What's the difference between AI and a random generator?

[Waller]

I'm not trying to be awkward I promise. I'm just trying to understand the technology.

We've all used random generators for years. Some random generators are totally random, others use inputs or some logic to construct their results. Some aren't even all that random!

If we ask an AI to write an NPC description or we ask a random generator to write one, other than the AI probably being better at it, what's the difference in process? What is the AI doing that a random (or non-random) generator isn't?

Is it just the data scraping part? If the AI did the same thing but didn't scrape data except for whatever data set the designer gave it is it then doing the same thing? Or is it fundamentally different? Or should we be looking askance at random generators too? There are generators which make maps and dungeons--is that similar to AI art generators, except for the scraping part?

I guess I'm asking is AI just a big fancy (semi-random) generator but with added data scraping?

 

[mamba]

They are entirely different. The random generator follows a set of predefined rules and picks predefined options from a pool at random within the confines of those rules.

No one knows what an AI does, seriously… we know how to create one, we know how to train it (reward it for good results, so it can adjust its internal logic to more reliably produce results we like), but how it actually determines what output to show us for a given input is a black box, no one knows how it gets to its results.

Without training on a wide dataset the AI is just a moron that randomly picks stuff and presents it as the result. It first has to learn what things are (eg when you ask for a picture of a penguin, for you to then get something that looks like a penguin and not a car or mountain lake). It does not really understand what a penguin is though, but it does understand that if you ask it for one and it shows you some upright black and white figure standing in snow, you are probably happy with that result.

A large dataset and extensive training are needed to get ‘reliable’ results and not just random garbage.

For us to understand how AIs get to their results there is the Explainable AI idea, which aims at having AIs that tell us how they arrived at their conclusions, but that is still pretty new

TL,DR: random generators follow rules established by the programmer, an AI comes up with and keeps refining its own rules based on how much we like its output for a given input

 

[dbm]

On the general topic of AI, check out this series of lectures from the Turing Institute:

 

[Umbran]

I guess I'm asking is AI just a big fancy (semi-random) generator but with added data scraping?

It is a good question.

No one knows what an AI does, seriously… we know how to create one, we know how to train it (reward it for good results, so it can adjust its internal logic to more reliably produce results we like), but how it actually determines what output to show us for a given input is a black box, no one knows how it gets to its results.

This is not a good answer. It is a false assertion, completely and entirely incorrect. I'm sorry, but you are just wrong.

So, let me try to give a layman's description of the tech currently under discussion....

Random Generator:
We probably all know this, but I will lay it out for contrast.

In a random generator, I create some list of outcomes, assign some probability to each, and then generate random numbers to pick from among those possibilities. Treasure tables in our DMG are a fine example.

Sometimes we can attach some logic to the process as well - like in a random dungeon map generator, room exits and entrances has to match up and be connected, and such.

But in all of these things, some person has directly created the list of outcomes and logic.

Generative AI:

The first thing to note is that "AI" isn't one clearly defined thing. It is a broad, non-technical description. What folks are up in arms about is generative AI, which is a specific technology. Other things we call "AI" may not work in the way I'm going to describe here.

Also, please note that I'm generalizing and simplifying here. They teach college level courses in how to do this, and I'm writing a few paragraphs on the internet.

First, you need a lot of examples of the sorts of things you want your system to produce, and you have to tag those examples with that you want the system to note is there. If you want it to produce images, you have to hand it a whole buttload of images, and the image of a penguin standing on a glacier by the ocean has to be tagged with "penguin" "glacier" and "ocean". The picture of the Dragonlance Companions of the Lance with be tagged with "Dwarf" "kender" "half-elf" "knight" "armor" and so on.

This is where the scraping comes in. The internet is loaded with images (and text) that is already tagged, telling you what is in it.

So, you get a huge number of examples - the more the better. Millions or even billions of examples, if you can get them. This is the "training set". One of the primary issues many of us have with Generative AI is that people who build these things take examples for their training sets and don't pay for using them - they pirate all the examples and aim to make a profit off the results. That's unethical - if you are making a profit based in part on someone else's work, they deserve a cut.

