subject matter specialists.

No APIs I am using train models based on the data I pass to them. Data may be retained for a short period by OpenAI and used only for human inspection purposes, but will then be deleted. This info can be found on their company site and I have gone over their staments with lawyers.

The vector database I built is local ot my machine. I considered a cloud-based version, but since there was no need for that, it was not necessary for me to work out what the company would do with my data. In fact, privacy is why I much prefer vector-indexing (RAG) solutions to fine-tuning. It’s much easier to control access.

I appreciate the concern, but this is in fact something I have strong experience with and have very high confidence that everything is kosher.

 

[Umbran]

From past interactions id summarize him as pretty much on the end ai any way possible bandwagon.

Mod Note:
Making it personal already on page 2, huh?

How about... you know... not? Please and thanks.

 

[FrogReaver]

I don’t know the actual ruleset at all so hard to know if the answers are actually good, but they are very detailed.

 

[CubicsRube]

All good. I just want to make sure your comments here are appreciated Gorgon.

 

[Gorgon Zee]

I don’t know the actual ruleset at all so hard to know if the answers are actually good, but they are very detailed.

LLMs are really good at pulling together facts from multiple sources into a coherent whole. They are effectively a highly effective pattern matcher with the ability to match meanings as well as pure syntax. However because they want to match patterns, they will often do so when patterns are not actually present. More technically, they choose some text that is low-probability given the context because it matches the desired pattern.

When I asked the LLM to give me information about cultures, once it starts on a given pattern it really wants to complete it, so when it cannot find the right answer in the text it is given, it adds in plausible text from its general knowledge. This is one source of “hallucination”.

A second issue is when you ask for a piece of info in the source, but it doesn’t know about it in the text, but does in general knowledge. Again,it fills in the gaps. You can see this in the “cultures” answer where it should be listing 7 specific adjective (with game meaning) for “distinctive features” but instead is giving a general description.

A good way to guard against this is to add text to the prompt instructions like “if you do not know the answer, respond with UNKNOWN. Do not give any response you cannot give a citation for”, or the like.

 

[Art Waring]

For what it’s worth, I’d mostly ignore @Art Waring

From past interactions id summarize him as pretty much on the end ai any way possible bandwagon.

Thanks for making this personal, when up to now, I have respectfully tried to follow the rules, and pay proper respect to other forum members.

Also, you have no right to put words in my mouth regarding my personal opinions on genai. I have stated numerous times now that I have no problem with the technology existing, provided that it is not used in a way that harms other people.

The fact that you have repeatedly missed this point, and painted the picture as to me being completely anti-ai is absurd. I have been vocal on the forums regarding genai because I have done a significant amount of research on the subject, and I don't appreciate your constant attempts to downplay the risks and the potential dangers.

I have even discussed individuals who are leading the forefront of ai development, especially in regards to ai alignment and safety. Just go look at Conner Leahy's company Conjecture and tell me that I am 100% anti-ai. Did you even bother to look at any of the evidence I posted in other threads? I think this proves that you never give me the same consideration that I have afforded you for a long time.

Thanks dude.

 

[FrogReaver]

Thanks for making this personal, when up to now, I have respectfully tried to follow the rules, and pay proper respect to other forum members.

You should probably still do that. I should do better at that.

Also, you have no right to put words in my mouth regarding my personal opinions on genai. I have stated numerous times now that I have no problem with the technology existing, provided that it is not used in a way that harms other people.

So let me ask instead of assume. Do you believe there is any way in which it can be used that doesn’t harm others? If so, an example or two would be helpful.

The fact that you have repeatedly missed this point, and painted the picture as to me being completely anti-ai is absurd. I have been vocal on the forums regarding genai because I have done a significant amount of research on the subject, and I don't appreciate your constant attempts to downplay the risks and the potential dangers.

Many of us have also done a lot of research, our conclusions just differ. I do believe there are risks and dangers. I don’t believe you accurately portray and/or contextualize them.

I have even discussed individuals who are leading the forefront of ai development, especially in regards to ai alignment and safety. Just go look at Conner Leahy's company Conjecture and tell me that I am 100% anti-ai. Did you even bother to look at any of the evidence I posted in other threads? I think this proves that you never give me the same consideration that I have afforded you for a long time.

This is starting to sound a bit too personal so I won’t be commenting in depth on this part. The short answer is yes, I read what you say and usually disagree with your take on its significance and your conclusions.

 

[Umbran]

LLMs are really good at pulling together facts from multiple sources into a coherent whole.

LLMs are really bad at knowing what is, or is not, a fact. LLMs don't understand the semantic content of the materials they are trained on.

 

[Gorgon Zee]

LLMs are really bad at knowing what is, or is not, a fact. LLMs don't understand the semantic content of the materials they are trained on.

That's true -- a better statement would be that LLMs are really good at pulling together information from multiple sources into a coherent whole. Whether that coherent whole is true is less certain.

However, if you provide information as input into the prompt, using techniques like vector-indexing / RAG, they are given a lot of weight. So if you do a good job of ensuring that the inputs are actual factual, then you strongly improve the odds that the output is. Except for the errors I made in extraction, the inputs I provide can be thought of as facts, but that is certainly not true in general.