What Connects Porn, Drugs, Dogs and GenAI?

Disclaimer:

The following article is intended for people who are interested in understanding how GenAI/LLM works. I am not an AI or LLM expert, what I write is based upon what I've learned myself and how I see these technologies, in order to make good use of these.        

Before we get to the "Generative" part of GPT, I'd like to make a couple of things clear. There is nothing new about LLMs (Large Language Models), in fact, we've all been training them for years. In order to understand better, how we've all been training, some of these models, completely for free - for over 20 years now. Let's take a little break from GPTs and LLMs - and talk about something else.

PornHub meets Medicine

Ok, I admit it - that is a very weird title, isn't it? what relation can PornHub have with Medicine? well, believe it or not, an urban legend tells a very funky story about prescription drugs, porn/adult content websites and ... a software company in the UK providing services to the NHS.

Putting aside the "sales" value of the above 1925 written prescription, let's have a look at the actual content of said prescription. Here we have an official paper, provided by a physician to a patient, to be handed to a pharmacist. Now, it is very clear that the written text is virtually undecipherable. In fact, there is an urban myth about people in medicine, having to go through a specific writing/reading class, in order to be able to read official papers.

Now, putting aside the urban myth, what does that have to do with PornHub? what does that have to do with AI and LLM training?

So, the story tells about a software company in the UK, that was hired in order to help the NHS digitize its archives and processes. They needed to process tens of thousands of paper and feed these into a computer. Now, we're talking about late 1990's - the height of the dot.com era (feels a little like the current AI bubble, no?). Now, the company need a way to decipher the somewhat impossible handwriting and OCR was not the same as today. So, a group of engineers came up with a novel idea - they'll take the documents and cut them into small chunks. The small chunks were relatively easy to read and decipher, but, they started feeding these as captcha for adult websites (via a different company). Every time someone was asked for a captcha, their database would "learn" a new option for a specific chunk. Once enough "teachers" affirmed the content of a specific "chunk", it was be stored and the system would move on to the next "chunk". Amazingly enough, it actually worked.

When you think about it - this is actually the buds of machine learning and pre-trained models - the concept is similar, the technology is ancient - but the result is very much identical.

You're not searching, you're training

In 2004, Google introduced the Google Search Autocomplete feature. Originally designed to make "search faster and easier" - I believe it had other uses. I'll be putting aside the more sinister use cases (after all, this is Google that adheres to the "do no harm" value - I hope). Today, when we look at the this feature, the database it collected from various searches can be used as a base for a language model - specifically, one that is aimed at searching information.

You see, one good thing about Google's search engine is that people don't really know how to communicate with it. So, people used natural speech to write questions. Unlike the old AltaVista and Yahoo engines, which were very much keywords based, Google was constantly learning and adapting itself. Google had to become really good at "understanding" what people are looking for - which means, they needed to implement a language model - long before anyone else did. Indeed, it wasn't as robust as an LLM - but the basic mechanics are in there.

Image captcha doesn't always work

These are probably the most annoying things ever, right? Have you ever wondered why some captchas will have "no choices", some will have very few and other may have 4-5 options to choose? The reason is simple, you are being used to train an image recognition model.

Previously we talked about language models, however, the same concepts can be applied to image recognition and image analysis as well. For example, let us examine the set of images below:

The link above is basically the search I conducted on google, which said "pomeranian white hair". Google has it wrong, because out of the 4 images, only 3 are pomeranian dogs, in fact, one of them is a long haired Chihuahua. Now, why did Google get it wrong? well, it didn't - Google looks for images based upon the meta-data, not based upon what the image actually looks like. Now, imagine that I would now take thousands of White Pomeranian images and feed them into an image analysis model - what I would eventually get, would be a set of vectors (similar that we used for the language model), that would jointly classify what a "White Pomeranian Dog" actually looks like.

So, let's put this to the test - using Midjourney, I've prompted it to do the following:

/imagine hyper realistic, pixel perfect, pointillist precision, 32k uhd, pomeranian dog, white hair.         

This was the result:

Ok, it seems like it has a good idea of what a Pomeranian dog is - but the concept of a white haired dog is a little "vague"? Let's try this again, and we'll try to be a bit more "specific". Although, it would seem that the 3rd image is relatively close to what we provided to the model.

/imagine hyper realistic, pixel perfect, pointillist precision, 32k uhd, fluffy white pomeranian dog.        

This was the result:

Ok, that seems a bit better - although, we're still a bit off. Why is that?

Well, potentially, the model used to generate the images was never trained specifically to generate Pomeranian images - or, there simply isn't enough data on this specific dog breed and color combination.

So, this is exactly where image based captcha needs our help. Have you ever noticed that sometimes, although you clicked all the right images on the captcha, it still thinks you're wrong? well, the reason is that probably, the one of the images wasn't yet properly trained, resulting in a false-positive result - in other words, get another captcha. And the training goes on and on and on...

