What Is Artificial Intelligence and How It Really Works

👁 117 reads

Artificial intelligence is software that solves tasks based on what it learned from data, instead of executing rules written explicitly by a human. That’s the entire definition. The rest of this article explains what follows from it, because this is exactly where most explanations become murky.

The Difference That Explains Everything

A regular program is a list of commands. Someone sits down and writes: if the amount exceeds one thousand dollars, mark as large order. The program does exactly that and nothing else.

A learning program gets examples instead: a thousand orders with information about which ones turned out to be problematic. It searches for patterns in them on its own and evaluates the next ones based on that. Nobody wrote a rule about one thousand dollars, yet the system can discover it, or find another one that nobody thought of.

The consequence is more important than it seems: the program’s author doesn’t know exactly why the system made a specific decision. They know what it was trained on. That’s why with AI we talk about testing and measurements, not about reading code.

Three Terms People Confuse

Artificial intelligence is the broadest term, covering everything that mimics abilities considered to require intelligence.

Machine learning is a way of building such systems by training them on data. This is the part that almost everything being built today relates to.

Language model is a specific type of such system, trained on huge amounts of text, predicting what should come next. All the well-known chatbots are language models.

So when someone says “AI,” nine times out of ten they mean a language model, which is the narrowest of these three concepts.

How It Works in Practice

During training, a language model read a huge amount of text and calculated how words relate to each other. It didn’t memorize sentences. It memorized relationships.

When you ask a question, the model adds successive fragments of the answer, each time choosing the one that best fits everything said before. It does this element by element, not sentence by sentence.

This is where the two things that surprise people most come from.

Why the same question twice gives different answers. Because the choice of the next element isn’t predetermined, but randomly selected from among those that fit. You change one word in the question and the entire rest goes a different way.

Why the model can write falsehoods with complete confidence. Because its task is to select what fits, not to verify whether it’s true. A false sentence can fit just as well as a true one. This isn’t a bug to fix in the next version, but a property of the mechanism itself.

What AI Doesn’t Do, Though It Looks Like It Does

It doesn’t understand in the human sense. It can correctly explain a concept it doesn’t “grasp,” just as a parrot can repeat a sentence.

It doesn’t remember you between conversations, unless the tool has a separate saving function. The model itself starts every conversation from zero.

It doesn’t calculate like a calculator. The result of an operation is also predicted, which is why with longer calculations errors occur in places where a calculator never makes mistakes.

It doesn’t have internet access unless given it. Models in their basic version answer from what they learned up to a certain date. Web search is a separate function that must be enabled.

Will AI Take Jobs

The question comes up in every conversation, so I’ll answer honestly: I don’t know and nobody knows. However, we do know what’s happening now, and that’s more instructive than forecasts.

The fastest changes are in professions involving text processing according to a pattern: rewriting, summarizing, translating simple content, first drafts of documents. They’re not disappearing, but one person does today what three used to do.

The slowest changes are in everything that requires responsibility for the outcome. A model can prepare a document, but won’t sign it. That boundary is currently clearer than technological boundaries.

Where to Start If You Want to Try

Not by installing anything. The biggest chatbots have free plans sufficient to check whether this is useful in your work. What exactly you get without paying, I’ve detailed in articles about free ChatGPT and in the Gemini pricing, where the free plan is unusually generous.

A good first task is something you can verify in five seconds. A summary of a text you know. Rewriting an email. Explaining a concept from your industry that you know everything about. You’ll see immediately where the tool is good and where it hallucinates.

Quality differs greatly between models, including in non-English languages. We measured this on six grammar tasks in Polish (a heavily inflected language): four models scored full points, and the price difference between them reached thirty-two times. The results are in the article about which AI model writes best in Polish.

Once you know what you want, the next step is learning to formulate instructions. I’ve compiled this in the article about what a prompt is and how to write instructions.

More About Why Models Hallucinate

I’ve explained the mechanism here briefly. If you want to understand it more deeply, along with how to defend against it in everyday work, I have a separate article about why AI hallucinates.

Plus the practical side of Polish text: errors that give away the machine in Polish text, meaning calques, numerals, and long dashes.

Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →

What This Means for Work

The most common question after such an explanation is: what about my job. I’ve compiled an answer based on sources, not forecasts, in the article about artificial intelligence and work. There’s also a calculator where you check your own tasks and see which can realistically be supported today.

Kinetyra