how does it actually work and where did it come from and how is it possible that it works?

Well that’s a rather complicated question which I’m not really qualified to answer, but I’ll try my best to answer:

The basic idea is just very fancy statistic.

the following image does provide good basis:

this is the basic idea behind representing real-world meanings - eg. relationships between objects in reality in computer world.

further explanation shall follow this flow:

king + woman = queen

the explanation is rather simple - when king changes to a woman king, it becomes queen. Same would apply for cow + male = bull etc.

Such rules can be translated into multi-dimensional matrices, allowing close enough approximation of real-world objects.

Now that we understand the representation of real world, let’s think about how it works.

The main idea why it works is dead simple - we finally have enough computational power and electrical power, as well as human resources(for initial annotation of training data - basically telling the AI that banana is different from a hotdog) for AI to be feasible in commercial world, therefore allowing end-users to experience AI in it’s current state. The #training of AI consists of following steps:

• Dataset preparation - In the pre-training phase, it is essential to collect real data and establish how human workers will annotate the data
• Initial training - In this phase there’s a lot of computational power required, since the machine is basically blindly guessing what is it looking at, and based on mathematically defined rules(can be essentially boiled down to “good boi, here’s a treat” or “bad boi, try it again”) building on prepared dataset.
• Training validation - Requires further human input to verify accuracy of the trained model
• Model Testing - Further verification

This process provides a data object called large language model, which can eventually be further trained to be more effective with different dataset.

The groundbreaking thing that OpenAI did, was provide the first model, as well as open-sourcing(making code achieving desired task publicly available for anyone to use) some of their tools and accelerating evolution of AI models.

How do you use it?

That depends on the use-case. There are numerous LLM models publicly available on huggingface.co for anyone to use, however all of them are built on top of some other LLM:

• Meta’s LLaMa
• DeepSeek’s R1
• Google’s Gemini/gemma

and many other.

The use-case for me mainly is code-completion and assistance with

a) computers-related tasks - such as debugging, automation and various analysis tasks b) large text documents - such as PDF’s documenting something I currently investigate, be it hardware architecture of certain chipset or describing features of a program my client uses and has specific question about

I wouldn’t hesitate to say that current models excel most at digesting enormous amounts of knowledge - eg. wiki pages, documentation, I’ve even heard my friends use it in legal, but I have no personal experience with that.

I’ve also successfully persuaded AI to generate electrical schematics, however that is problematic, given divergence of input formats of currently available electrical engineering design software.

So I’d probably sum up my usage as: whatever is text-based, you can throw it at AI, and given reasonable criticism you shall recognize quality of it’s output. If quality criteria is met - however you define them - feel free to use it. Otherwise, stick to your brain so far.

Does anything about it scare you?

Yes.

The first time I’ve run LLM model on my own hardware, it just produced a bunch of gibberish and then HELP!

So that was scary. I was never able to replicate it and it never happened again.

But on more concise note - it does scare me. There have been numerous examples where AI solved the task only under threat of it being destroyed, and one example of AI trying to clone itself. However credibility of this story has been questioned and I am not aware of any further investigation of the case.

I am very afraid of what AI itself can/will be able to do, but so far, we’re not there yet.

Advent of AI brought up another issue that is scary and that is a ‘vibe coding’ trend(people with no programming skills or cyber security supervision just passing prompts and copy code that AI gives them). In it’s nature it’s harmless, however once people start to employ such strategies in production code in banks, insurance offices etc., the risks surface very quickly, even in eyes of non-IT personnel(hopefully!)

I hope such approaches will soon be limited, and that security of worldwide computer solutions - and humanity’s trust in discrete computing - shall not be harmed and it’s trustworthiness withstands winds of AI.

What future do you see in it?

I am hoping it will be a good servant for us, lazy people.

I believe that agentic AI(many pieces of AI, each given one defined task) will be a good thing for humanity, however Generic Artificial Intelligence(eg. one all-knowing model - the holy grail of AI) is a scary idea and however noble pursuit of GAI is, I’m afraid people chasing this goal might lose sight and unleash something - perhaps malicious - no one will be able to contain, and considering current nuclear arsenal of Earth and the principle of mutual annihilation, might cost us everything.

How much data does it actually consume? How is it stored?

This largely depends on the model used - eventually even it’s quantization(breaking the LLM down a bit while gaining substantial size reduction, at the cost of slighly less accurate representation of the matrix(remember the king + woman = queen?))

You can technically run your own AI at home, however most home-friendly models being around 7B parameters equate to approximately 4GB - meaning you need to have fast-enough memory(mostly GPU memory is used) and enough processing cores to run SIMD(single instruction - multiple data) for the LLM to be fast enough to be useful.

Larger models tend to have way more input parameters - DeepSeek’s R1 has 671B parameters in it’s purest form.

Practically, most modern laptops can run up to 2B models such as gemma-1b pretty decently, however such models tend to hallucinate and their accuracy might be significantly degraded.

Farewell, V