Deadpool meets Monty meets Ryan Reynolds

Honestly, I love Ryan Reynolds - really, I find him funny and lighthearted - and I can't think of a movie that he participated in that I didn't like. There is simply something really simple and down-to-earth about how he portrays his characters.

When Marvel came out with Deadpool, the word was that the movie version of Deadpool was actually written and evolved around Ryan Reynolds previous roles, simply because they fit really well.

So, I decided it would be really funny to see what would a conversation between Deadpool, Monty (from the movie Waiting) and Ryan Reynolds himself may look like. Why would something like this work with a language model? very simple, all three characters have an ample enough of language data - so the generative parts (which we'll get to in Part 2 of this article series) will actually work the right way.

So, here is the prompt I gave ChatGPT 4 (in our case, a seriously loaded gun for these 3 extreme characters):

Create a discussion between the following characters: Deadpool (as portrayed by Ryan Reynolds), Monty (as portrayed by Ryan Reynolds in the movie waiting) and Ryan Reynolds himself. The discussion should be about the usage of Artificial intelligence and Generative technology in creating fictional characters.        

The result was - quite impressive, so here it goes:

Setting: A modern living room styled like a gamer's paradise, with LED lights and multiple screens showing AI-generated visuals. A coffee table is scattered with controllers and VR headsets.

Characters:

1. Deadpool: Clad in his unmistakable red and black costume, mask cheekily perched on his head, revealing an irreverent grin.

2. Monty: Wearing a slightly wrinkled server’s uniform, his expression is one of permanent, sardonic amusement.

3. Ryan Reynolds: Casually dressed in a sharp blazer and a t-shirt featuring a clever pun, ready to moderate with a twinkle in his eye.

---

Ryan Reynolds: Alright, let’s dive into a hot topic—AI’s role in creating fictional characters. Thoughts?

Deadpool (*with exaggerated enthusiasm*): Hell yeah! Let AI take a crack at it. Maybe it’ll spit out someone who doesn’t give a f--- about fourth walls or box office gross. Could it design someone with more sex appeal than me? Doubt it, but I’d love to see it try.

Monty: Oh, sure, because what we need is a machine deciding to amp up the angst or joy in our lives. Can’t wait for AI to program the existential dread of a never-ending shift at Shenanigans. Really looking forward to that innovation.

Deadpool: C'mon, Monty, don’t be such a downer. Imagine an AI that cranks out characters who can out-drink Thor and out-smart Stark. And maybe, just maybe, it can program them not to give a crap about anyone's approval.

Ryan Reynolds: It’s a compelling point. AI could take storytelling to new heights with analytics and algorithms. But where do we draw the line between creative help and creative takeover?

Monty: Right, because what’s more inspiring than a computer algorithm telling you your life’s story? Maybe it can calculate the exact moment my life turned into a perpetual joke.

Deadpool: Oh boo-f---ing-hoo. Look, AI could literally invent the next big anti-hero or, dare I say, a more twisted version of me. Bring on the AI-generated chaos, I say!

Ryan Reynolds: That brings us to the ethics of it all. If AI can generate a character, who truly owns that creation? Is it the programmer, the public, or the AI itself?

Monty: How touching. We’re worried about AI’s feelings now? Maybe we should get it a therapist too, talk about its mommy issues and fear of obsolescence.

Deadpool: Ha! But seriously, if an AI pumps out something that hits it big, I want in on that action. Copyright debates would be way more entertaining with me in the courtroom—just saying.

Ryan Reynolds: True, there's a need for collaboration. AI might enhance creativity, not replace it. We use tools to build things all the time, why not creative tools?

Monty: As long as it’s not making the drinks. Last thing I need is a robot judging my pour.

Deadpool: Screw the ethics, I’m in if it means faster sequel turnarounds and less screen time for the goody-two-shoes heroes. More explosions, less moral compass!

Ryan Reynolds: Maybe we're onto something. AI as a partner in the creative process could open up new avenues we've never even considered.

Monty: Sure, until it decides to replace us all. But until then, cheers to our future AI overlords.

Deadpool: Bottom line—if it can make Cable smile, it’s worth a shot. Let’s toast to AI: may it serve us well and never become self-aware.

---

Conclusion: The banter closes with a mix of humor and a touch of skepticism. The trio acknowledges the potential of AI in creativity while playfully jousting about the implications of such advancements, blending the serious with the satirical.

Honestly, I can actually imagine these characters in the above discussion - and the responses seem almost "legit" for them.

Conclusion - for now...

So, GenAI is impressive - but the results aren't always consistent or make sense. While the learning and modeling part makes sense, we still need to get a better understanding of how the "generative" part works, in order to truly understand why results are what they are.

In part 2 of this article series, we'll discuss the "Transformer" part of the GPT - which is the "thing" that is responsible for "generating" responses, in a (hopefully) predictable manner.