Anyway, so you pick up an example (let us say an image) from your training set, and tell the machine to give you an example of something in that image - say, a penguin. The machine doesn't know what the heck a penguin is, so it spits out random drek. You calculate the difference between what it put out, and the example (in the easiest-to-understand approach, you can calculate the differences pixel-by-pixel, but there are more sophisticated ways too). The machine takes that difference, adjust bits on the back end so that when you feed it "penguin" the result is close to the example you showed it.

You do this a bunch of times, with different examples, and eventually it "learns" to produce a credible image of a penguin. Note that it doesn't know what a penguin is. It just takes the input string "penguin" and crunches numbers and spits out the result. The big thing here is that despite Mamba's assertion, there are people who know what that number crunching looks like. I have a textbook on early forms of that number crunching on my shelf behind me. You can crack open the black box, and see why it responds the way it does.

There's a lot of subtlety, because you aren't just teaching it to respond with something that looks like a penguin. You are, at the same time, teaching it how to respond with something that looks like a poodle. And a puffin. And Paraguay. And pumpkin. And Marcel Marceau. and everything else in your training set.

But, anyway, eventually the machine responds with credible versions of the things you want. In a sophisticated system, there may be some randomization of elements, but the overall effect is more "stimulus & response" than "random generator".

 

[MatthewJHanson]

In addition to the technical difference people have mentioned, another element is that random generators are never going to put anybody out of a job.

We're already seeing companies use AI art in their products, and AI writing is all over clickbait sites. AI writing isn't good enough to replace writers in TTRPGs yet, but it may only be a matter of time.

You could put together a "random art generator" that say choose random hair, eyes, clothes, etc for some characters, but I doubt many people would use it for a published RPG product.

 

[mamba]

This is not a good answer. It is a false assertion, completely and entirely incorrect. I'm sorry, but you are just wrong.

nothing you wrote contradicted what I wrote (in my whole post, not just the part you quoted).

The machine takes that difference, adjust bits on the back end

that is the ‘we do not know what logic it uses’ part…

You described the training part in more detail than I did, that still does not mean that we have any idea what internal logic the AI uses to show you a penguin when you ask for one.

As you said there is a lot of detail and we keep it very short and simple out of necessity. The main difference I see is that I kept it even shorter

 

[dragoner]

Transparency for one, and usually the random generators I use are on paper.

 

[Epic Meepo]

that is the ‘we do not know what logic it uses’ part…

You described the training part in more detail than I did, that still does not mean that we have any idea what internal logic the AI uses to show you a penguin when you ask for one.

I worked for three years as a professional AI trainer at a tech start-up (using non-pirated data sets), and I have to largely disagree with your assertion.

Software engineers who program AI professionally absolutely do know the logic their AI uses to produce results, to the same extent they know the logic used by any non-AI software they write. Specifically, they understand the intended logic perfectly, but they need to debug and refine their code over time to ensure it implements that logic correctly. And even bugs introduced by human error produce results one can predict once those bugs are identified.

Generative AI isn't magic. It's a calculator that takes all the information in a massive data set as input instead of taking a few small numbers as input. Due to that difference in scale, it's tedious to track every single step performed by an AI, but tracking that internal logic or, more practically, a small subset thereof is entirely feasible. In some cases, making a log of certain logical steps taken by AI software is actually necessary for debugging or refining the code.

 

[Umbran]

that is the ‘we do not know what logic it uses’ part…

But, we totally know what logic it uses.

You described the training part in more detail than I did, that still does not mean that we have any idea what internal logic the AI uses to show you a penguin when you ask for one.

Again, we do know - I simply didn't describe it. It is typically a "neural network" with a known topology (number of layers and connection between layers). Each neuron has some number of inputs and outputs - The weight given to each input is developed by the training process, the calculation done on the weighted inputs is known, and determines the output to the next layer.

There is no mystery here. If one wanted to, one could take all the weights, and walk through the entire operation of of the network by hand. We don't, not because we don't know the logic, but because it would be, for people, a long and tedious process.

