Skip to content
Back to all articles
By TJ SoftUpdated:

How Does Artificial Intelligence Work? A List of Key TJ Soft Articles.

Jak działa sztuczna inteligencja? Spis kluczowych artykułów TJ soft.

The TJsoft blog already contains several dozen articles about artificial intelligence. Some focus on practical applications, while others are quick news updates (which become outdated fairly quickly). Many articles describe the general principles behind how artificial intelligence works; these will remain useful for many years and are a good complement to any texts about AI. If a new language model has a larger context window, it is useful to understand what this window actually is.

The following text is intended to facilitate the search for useful content; it is an initial attempt to create a Wikipedia by artificial intelligence.

Table of contents

Essential Articles

How does artificial intelligence work? Machine learning is the foundation, and autocorrect makes it easy to understand intuitively.

In this text, I explain how artificial intelligence works, from simple examples such as autocorrect to the complex mechanisms used in language models. The article is written in such a way that people who have no idea about artificial intelligence can understand it. Although before reading it, it is good to have some familiarity with, for example, ChatGPT.

I explain what machine learning is, how neural networks work, how models predict the next words, and how they differ from ordinary programs. I also describe concepts such as temperature, as well as the stages of model training (pre-training and fine-tuning).

I explain how AI recognizes images and what its limitations are. The text guides the reader step by step to understand where models’ responses come from and why they sometimes make mistakes, as well as what this means for the user in practice.

How to Start Using Artificial Intelligence in 2025 If You Haven’t Done So Before

If you think of artificial intelligence as something difficult, complicated, or “not for you,” that’s exactly why this text is worth reading. It shows that in 2025, using AI is just as simple as typing something into Google. You don’t need to install anything, you don’t need to know English, and you don’t even need to create an account. Just click and ask.

What will you find inside?

  • How to start from scratch: no stress, no technical knowledge required.
  • First practical applications: improving texts, searching for things whose names you don’t know, help with repairs, cooking, learning, and work organization.
  • The best models and tools: ChatGPT, Grok, Gemini, NotebookLM—what to choose and when.
  • How to write to AI to get sensible answers, and what to avoid.
  • Images, videos, audio, transcriptions—how to get more out of AI than just text.
  • Using AI safely: how not to fall for hallucinations, how to protect your data, and how to ensure legal security.

Why is it worth taking an interest in artificial intelligence?

Artificial intelligence can genuinely save you time, money, and stress. You can learn new things faster, create, fix things, and handle everyday tasks. If you approach it thoughtfully and with an open mind, artificial intelligence will become your second brain. And it all starts with a single click.

This text isn’t a dry guide, but an accessible introduction for everyone. Even if you’re at the “I have no idea what this is about” stage.

Articles on the Technical Foundations of AI

Artificial Intelligence Concepts. Part Two.

This text serves as a practical guide that helps organize vocabulary related to artificial intelligence and makes it easier to navigate the world of language models. It was created as a companion to other articles, so it works best as a supplement to basic knowledge. If you are just beginning your journey with AI, it is worth reading the introduction explaining how language models work, in which I explain concepts such as tokens, vectors, embeddings, layers, and temperature step by step. This will make this glossary a natural extension of that article and help you understand more quickly the meaning of technical terms that appear in the world of artificial intelligence.

There are so many terms related to artificial intelligence that it is difficult to fit them all into a single article, even if each term is described in one paragraph. In this case, the article focuses on more fundamental topics.

Transformers in Artificial Intelligence: How They Work and Why They Changed the World

In this article, I explain what transformers are and why they sparked the greatest revolution in artificial intelligence. I show how the underlying mechanisms of attention, tokens, embeddings, BERT, and GPT work, as well as how they moved from laboratories into search engines, translators, chatbots, and multimodal models. This raises the question of whether transformers truly “understand” language or merely pretend to do so very cleverly.

What exactly is a neuron in artificial intelligence?

