Right!

It's time for part 2 of my series on ChatGPT (and AI+LLM's in general).

Let's begin by 'eating our vegetables', ie that in order to have the juicy discussion about ChatGPT we need to understand some potentially boring and confusing topics. This is necessary because otherwise we can't really have a meaningful discussion about the whole topic.

When I say ChatGPT am often referring to the breakthrough product by OpenAI, which holds the world record for ... well let me just quickly quote wikipedia:

By January 2023, ChatGPT had become the fastest-growing consumer software application in history, gaining over 100 million users in two months.[6][7] As of 2025, ChatGPT's website is among the 5 most-visited websites globally,[8][9] and has over 800 million active weekly users.[10] It has been lauded as a revolutionary tool that could transform numerous professional fields.

However most of the time the same applies to Gemini, Claude, Grok, really anything that uses what is called an LLM, which brings me to the main reason I am talking about it. I am not an expert on how they work, very few people in the world really are, as far as I understand it, they kind of got invented almost by accident or at the very least the people at OpenAI who were tinkering with the idea of this 'LLM' had NO idea that it would explode into the world changing beast that it has become. In fact they didn't even quite invent it, they used something publicly released by Google called ...

You can literally read the paper on the place where these kind of things are published, arxiv. Appently, this is going to be a series of posts, because it has been demanded politely requested by someone who shall remain Anonymous via a WhatsApp voicenote, that there needs to be several of these.

Anyways, we don't need to quite understand how it ALL works, all we need to understand is that they all have a similar underlying way that they work, and so what applies to one 'model' ie ChatGPT can also apply to your flavour of choice, perhaps Claude, perhaps Gemini. ChatGPT is certainly the most popular one and it's CEO, Sam Altman is certainly the best at publicity and gathering attention to the product.

The Good...

I made the above entire website, a portfolio of some of my inventions, written in both English and translated to Danish, in one fucking prompt using Claude Code CLI, which is kind of like a ChatGPT but he lives inside a box that can access anything you want him to in your computer. It/he can do anything someone else could like clicking around, making new files, signing up for a web hosting service, reading documentation on how to make cool shit, the whole thing. Basically my flipping job and it did the whole thing in one sentence, okay it was quite a long sentence, but you can literally click through that image or this link and check out that site and be blown away, even the prompt to build that site is on that site. How Meta!

It took me maybe 10 minutes. 3 years ago such a thing probably would have taken me a week.

Tricked you!

We cannot get into the good parts of AI, of which there are many, before I take this pedastal and high horse and become a total party pooper, first showing the most massive flaw that anyone and everyone using them MUST understand. Almost like a drivers license, you wouldn't feel safe in a world where just anyone can drive a car, you need to first have a driver's license.

As an aside, in Denmark, they have a list of countries that they recognise as valid drivers licences and when one moves here from one of those countries they essentially just pop over to the dansk version of home affairs and say, here is my previous drivers license, gib me a danskish one, ah tusind tak, kay-thanx-bai. South Africa is most decisively NOT on that list, and if you want to drive in Denmark you have to follow these steps.

0) Find your South African drivers license, which can be printed by only one random printer located in Bloem
1) Flush said card down the nearest toilet.
2) Join the queue with all those teenagers and do a written test, do many hours of lessons with a certified instructor, take the drivers test,
3) ???
4) Get your danish license. Honestly can you blame them though?

We South Africans might be 'good' drivers but only from the perspective of driving amongt the chaos of Gauteng traffic, watching out for smash and grabs at every robot ... (for my non South African readers ... I don't even have the fucking mental space to explain what that even is, I'll tell you when you are older) ... but we probably need some adjustments and rules in learning how sane responsible Danish citizens drive those same cars, especially around the cyclists who rule the road like the roving gangs in Mad Max.

How did I end up ranting about drivers licences though?
scrolls up 1000 lines where I scream into the void*
Ah right, needing to know some 'things' before you can drive.

We are back on track although ... knowing myself, we will not stay on track for long, so here is an unsubscribe button in case you want off this ride.

I am seeing a total lack of awareness about the thing I am about to discuss from so many people so I feel compelled to bring it up first because yall drive off into the sunset with your AI companions thinking that I have given you my blessing to do so.
I have not.
I know your Franjelicas™ Chilla™ is already in the cupholder but wait. Get out of the car. Not yet. We need to talk first.

How the sausage is made...

To understand the most important part of this post I need to dump some information on you, and some of you... not saying who but they sometimes call me Wooksil, might already be saying that this is too technical and that you just get flashbacks to high school algebra, but I assure you, as complex as these things may be, there is a simple enough underlying way that they work and once you understand that, you can benefit from them much more, and understand why they tend to do the things that they do.

