what sign up
nixm nixm

← learn

How does AI work? The plain-English version

October 14, 2026

Modern AI works by learning patterns from enormous numbers of examples instead of following rules a programmer wrote. A text model read a large share of the internet and practised guessing the next word until it got very good at it; when you type a prompt it continues from there, one small piece at a time. Image and video models do the same with pixels and frames. That is why it can write, draw and explain almost anything, and why it sometimes states something false with total confidence: it produces what is likely, not what is checked.

Two kinds of "AI"

For most of computing, software did what it was told: a programmer wrote rules, the machine followed them. Spam filters, chess engines and your thermostat worked like that, and people called some of it AI. The AI that changed everything in the 2020s is different: nobody wrote the rules. The system was shown examples, millions or billions of them, and adjusted itself until it produced the right answers. That is machine learning, and everything below is a version of it.

Learning from examples: the core idea

Imagine teaching someone to recognise cats by showing them ten million photos, each labelled cat or not cat, and letting them guess each time. Every wrong guess nudges their sense of "cat-ness" a little. After enough photos, they are very good at it, and they could not tell you the rule they are using. That is a neural network: a very large pile of numbers (parameters) that get nudged, guess by guess, until the outputs are right. "Training" is the nudging. It takes thousands of specialised chips running for months, which is why only a few companies build the biggest models.

How a text AI works

A language model like the ones behind ChatGPT, Claude or Gemini is trained on one task: given some text, predict the next word. It practised that on a huge amount of writing. To predict the next word well across all of it, the model had to absorb grammar, facts, how arguments go, how code is structured, how a recipe is laid out: not as rules, but as a compressed sense of what tends to follow what.

When you type a question, the model does not look it up. It takes your words, produces the most likely next word, adds it, produces the next, and so on until the answer is complete. Each step is a prediction. The whole paragraph is a chain of predictions that looked right at every point. A second phase of training, where people rate answers and the model is adjusted toward the ones they prefer, is what turns a raw predictor into something that answers questions helpfully instead of just continuing the text.

How image and video AI works

Image models learn from a large pile of pictures with captions. The common method starts from pure noise and, step by step, removes noise in the direction of "what an image with this caption would look like", until a picture appears. Video models do the same across frames, keeping each one consistent with the last, and recent ones generate the sound at the same time. Describe a fox in snow at dusk and you get one that never existed, built from the statistics of every snowy fox it saw. In nixm this is the same conversation as text: ask for the picture, then ask to animate it, and the clip comes back with sound.

Why it is so good

Because "predict the next piece" turns out to be a very deep task. You cannot predict the next line of a legal contract without having absorbed how contracts work, or the next line of a Python function without absorbing Python. Scale that across most of human writing and the model ends up with a usable grasp of most of what humans write about. Nobody fully expected this; it is the surprise of the decade.

Why it gets things wrong

Because likely and true are not the same thing. If the most plausible continuation of "the 1987 ruling in" is a case name, the model produces a case name, whether or not that case exists. It has no separate store of verified facts to check against, no sense of its own uncertainty beyond the statistics of the text, and no way to know that the world changed after its training data ended. Tools fix some of this: web search gives it current facts to read, code execution gives it real arithmetic. But the core remains a predictor, and the person reading the answer is the checker. The limitations page lists what follows from that.

What "training on your data" means

When a company says it trains on your conversations, it means what you typed goes into the pile of examples for the next version of the model, so that the model's sense of what follows what includes your words. That is why the privacy question matters and why nixm does not do it: nothing you say is used to train anything, no transcript is kept, and the conversation dissolves after seven days.

frequently asked

How does AI work in simple terms?

It learns patterns from millions of examples instead of following written rules, then applies those patterns to new input. A text AI learned to predict the next word across most of the internet and answers you by predicting one word at a time.

How does ChatGPT work?

It is a language model trained to predict the next word, then tuned with human feedback to answer helpfully. Each reply is generated one piece at a time, each piece chosen because it is likely given everything before it.

How does AI image generation work?

The model learned from a large set of captioned pictures. It starts from random noise and removes it step by step toward what an image matching your description would look like.

Why does AI make mistakes?

It produces what is likely, not what is verified. When the most plausible continuation is false, it states the false thing in the same confident tone. Check anything that matters.

What does training an AI mean?

Showing the model huge numbers of examples and adjusting its internal numbers after each guess until its outputs are right. It takes months on thousands of specialised chips.

Does AI learn from my conversations?

Many consumer tools use your chats as training examples unless you turn it off. nixm does not: nothing you say is used for training, no transcript is kept, and the conversation dissolves after seven days.

try it now — no account, no transcript kept. even the signal dissolves in seven days.

seed a signal →

Enter your email and receive a link to sign in. No passwords, just a secure signal to connect.

why connect

  • 20 free credits to start
  • your signals, remembered
  • buy more credits anytime
  • image & video generation features
or
+