Anyone even minimally interested in artificial intelligence has surely heard of deep learning or neural networks. This brings us to the title question: what exactly is this neuron in AI? In short, a neuron is simply a mathematical function that determines what information to pass on.

Levenshtein Edit Distance in the Context of Artificial Intelligence and Autocorrection

Edit distanceThis innocuous-looking tool is one of the foundations of how autocorrect, search engines, and NLP systems work. When you type “komputerz,” no magic happens—the algorithm is simply trying to determine which word is closest to your mistake and suggest the correct form. By understanding the mechanism behind correcting typos, as well as the mechanism behind suggesting the next word on a smartphone keyboard, it becomes easier to understand how large language models work.

Backpropagation, backpropagation, backpropagation algorithm, backpropagation of errors

Language models use many layers. During AI training, after passing a task through all the layers, we obtain an output. This output is then compared with the model (manually verified) output, and in this way we determine the error value. Knowing this error, we need to modify the parameters in many layers so that they produce better results. This process is called backpropagation.

RAG (Retrieval-Augmented Generation)

RAG is a technique that makes artificial intelligence stop “guessing” and start using real information. Before the model generates an answer, it first searches for the necessary data in documents and only then creates an answer based on them. This makes it more accurate, more reliable, and helps it avoid hallucinations. The article explains how this mechanism works and why it has become one of the most important elements of modern AI systems.

By default, all the knowledge of language models is hidden in encoded connections between words. Until we ask AI, we do not know what “it knows.”

What is the gradient descent method?

Gradient descent is one of the fundamental techniques used in machine learning and artificial intelligence. It allows models to gradually adjust their parameters so that their predictions become increasingly accurate.

With this method, the model learns from its mistakes: first, it guesses an answer, compares it with the correct one, and adjusts its parameters so it can get closer next time. The process is repeated thousands of times, and each change is a small step “downward,” toward the best possible result. Because the calculations are too complex to find the ideal parameters immediately, gradient descent relies on approximations and gradual improvement.

Context window

Have you ever wondered why ChatGPT handles short answers quite well, but writes poorly when producing longer texts? The cause of this problem is the context window, and the fact that enlarging the window requires enormous computing power.

How Does Artificial Intelligence Really Learn? Training Artificial Intelligence for Non-Technical People

To fully understand how language models work, it is good to know how they are trained. Training models consumes much more energy and computing power than using ready-made models. In short, creating artificial intelligence involves assigning appropriate parameters to words (or, more precisely, tokens). The idea is for words with similar meanings to be “close to each other.”

For non-technical readers, this is an excellent starting point for understanding concepts such as vectors, fine-tuning, and the attention mechanism. The text also explains why AI can hallucinate, where unusual associations come from, and what role a knowledge base (RAG) plays in providing up-to-date information.

Let’s assume that we don’t like the existing language models and want to create something of our own, from scratch. Creating a new model that matches GPT 5.2 under home conditions is practically impossible. However, adapting existing models (fine-tuning) and adding our own knowledge base (RAG) is relatively simple, and for narrow applications, anyone can create a tool that will work better than the latest ChatGPT or Grok.

What Exactly Is Temperature in Artificial Intelligence?

This article explains what temperature really is in language models and why the popular comparison to a “creativity slider” can be misleading. Temperature is a technical parameter that regulates the randomness of the words selected by the model—with a low temperature, AI chooses the most probable words (predictable responses), while at a high temperature, it allows less obvious options (more chaos). The problem is that high temperature does not mean genuine creativity—the model does not create anything new, but merely selects from options it has already learned according to a modified probability distribution. Unlike human creativity, which requires intention, purpose, and awareness of context, temperature simply increases the likelihood of unusual combinations of words, which often leads to nonsense instead of valuable ideas.

In other words, creativity is more than just writing style. Whether the text should be formal or casual can easily be specified using the appropriate commands. It would also be easy to create an overlay with sliders for setting this (although it would not be very practical). However, creativity itself stems from the quality of the model—the more advanced the model, the more creative it is. Creativity is not just fairytale imagination, poetry, and butterflies; it is also important for tedious and technical tasks.

