What Is a Neural Network in Plain Words and How It Works

A neural network is a program that was never given ready-made rules. It derived them from examples. It was shown millions of texts, images or audio recordings, and it tuned billions of numeric coefficients inside itself until it could guess how a phrase continues, or assemble a picture from a description.

Everything else follows from that: why a neural network answers in human language but sometimes lies with full confidence, why the same prompt gives different answers, why it retells well and counts badly. Below is the breakdown: what it is made of, how it learned, and what happens at the moment you press send.

What a neural network is, in plain words

Picture a table with several numbers as input and one number as output. You do not know the formula that links input to output, but you have ten thousand filled-in rows. A neural network is a way to find that formula automatically, nudging coefficients until it matches the rows you already know.

The whole trick is scale. If the input is not three numbers but a 512 by 512 pixel image, and there are not ten coefficients but hundreds of billions, then the "formula" starts describing things like "there is a cat in this photo" or "after the words 'Paris is the capital of' usually comes 'France'". That is the same machinery behind every model you can open on Moleculs.ai, from a chat assistant to an image generator.

The word "neuron" is historical. The brain analogy gave the field its name, but there are no cells and no impulses inside: there is multiplication, addition and a simple nonlinear function. The model does not understand meaning, it finds stable patterns in data.

Neural network vs a regular program

A regular program is written as rules: if the order total is over a thousand, apply a discount. That logic reads line by line, you can test it and explain it to anyone. A neural network contains no rules. It has a structure and a set of weights, and nobody, not even the model's authors, knows why one particular weight equals 0.0731.

The practical conclusion is simple: do not put a neural network where you need one guaranteed answer. Let a calculator handle the tax math, and let the model write the email about it.

What it is made of: neurons, layers and weights

A neuron is a tiny function. It takes the numbers from the previous layer, multiplies each by its own weight, sums them up, adds a bias and passes the result through a simple rule like "zero out anything negative, keep the rest". One such operation is useless. Millions of connected ones are not.

A layer is a set of neurons working in parallel. Layers stack: the first one finds coarse brightness edges in an image, the next assembles corners and strokes from them, then fragments like an ear or a wheel, and the final layers produce the answer. Language models build abstraction the same way: from characters to words, from words to the meaning of a sentence.

Weights are the model's memory. Everything a trained network knows lives only in the numbers on its connections; there is no separate fact database inside. When people say "a 70 billion parameter model", they mean the count of those numbers.

[[СХЕМА: слои нейросети: вход - пиксели фото, промежуточные слои - линии, углы, фрагменты, выход - метка «кошка» | How features get more complex from layer to layer]]

How training works: examples, errors, adjustment

Training is a four-step loop repeated millions of times.

1. The model gets an example with a known answer: a photo and the label "cat".
2. It produces its own answer, random at first.
3. The error is measured: how far the answer is from the correct one.
4. Every weight shifts a little in the direction that reduces the error.

The key phrase is "a little". The adjustment step is kept small, otherwise the model swings around and never converges. That is why training a large model takes weeks on hundreds of accelerators instead of an evening on a laptop.

Overfitting is a separate trap. With too few examples the model memorises them and falls apart on anything new. The cure is volume and variety: for a narrow recognition task the order of magnitude is thousands of labelled examples per class, for a general-purpose language model it is terabytes of text.

Adapting a model to your own task is much cheaper. Retraining all the weights is almost never necessary. The usual approach is to take a finished model and train small additions to its weights, called adapters, with LoRA being the best known variant. A few hundred examples are enough to set the style or answer format you need without touching the core knowledge.

How a language model works: predicting the next chunk of text

A language model solves one task: given the text so far, predict the next small chunk. That chunk is called a token, and it ranges from a character to a word. A long or unusual word often splits into two or three tokens.

Your prompt is turned into tokens, the model computes probabilities for every possible next token, picks one, appends it to the text and computes again. A five-paragraph answer is several thousand such steps in a row.

The pick is not always the most probable option: some choices are made at random, and how much randomness there is depends on a parameter called temperature. That is why the same question gives different answers. For drafts and brainstorming the randomness helps; for pulling data out of a document you want it low.

Prompting practice follows from the same mechanism. The model leans on whatever is already in the text, so the more precisely you set the role, the format and the constraints, the less room it has to drift. You can verify this in a minute: send one question twice, adding "answer in three bullets of 20 words each" the second time, and compare.

How models draw images, make video and sound

Image generators are built differently, though training looks similar. The model learned on pairs of image and caption, with the images deliberately corrupted by noise, and its job was to remove the noise. At generation time the process runs in reverse: start from pure noise and, over a few dozen steps, clean it into an image that matches your text.