As you said there is a lot of detail and we keep it very short and simple out of necessity. The main difference I see is that I kept it even shorter

No, the main difference is that you were inaccurate.

 

[Art Waring]

But, we totally know what logic it uses.



Again, we do know - I simply didn't describe it. It is typically a "neural network" with a known topology (number of layers and connection between layers). Each neuron has some number of inputs and outputs - The weight given to each input is developed by the training process, the calculation done on the weighted inputs is known, and determines the output to the next layer.

There is no mystery here. If one wanted to, one could take all the weights, and walk through the entire operation of of the network by hand. We don't, not because we don't know the logic, but because it would be, for people, a long and tedious process.

Hi folks.

In the spirit of solidarity, I am presenting evidence from experts in the field of AI, in the hopes that we can better understand exactly what current ai models are, and how they work.

Yes, they do consist of layers, but the larger and more complex the model, the more "hidden" layers exist that are currently unable to be observed by humans. Conner Leahy goes into this in detail over his many articles, blogs, and videos on youtube. "Narrow" ai models do not typically have hidden layers (their terminology, not mine), but the advanced models all have hidden layers, including the GPT's and the upcoming GPT-5.

Deep Dive Final AI Report of 2023 from opensource.org

The Mysterious “Black Boxes” of AI/ML Models [Pg 13]
As depicted earlier in our brief description of deep learning, many AI models consist of “hidden layers” of algorithms that identify features and patterns in data and images without humans having to provide specific directions to the computer. Because many AI computations occur within the computer and essentially without human control and observability, AI models are often described as mysterious “black boxes.” We can’t always explain why an AI model decides what it decides. Therefore, we can’t prove that the machine is neutral and wise. This is often referred to as the explainability or interpretability problem.

A quote from Conner Leahy (CL 42:01):

We are at the point where we have these super complicated systems... and we just have no idea what's going on inside of them. It's not possible currently that we say "Oh, we see a failure case in our model. That's not good" and then we go into the model and we fix it. We can't currently do this."

A quote from Alek Tarkowski (AT 06:58)

The issues [with AI] are the same as with all the automation, but the ways of addressing them are a lot harder, because this beautiful symbol of the black box that hides things inside... complexity [of AI systems] basically makes them much harder to analyze, assess their impact, and so on.

And a final quote form the chapter on AI safety:

Explainability is just one of many AI challenges yet to be solved. Connor Leahy described several more (CL 32:50+) :
• The Alignment Problem: How do we get an AI system to do what we want and align with countless unspoken values and innumerable hypothetical scenarios that may occur? For example, if I train a robot to bring me a cup of coffee, how do I ensure it won’t run over grandma and my cat in the process?
• The Stop Button Problem: How do we make robots that are entirely indifferent to being shut off?
• The AI Security Problem: How do we keep AI systems safe from hackers?
AI technology is relatively young, and so it’s not surprising that the list of AI problems to be solved is substantial; nevertheless, some experts are optimistic, believing that we’ll figure out the solutions to these problems eventually.

If the leading experts cannot currently achieve either Explainability or Interpretability (it's one of the current issues they are trying to fix), then its still an ongoing problem.

If folks might not know who Conner Leahy is, he was the first person to reverse-engineer Chat GPT and create his own GPT-AI at the age of 24 (when at the time, OpenAI was closed-source and they did not share their datasets with anyone). He is a respected expert on AI technology and AI alignment, and is the current CEO of Conjecture AI, which is a company focusing on creating advanced ai systems, while also developing AI Alignment and AI Security (two very important fields of study, IMHO).

Thank you for your time,
A

P.S. EDIT: To restate my position, I am absolutely against the unethical use of ai tools, including data scraping without consent, credit, or compensation for artists. The music industry has proven that you can create ethical models and still pay artists, the fact that they refuse to pay illustrators for their work is a harmful exploitation of loopholes in the law. Artists deserve protection under the law, just like musicians, writers, and actors do.