Practical Applications of AI

ChatGPT – A Practical Guide and Program Development.

Most guides on ChatGPT focus on describing the language model itself. In this article, I have focused on the program’s structure and its features (such as Projects, file uploads, Canvas, the DeepSearch feature, keyboard shortcuts, etc.

How to Start Using Artificial Intelligence in 2025?

I described this article on how to start using AI in the introduction to the article you are currently reading. I decided it would be useful to include it in this section as well, so that all texts in this category would be in one place.

In 2025, using AI is as easy as typing something into Google. You don’t need to install anything, create an account, or know English. Just visit chatgpt.com or grok.com and start typing. This guide shows you how to go from having no experience at all to making practical use of artificial intelligence in everyday life.

Where to start:

  • Treat AI like an all-knowing assistant that you need to provide with context
  • Start with simple tasks: improving texts, searching for things (especially when you don’t know the name), recipes, repairs
  • Add photos when you don’t feel like describing something or don’t know how to
  • Ask how long a repair will take if you do it yourself and whether it’s worth doing it yourself

Practical applications:

  • Cooking: find a recipe for a dish whose name you don’t know, and adapt it to the ingredients you have
  • Repairs: take a photo of the issue, describe your skills, and find out whether you can fix it yourself
  • Writing: overcome writer’s block by dictating your thoughts, edit texts, and shorten or lengthen passages
  • Learning: Deep Search mode will search dozens of websites and provide you with a ready-made study in 10–20 minutes
  • Notebook LM: upload YouTube videos and “talk” to them, create notes and flashcards

Graphics and video:

  • Image generation: simply say “create an image” in ChatGPT, Grok, or Gemini
  • Grok creates videos with sound for free (a dozen or so per day)
  • Nano Banana Pro (Google) excels at charts and data
  • You can modify real photos, e.g. place a product in a new environment

Security:

  • Always verify facts – AI may hallucinate (make up information)
  • Do not paste passwords or private data into cloud-based models
  • Avoid leading questions (“what happened at company X on March 5?”)
  • Add to prompts: “if you don’t know something, say so; don’t make it up”

Helpful tricks:

  • Ask AI to write a prompt for itself—it works surprisingly well
  • Use the formula: Role + Topic + Format (“You are a math teacher; explain integrals in the form of an article”)
  • Dictate texts in ChatGPT—it handles punctuation very well
  • In a new thread, paste the AI’s response and ask, “Is this true?”—a simple way to verify it

Python – the language of artificial intelligence and a useful tool

This article shows how Python became the foundation of modern AI and why anyone can use it today.

The article explains the phenomenon of Python in the world of artificial intelligence. It combines simplicity with power—the code resembles natural English while providing access to powerful tools. Libraries such as NumPy, Pandas, and TensorFlow allow scientists to focus on creating algorithms instead of struggling with complicated syntax.

AI plays a key role here, completely transforming the way we work with Python. Today, anyone—even without programming knowledge—can create a useful program in just a few minutes. Need a script that sorts files? A bot that checks prices in stores? A program for automatically sending emails? Simply describe to the AI what you want to achieve, and you’ll receive ready-to-use code. It’s like having a personal programmer who never gets tired.

The article walks through practical applications—from automating tedious office tasks and processing data to creating intelligent bots. It shows that you don’t have to be a genius to get started. Just three hours are enough to learn the basics and start creating solutions to real-world problems. Python is no longer the domain of a select few—it has become a tool accessible to everyone.

Vibecoding

A guide to creating programs and task automations using AI, aimed at people without programming experience.

The text explains that the true power of vibecoding does not lie in creating complex commercial applications, but in streamlining boring, repetitive tasks—especially in spreadsheets. Most people use Excel, and with AI, it is possible to create advanced formulas and macros that save dozens of hours each month.

A gradual learning path is presented. It begins with play—creating simple games in ChatGPT to explore what is possible. It then moves on to what truly transforms work: spreadsheets. AI can help create a formula for splitting columns, a Gantt chart, or an attendance planning system. Tasks that previously took hours can now be completed in a dozen or so minutes. The key is realizing that something can be automated at all—the rest ChatGPT will do.