The way that they work is that they took ALLLLLLL the text ever created by humanity, or very close to that, seriously, they got their hands on pretty much every piece of text content on the open internet, in libraries (that could be scanned) and nobody-quite-knows-just-what-else and ran it through a very complex algorithm that I don't even plan to attempt to understand, an algorithm that can only run on a special type of computer called a GPU, in particular this GPU

Or actually more accurately a bunch of these thingies.

But even more than that. Way more, like 40,000. And the program didn't run for like a few seconds, like opening Google Chrome, it didn't even run for several minutes, like when Windows 'decides' that it's time for a software update -please-dont-turn-off-your-pc-or-it-might-never-turn-on-again just as you are about to pack up your laptop for the day, seriously, like does windows have the ability to detect when you have had a particularly hard day and waits until you have 'reached the end of your tether' and are about to go home and says, "RIGHT! IT'S TIME FOR AN UPDATEEEEEEEE, NO THIS CAN'T WAIT THANK YOU VERY MUCH!" - no, not that long either. 90 DAYS. Of running 24/7. It uses so much electricity that it's pretty incomprehensible. It's like if you left all the lights on in the house and went to a movie and Daddy then came home from work and explained to you that you don't pay the electricity bill but if you did you would turn off each lightbulb every time you left the room, but times infinity.

The result of all that? A 'tensor checkpoint'. If you examined it, it would look like a bunch of numbers, and another file, split into thousands of smaller files (smaller as in terabytes each one, more than can fit on the 2000 dollar iPhone 17 Pro Max Ultra) and some other files that contain lots of words. Those files are then refined into smaller things, quantization and distillation and now I have to yada yada Seinfeld this because it goes beyond my understanding entirely. It's like PhD level thingies.

They end up with a 'model' which is a series of weights, whatever that means. Just think of it as a black box. Many experts literally do, and actually describe it as such.

With that black box, you can give it pretty standard input of some text, and things called a temperature, that is is used to determine how 'random' the output will be, how many words it must stop after, a few other things, (eg top k + top p values), which sounds complicated but really are just various volume dials of how random or predictable to be. It will then output text in response to that.

So you could give it something like this

Temperature: 0.1
Words: 1
Mary had a little ...

and it would output

lamb

But you can also give it

Temperature: 0.95
Words: 20
Mary had a little ...

And it would output

banana. She loves yellow. Mary is alien blue cheese garlic a farm that made...

If that is where this ride ended then it would have been no more than just an amusing computer experiment that would have been mildly amusing to some PhD students.

But it turns out that if you train it, as they did, on ALL of the text ever written, and give it some sensible values for temperature, not too random so it spits out nonsense about Mary the alien who grew up on a cheese farm, but also not a temperature too low that it just ends up sounding like the autocomplete on your phones keyboard like this...

... then you can get some pretty amazing results.

Mostly in the form of this.

You are a helpful AI assistant created by Acme Inc. Your job is to be a supportive, attentive listener with expert advice on psychology, CBT and life coaching. You are responding to David, who has recently moved to Copenhagen. The first message from the user is "____

then it sends your text that you gave it, aka 'The Prompt'

I am feeling really sad today.

and it will output

I'm sorry to hear that David. Sometimes when we move to a new country, especially one that has dark winters with lots of rain it can be tough on anyone...

And if it can do THAT, then the limits are maybe endless, nobody quite knows.

So that is how this type of AI, which are also called LLM's, Large Language Models work.

One quick aside because I find this fascinating...
The company that makes the machines that can train these thingies, is called Nvidia.

... Yes that Nvidia, the one that also makes the rainbow vomit blinged out thingy in the computer of your average teenagers bedroom that they play Fortnite on until 4am). That Nvidia is currently the only one that can make them has gone up in valuation since all this began from 120 billion dollars, to almost FIVE TRILLION DOLLARS.

By far the most valuable company of all time. More valuable than (on paper at least) - (yes even adjusted for inflation) - than the Dutch East India Trading Company which ruled pretty much most of the planet for a brief 200 year stretch back, back when my current apartment building was built.

And yet if you use it right, and understand some tricks and simple patterns then it can do some brilliant things, like that portfolio website

I can, and very much plan to show you some of that stuff but this post has just crossed over 2200 words and I REFUSE to use ChatGPT to condense it for me, stripping away all of my personality, so that will have to wait for the next edition.

After that, I would absolutely love to show you some of the incredibly cool things that you can get it to do, and showcase some of the things I have done with it, because there are so many possibilities that I have seen them myself over the past 3 years and alongside the bad, there is also so much potential for good.