That is why image models have a step count, and why the result changes from run to run even with an identical prompt. It is also why small details like fingers, lettering and reflections are harder for them than the overall composition: there were fewer stable patterns behind those details in the data.

Video models add time. Neighbouring frames have to agree on objects and lighting, which is why clips are short and each run is expensive. Sound splits into two different tasks: generating music or effects from a description, and transcribing speech into text. Those are separate models, not one.

The difference shows up in run cost inside one subscription too: a cheap text model costs 1-2 credits per run, image generation usually costs tens of credits, and video is noticeably more. The cost of each model is shown next to it, and the plan quotas are on the pricing page. If you want to feel the difference in practice, start with something cheap, like Nano Banana for images, and see how many attempts it takes to get an acceptable result.

Why a model invents facts, and how to spot it

A language model is optimised for a plausible continuation of text, not for truth. If the fact is not in its weights, it will still produce a coherent sentence, because coherence is what it was trained on. That is how you get nonexistent legal clauses, invented quotes and books nobody ever wrote.

Signs that give a fabrication away before you check it:

• very specific numbers and dates with no source, especially round ones;
• links and titles that look perfect in format;
• an answer about recent events when the model did not search the web;
• names and job titles in a narrow field;
• a confident tone where you yourself expected "it depends".

A working habit: verify separately anything that can be confused. Dates, amounts, names, clause numbers, quotes. It also helps to ask for uncertainty markers right in the prompt, for example "if you are not sure about a number, write 'needs checking' instead of the figure". That does not remove the need to check, but it shortens the job.

What neural networks still cannot do

Count reliably. The model predicts symbols, it does not execute arithmetic, so long calculations and table reconciliation produce errors. Check the math with a calculator, or ask the model to show its steps so the mistake is visible.

Know what was not in the data. There is a cutoff date in training, and without a web search recent events simply do not exist for the model. For current facts you need models that search and cite, and even then it is worth opening the source.

Hold an endless conversation in its head. Context is limited: old messages get pushed out and the model "forgets" what you agreed on at the start. The fix is to repeat the important constraints in each new prompt.

Take responsibility. A neural network cannot tell where a mistake costs a penny and where it has legal consequences. The decision and the accountability stay with a human.

Try it yourself in four steps

The fastest way to get the principle is to ask one question and watch how the answer changes with the wording. Here is the order.

Step 1. Open any text model. You can start without an account: a guest can send one prompt and get an answer. Everything after that needs registration, by email or through Telegram.

Step 2. Pick a cheap model. You do not need a flagship for first experiments. Economy text models cost 1-2 credits per run, which makes them convenient for testing how the wording affects the answer.

[[СКРИНШОТ: текстовая модель с готовым запросом в поле ввода | A ready prompt in the input field before you run it]]

Step 3. Send a prompt built on four lines. This is the minimum that removes most of the noise from the answer:

Role: you are a teacher explaining to a school student.
Task: explain what a neural network is and how it is trained,
using the example of recognising cats in photos.
Format: five paragraphs of 3-4 sentences, then three takeaways as a list.
Constraints: no jargon; do not use the words "algorithm", "model",
"data"; if an explanation needs a formula, replace it with an
everyday comparison.
Copy

If you just want to paste it somewhere and press send, any GPT model will do.

Step 4. Run two control tests. First: send the same prompt again and compare the wording. The meaning will match, the text will not, and that is the randomness of token selection. Second: delete the "Format" line and watch the answer go shapeless. After those two runs, "the model predicts a continuation from context" stops being an abstraction.

The same framework carries over to other task types: it works for long-form writing, and, with a shot description instead of a format line, for images.

Mini glossary

Prompt - your whole request to the model: role, task, format, constraints, attached files. Everything you send is one piece of text as far as the model is concerned.

Token - the chunk of text the model operates on, from a character to a word. The model predicts its answer token by token, which is why a long answer takes longer than a short one.

Context - the amount of text the model keeps in front of it: your prompt, attached files and previous messages. In current models it is large but not infinite; when it runs out, the beginning of the conversation is pushed out.

Parameters - the numeric weights inside the model where everything it learned is stored. More parameters usually means better quality and a higher run cost, but not automatically: architecture and training data matter just as much.

Multimodality - the ability to accept more than text: images, files, audio or video. The set of inputs differs by model. Some read only text and images, others also audio and video, and the limit on attached images varies too. On Moleculs.ai that is listed for each model, and it is a list worth reading before you send the prompt rather than after.