The subsequent stages include creating websites, installing Python and writing local programs, and finally using tools such as Bolt or Figma to build fully fledged applications. The article contains practical tips: save working versions, one command—one change, and if something doesn’t work after three attempts, roll it back.

The most important lesson? You don’t have to be a programmer. You just need to know that something is possible and ask AI for help.

Google Gemini can create AI presentations. This is a major game-changer.

Creating presentations with artificial intelligence is one of the best practical applications of AI. Google Gemini offers an automatic presentation generation feature that can significantly speed up your work, especially when you already have the text and research prepared.

How does it work? Simply select “Canvas” mode in Gemini and ask it to create a presentation. Important: we must start in Gemini, not in Google Slides. Gemini is not yet able to create presentations directly in Google Slides, even in the premium version.

Main disadvantages:

  • Minor technical errors may sometimes occur, such as overlapping text
  • AI may misinterpret the content and omit important information
  • Presentations have a schematic, repetitive layout
  • Image licensing issues — Google uses images from various sources, which may make it impossible to share presentations publicly without replacing the graphics

Key advantages:

  • Huge time savings when creating the first version
  • Ability to edit in Google Slides
  • Ideal for internal presentations or as a learning tool (creating visual notes)
  • Thanks to the speed of creation, making presentations for yourself as a learning aid becomes worthwhile

The tool works best as a starting point—AI creates a solid foundation that can then be refined manually. Other options include Gamma.app, Napkin.ai, and features in ChatGPT and Claude.

How to Talk to ChatGPT and Whether It’s Worth Learning Prompts

The most important rule: using AI is just a regular conversation, not magic spells. Learning ready-made prompts from lists is a waste of time; it’s better to simply talk to the model and clarify what you need.

Why ChatGPT Makes Mistakes

Most problems stem from a lack of context. A simple formula—who I am + what I need + who should help me—solves the issue in 90% of cases. You can add the format (email, list), style (casual, formal), and what to avoid.

What Really Works

A few tried-and-tested ways:

  • Set the scene: “You are a mechanic repairing a car from the ground up—describe the process”
  • Mix styles: “Write as if Ludwig von Mises met a contemporary journalist”
  • Add motivation: “Do your best and you’ll get 50 PLN” (yes, there is research confirming that this works)
  • Use reinforcements: “make sure,” “remember”
  • Ask for a prompt: “I want to achieve X—write a good prompt for me”

The fastest method? Write something yourself, even chaotically, and ask AI to improve it. A 4-hour task turns into 20 minutes. Don’t worry about grammar—the model understands even a mess.

What about “prompt engineering”?

It’s a relic from the days when models were weak. Today, it’s mainly a title for people who don’t know what to call themselves. If you test AI tools but don’t build systems, you’re more of an “AI operator.” Typical prompt engineering applies when you create specific, complex things—for example, advanced AI graphics, where each image is described using a dozen or so parameters.

Artificial Intelligence in University Studies – A Brief Guide

For today’s student, AI is not a tool for cheating, but a solid tool that supports learning. It does the drudge work: searches sources, formats references, and writes scripts, leaving time for thinking and creativity. The key is to use tools connected to reliable sources.

Deep Search in ChatGPT or Perplexity helps you quickly get oriented in a new topic. Consensus.app searches only peer-reviewed academic publications, showing whether there is agreement on a given issue. Google’s NotebookLM works exclusively with the files you provide; you can upload a link to a lecture and “talk” to the video without scrolling through it. AI also works excellently as an editor: you paste in a draft and ask it to improve the style without changing your arguments. Thanks to “vibecoding,” even a humanities student can create advanced Excel formulas or Python scripts by describing the problem in Polish.

The most important rule is transparency. Using AI as an assistant is acceptable as long as you acknowledge it: concealing or submitting work generated entirely by a machine is asking for trouble. Safe uses include proofreading your own text, explaining difficult concepts, helping structure your work, or conducting research using dedicated tools. Risky uses include generating entire chapters through “copy and paste,” citing unverified sources, or claiming as your own conclusions invented by the machine.

The existence of AI is somewhat changing the role of human memory. For many tasks, the information itself is less necessary; what matters more is the context needed to find it. On the other hand, this creates a certain trap, because sometimes it is better to have specific information—or even hard data—in your head, as asking ChatGPT about it every time is a waste of time. That is why taking notes by hand on paper still fosters creativity, so don’t throw away your traditional notebook. Curiosity remains the most important thing: AI will answer any question, but you are the one who has to come up with the idea to ask.

Google Finally Makes Good AI

This article traces Google’s return to the forefront of artificial intelligence. Although Google was a pioneer in AI (neural machine translation, Google Assistant), its initial responses to ChatGPT—Bard and early versions of Gemini—proved disappointing. The breakthrough came in early 2025, when Google stopped imitating ChatGPT and began creating unique solutions that leveraged its strengths: integration with its service ecosystem (Drive, YouTube, Maps, Gmail) and the development of its own increasingly powerful text models. The key to success was not merely catching up with the competition, but charting its own paths forward—creating specialized tools instead of a single universal chatbot.

The article discusses three main Google AI products in detail. NotebookLM is a research assistant that accepts up to 50 sources (PDFs, Google Docs, links, and YouTube videos), with the option to select which materials are active for each question, and offers podcast generation, mind maps, summaries, and FAQs. Google AI Studio is a free developer platform particularly well suited for processing audio and video into transcripts. Gemini offers Deep Research, which creates long, coherent texts, as well as unique integrations with YouTube (searching channels), Drive, and Maps (planning routes based on criteria such as “a restaurant with a playground”), along with Canvas—an editor for texts of up to 2,000 words with export to Google Docs, better than the ChatGPT version. Google’s strategy is to create multiple specialized tools instead of one universal application, providing greater functionality but making integration between products more difficult.

I added this text in May 2025. Since then, Google has developed significantly. Gemini 3 surpassed ChatGPT’s latest model, prompting OpenAI to declare code red and quickly release the GPT-5.2 model. This is not merely about reaching the top of some benchmark—Grok, Claude, and Gemini have all occasionally managed to outperform ChatGPT there; it is about a range of benchmarks, user opinions, and the number of app downloads. Gemini 3 performs significantly better than ChatGPT at frontend coding, especially when using the Antigravity tool. Additionally, Google’s services are becoming increasingly well integrated.

AI Ecosystem

To fully understand the situation surrounding artificial intelligence, it is useful to know the background of companies’ activities. Development is taking place all over the world, not only in the USA. AI is not just ChatGPT.

Polish AI Companies

This article presents an overview of the Polish ecosystem of companies developing artificial intelligence–based solutions, divided into five main industry categories. In medicine and biotechnology, notable companies include Wrocław-based Infermedica (medical triage platforms, with an app that has been downloaded 500,000 times), Warsaw-based Molecule.one (generative AI for designing drug synthesis), DeepFlare (mRNA vaccines), and Labplus (automation of laboratory analyses). In finance and cybersecurity, key players include Authologic (digital identity verification), Nethone (fraud detection through behavioral analysis, blocking 95% of unauthorized access attempts), the Kaspro chatbot, and the STIR System (which blocked PLN 180 million in suspicious transactions). Industry and logistics are represented by Nomagic (intelligent warehouse robots using computer vision), AI Clearing (construction site monitoring using drones and 3D scans), and Surveily (computer vision for occupational health and safety).

In customer service and marketing, the leading companies include Krakow-based Synerise (real-time customer behavior analysis, valued at over €100 million), Warsaw-based Cosmose AI (offline behavior analysis using phone and WiFi data, serving 20 million locations for LVMH and L’Oréal, valued at PLN 2 billion), Szczecin-based Tidio (an AI system for e-commerce with over PLN 100 million in funding), as well as SentiOne and VoiceLab (voicebots and NLP). In generative AI, Eleven Labs stands out—it was founded by Poles in the United States, is valued at $3.3 billion, and develops advanced speech synthesis technology with its main R&D team in Poland. The article demonstrates that Poland’s AI scene has transformed into a dynamic ecosystem competing globally—from medicine and Industry 4.0 to voice technologies—with many companies achieving international commercial and technological success.

State of Warsaw AI — summary

State of The Warsaw AI 2024 is a comprehensive report on artificial intelligence in Warsaw, showcasing the capital as an AI leader in the Central and Eastern European region.

Key figures:

  • There are approximately 1,000 AI companies operating in Poland, 46% of them in Warsaw
  • 15% of Polish companies are implementing AI technologies (approximately 45,000 businesses in Warsaw)
  • Poland plans to allocate €119.8 million to AI development between 2024 and 2030
  • Approximately 6,500 AI specialists work in Warsaw, and their number is growing by 54% annually

Key findings: Most AI companies (60%) are less than 5 years old, and 70% are not yet ready for expansion. The sector is young and dynamic, dominated by startups. The main areas of development are data analysis, machine learning, and generative AI.

Challenges: The biggest problem is the lack of cooperation between the public and private sectors (65.5% of respondents). Only 29% of companies benefit from public assistance due to bureaucracy and a lack of information.

Where to gain knowledge about artificial intelligence?

“Where to Gain Knowledge About Artificial Intelligence When You Have Little Time?” is a guide to the best sources of AI knowledge for people with limited time.

The article highlights reliable places where you can quickly catch up on what you’ve missed. Among Polish sources, it recommends the YouTube channels Adrian Kilar and Netgonet, which offer concise reviews of the latest developments every two weeks, while the Bielik community on Discord brings together enthusiasts developing a Polish language model. Among official portals, it recommends gov.pl/ai and Haimagazine.com/pl.

On the English-language side, Bush AI Cave stands out—the best news aggregator, publishing one-sentence summaries of the most important events every day. For those seeking a deeper understanding, it recommends Two Minute Papers (several-minute discussions of scientific research) and AI Explained (analysis of key events). Matt Wolfe showcases practical tools, while Lex Fridman conducts hours-long interviews with industry leaders such as Sam Altman and Yann LeCun.

The article also lists essential tools: LLM Arena for testing AI models without creating an account, There is an AI for That as a database of thousands of AI applications, and The Rundown AI—a daily five-minute newsletter featuring the most important news.

History of Artificial Intelligence

The history of AI begins with ancient legends—the Greek bronze Talos or the Jewish clay Golem—which reveal humanity’s age-old desire to create artificial life. Aristotle laid the foundations by formalizing logical reasoning, while Leibniz created the binary system. In 1843, Ada Lovelace foresaw that machines could manipulate not only numbers but any symbols—a vision of the general-purpose computer.

The real revolution began with Alan Turing, who proposed his famous test in 1950: if you cannot distinguish a machine from a human in a conversation, then the machine is intelligent. In 1956, the field of AI was officially born at the Dartmouth conference, and its founding fathers believed they would solve the problem of intelligence within a year. They were very wrong.

Golden years arrived, followed by painful disappointments. In 1966, ELIZA proved how easy it is to deceive people with a simple chatbot, while the Shakey robot showed how difficult it is to operate in the real world. Then came the “AI winters”—periods when enthusiasm froze and funding dried up. The turning point was 1997, when Deep Blue defeated Kasparov—not through intelligence, but through brute computing power.

The 21st century has brought three waves of revolution: deep learning (2012, the AlexNet breakthrough), the Transformer architecture (2017, “Attention Is All You Need”), and finally the generative explosion (2022), when ChatGPT went mainstream. It was AI’s “iPhone moment” – technology wrapped in a simple conversational interface.

The irony of history is that AI pioneers such as Weizenbaum and Hinton became its greatest critics, terrified by the power of their own creations.